{
    "paper_id": "2022",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T01:11:41.791234Z"
    },
    "title": "Lifelong Pretraining: Continually Adapting Language Models to Emerging Corpora",
    "authors": [
        {
            "first": "Xisen",
            "middle": [],
            "last": "Jin",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Southern",
                "location": {
                    "country": "California"
                }
            },
            "email": "xisenjin@usc.edu"
        },
        {
            "first": "Dejiao",
            "middle": [],
            "last": "Zhang",
            "suffix": "",
            "affiliation": {
                "laboratory": "AWS AI Labs",
                "institution": "",
                "location": {}
            },
            "email": "dejiaoz@amazon.com"
        },
        {
            "first": "Henghui",
            "middle": [],
            "last": "Zhu",
            "suffix": "",
            "affiliation": {
                "laboratory": "AWS AI Labs",
                "institution": "",
                "location": {}
            },
            "email": "henghui@amazon.com"
        },
        {
            "first": "Wei",
            "middle": [],
            "last": "Xiao",
            "suffix": "",
            "affiliation": {
                "laboratory": "AWS AI Labs",
                "institution": "",
                "location": {}
            },
            "email": "weixiaow@amazon.com"
        },
        {
            "first": "Shang-Wen",
            "middle": [],
            "last": "Li",
            "suffix": "",
            "affiliation": {
                "laboratory": "AWS AI Labs",
                "institution": "",
                "location": {}
            },
            "email": "shangwenl@amazon.com"
        },
        {
            "first": "Xiaokai",
            "middle": [],
            "last": "Wei",
            "suffix": "",
            "affiliation": {},
            "email": "xiaokaiw@amazon.com"
        },
        {
            "first": "Andrew",
            "middle": [],
            "last": "Arnold",
            "suffix": "",
            "affiliation": {
                "laboratory": "AWS AI Labs",
                "institution": "",
                "location": {}
            },
            "email": "anarnld@amazon.com"
        },
        {
            "first": "Xiang",
            "middle": [],
            "last": "Ren",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of Southern",
                "location": {
                    "country": "California"
                }
            },
            "email": "xiangren@usc.edu"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Pretrained language models (PTLMs) are typically learned over a large, static corpus and further fine-tuned for various downstream tasks. However, when deployed in the real world, a PTLM-based model must deal with data distributions that deviate from what the PTLM was initially trained on. In this paper, we study a lifelong language model pretraining challenge where a PTLM is continually updated so as to adapt to emerging data. Over a domain-incremental research paper stream and a chronologically-ordered tweet stream, we incrementally pretrain a PTLM with different continual learning algorithms, and keep track of the downstream task performance (after fine-tuning). We evaluate PTLM's ability to adapt to new corpora while retaining learned knowledge in earlier corpora. Our experiments show distillation-based approaches to be most effective in retaining downstream performance in earlier domains. The algorithms also improve knowledge transfer, allowing models to achieve better downstream performance over the latest data, and improve temporal generalization when distribution gaps exist between training and evaluation because of time. We believe our problem formulation, methods, and analysis will inspire future studies towards continual pretraining of language models.",
    "pdf_parse": {
        "paper_id": "2022",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Pretrained language models (PTLMs) are typically learned over a large, static corpus and further fine-tuned for various downstream tasks. However, when deployed in the real world, a PTLM-based model must deal with data distributions that deviate from what the PTLM was initially trained on. In this paper, we study a lifelong language model pretraining challenge where a PTLM is continually updated so as to adapt to emerging data. Over a domain-incremental research paper stream and a chronologically-ordered tweet stream, we incrementally pretrain a PTLM with different continual learning algorithms, and keep track of the downstream task performance (after fine-tuning). We evaluate PTLM's ability to adapt to new corpora while retaining learned knowledge in earlier corpora. Our experiments show distillation-based approaches to be most effective in retaining downstream performance in earlier domains. The algorithms also improve knowledge transfer, allowing models to achieve better downstream performance over the latest data, and improve temporal generalization when distribution gaps exist between training and evaluation because of time. We believe our problem formulation, methods, and analysis will inspire future studies towards continual pretraining of language models.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Pretrained language models (PTLMs) have achieved remarkable performance on benchmark datasets for a range of NLP tasks (Liu et al., 2019b; Brown et al., 2020) . However, when deployed in the wild, NLP systems must deal with emerging data that have constantly shifting data distribution, different from the text corpora they were initially pretrained on -for example, when new data domains are introduced (upper part of Fig. 1 ) (Gururangan et al., 2020) , or when the language uses and vocabulary change over time (lower part of Fig. 1 ) (Lazaridou et al., 2021) . Fine-tuning from a Figure 1 : Two data streams created for studying lifelong language model pre-training. We focus on evaluating knowledge retention on the domain-incremental research papers stream; we focus on adaptation to the latest data and temporal generalization on the chronologically ordered tweet stream. static and possibly \"outdated\" PTLM may limit the model performance on downstream tasks, as the PTLM may no longer provide an effective model initialization (Beltagy et al., 2019; M\u00fcller et al., 2020 ). Here we look to understand whether continuously adapting a PTLM to emerging data can yield gains on various downstream tasks, and how to achieve better downstream performance for such lifelong PTLM adaptation.",
                "cite_spans": [
                    {
                        "start": 119,
                        "end": 138,
                        "text": "(Liu et al., 2019b;",
                        "ref_id": "BIBREF32"
                    },
                    {
                        "start": 139,
                        "end": 158,
                        "text": "Brown et al., 2020)",
                        "ref_id": "BIBREF6"
                    },
                    {
                        "start": 428,
                        "end": 453,
                        "text": "(Gururangan et al., 2020)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 538,
                        "end": 562,
                        "text": "(Lazaridou et al., 2021)",
                        "ref_id": null
                    },
                    {
                        "start": 1036,
                        "end": 1058,
                        "text": "(Beltagy et al., 2019;",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 1059,
                        "end": 1078,
                        "text": "M\u00fcller et al., 2020",
                        "ref_id": "BIBREF38"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 419,
                        "end": 425,
                        "text": "Fig. 1",
                        "ref_id": null
                    },
                    {
                        "start": 529,
                        "end": 535,
                        "text": "Fig. 1",
                        "ref_id": null
                    },
                    {
                        "start": 584,
                        "end": 592,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "A number of recent works make attempts on adapting PTLMs to a new data domain. Gururangan et al. (2020) ; adapt language models to corpora of different genres and topics and observe performance improvement in domainspecific downstream tasks. Arumae et al. (2020) further show that by regularizing the parameters of PTLMs, the downstream tasks performance on the general domain can be preserved. Another line of works focuses on temporal domain shift (Hombaiah et al., 2021) , which analyzes the effect of pretraining over up-to-date data to the downstream tasks. R\u00f6ttger and Pierrehumbert (2021) further study vocabulary composition approaches for improving adaptation to up-to-date corpora. However, these work focus their study on adapting PTLM to a single new domain; while in practice, corpora from distinct domains and time stamps may emerge sequentially. Whether one can maintain a single, up-to-date PTLM remains an open problem. Related to this, Lazaridou et al. (2021) study adaptation of PTLMs over temporal data streams, but solely focus on language modeling instead of fine-tuning performance. It is also important to understand multiple aspects of the utility of lifelong PTLM pretraining, such as knowledge retention over all the seen data, and study what methods can improve the utility of PTLMs in such a continual pretraining process.",
                "cite_spans": [
                    {
                        "start": 79,
                        "end": 103,
                        "text": "Gururangan et al. (2020)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 242,
                        "end": 262,
                        "text": "Arumae et al. (2020)",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 450,
                        "end": 473,
                        "text": "(Hombaiah et al., 2021)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 563,
                        "end": 595,
                        "text": "R\u00f6ttger and Pierrehumbert (2021)",
                        "ref_id": "BIBREF46"
                    },
                    {
                        "start": 954,
                        "end": 977,
                        "text": "Lazaridou et al. (2021)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we formulate a Lifelong Language Model Pretraining task to simulate practical scenarios of maintaining and adapting a PTLM over emerging corpora, create a testbed (along with pretraining data streams and downstream tasks) for studying continual pretraining algorithms, and present a systematic evaluation protocol for measuring the progress made on this challenging problem (see Figure 2 for an illustration). We consider two types of text corpus sequences when constructing pretraining data streams, each of which simulates a representative use case and that has slightly different focuses on the evaluation: continuously learning a single model that is applicable to both old and new domains; and improving the model's ability to handle latest data. Specifically, we construct 1) a domain-incremental text stream that consists of academic papers published in four research fields, and 2) a temporal tweet stream that consists of tweets collected from four different years. By conducting systematic experiments on these two data streams, we look to answer a series of analysis questions: 1) whether continual pretraining retains fine-tuning performance over earlier corpora compared to traditional offline pretraining, 2) whether pretraining improves downstream performance on the latest data, and 3) whether pretraining improves temporal generalization where training and evaluation have distribution gaps because of time.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 394,
                        "end": 402,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "To address the research questions above, we conduct a systematic evaluation of existing continual learning (CL) algorithms, spanning over modelexpansion based, memory-based, and distillationbased approaches. Our results show distillationbased approaches are most effective in knowledge retention in the research paper stream, while simultaneously improve adaptation to latest data and temporal generalization in the tweet stream. We believe our problem formulation, evaluation setup, methods and analysis can inspire more future work on continual pretraining of language models.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Here we present the problem formulation for lifelong pretraining of PTLM, provide details about the data stream construction process and downstream tasks, and introduce the evaluation protocol.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Problem Formulation",
                "sec_num": "2"
            },
            {
                "text": "We consider the scenario where one needs to deploy and/or maintain NLP models over a sequence of T data domains. At each time step t the model visits an unlabeled text corpus D t from a domain with a data distribution P (D t ). The data distribution P (D t ) evolves as the time step t, forming a data stream D 1..",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Lifelong Pretraining of PTLMs",
                "sec_num": "2.1"
            },
            {
                "text": "T = {D 1 , D 2 , ...D T }.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Lifelong Pretraining of PTLMs",
                "sec_num": "2.1"
            },
            {
                "text": "In practice, the data domain shift can refer to the topic change of the text content (from computer science research papers to biomedical papers), or temporal evolution of the text (from past to recent tweets). The task of lifelong pretraining of PTLM looks to continuously adapt a language model f as the model visits (unlabeled) text corpus D t from the data stream D 1..T , in order to provide a good model initialization for fine-tuning on downstream tasks from the same domain. With slight abuse in notations, we also use D t to directly refer to a data domain.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Lifelong Pretraining of PTLMs",
                "sec_num": "2.1"
            },
            {
                "text": "Here, we assume a language model f is updated sequentially over each pretraining corpora D t , without accessing the full earlier corpora {D i } i<t in the data stream D 1..T . This aims to capture practical constraints such as privacy restriction for storing earlier data, or computation budget for training over all the text corpora in D 1..T . We use f t to denote the language model right after updating on the domain D t . In our study, f is a RoBERTa-base transformer (Liu et al., 2019b) and the model (f 0 ) is initialized with pretrained RoBERTa weights.",
                "cite_spans": [
                    {
                        "start": 474,
                        "end": 493,
                        "text": "(Liu et al., 2019b)",
                        "ref_id": "BIBREF32"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Lifelong Pretraining of PTLMs",
                "sec_num": "2.1"
            },
            {
                "text": "The utility of the PTLMs {f t } is evaluated based on their fine-tuned model performance on various downstream tasks. After updating on a domain D i , the model f i can be fine-tuned over downstream tasks from visited domains D t where t \u2264 i. We note the set of downstream tasks related to domain D t as S t = {S j t } Nt j=1 , assuming the number of downstream tasks is N t . Note that in the finetuning stage, model f t has no access to any of the pretraining corpus D 1..T .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Lifelong Pretraining of PTLMs",
                "sec_num": "2.1"
            },
            {
                "text": "\u2026 ! \" # # ! Knowledge Retention Train \" Old Old Test \u2026 $ Fine-Tuning & Evaluation Train Latest Latest Test Adaptation to Latest Data Temporal Generalization \" Train \" ! Old Latest Test ! ! Figure 2:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Continual Pretraining",
                "sec_num": null
            },
            {
                "text": "Training, evaluation setups, and metrics of lifelong language model pretraining. The model sequentially visits each corpus, and is fine-tuned on downstream datasets related to the domains of pretraining. We evaluate knowledge retention and adaptation to new data with downstream fine-tuning performance on old and latest domains respectively. Besides, we evaluate temporal generalization where training/test examples are drawn from different time steps.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Continual Pretraining",
                "sec_num": null
            },
            {
                "text": "We construct data streams to simulate two representative scenarios of data domain shifts in practice (also see Fig. 1 ): one domain-incremental stream to simulate the sequential changes of research paper areas; and one chronologically-ordered stream to simulate tweets emerging over time.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 111,
                        "end": 117,
                        "text": "Fig. 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Data Streams & Downstream Datasets",
                "sec_num": "2.2"
            },
            {
                "text": "Domain-incremental Paper Stream. This paper stream consists of the full text of research papers published in four research areas: biomedical, computer science, material science, and physics, filtered from the S2ORC dataset 1 , which are presented sequentially to the model. For each domain, we evaluate downstream performance over two datasets. The downstream tasks span over various tasks such as relation extraction and named entity recognition, and are summarized in Table 1 . We detail these datasets in Appendix D.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 470,
                        "end": 477,
                        "text": "Table 1",
                        "ref_id": "TABREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Data Streams & Downstream Datasets",
                "sec_num": "2.2"
            },
            {
                "text": "Chronologically-ordered Tweet Stream. This tweet data stream consists of tweets from the year 2014, 2016, 2018 and 2020, collected by the Archive Team 2 and preprocessed following Nguyen et al. (2020). These four tweet corpora are presented sequentially to the language model following the chronological order of the tweet year. For downstream tasks, we hold out 1M tweets from each year's corpus to construct multi-label hashtag prediction datasets (Gong and Zhang, 2016) and single-label emoji prediction datasets (Barbieri et al., 2018). On two datasets, we report label ranking average precision scores (a multi-label version of MRR) of models (Azeemi and Waheed, 2021) and Macro-F1 respectively. The detailed dataset construction process is included in Appendix D.",
                "cite_spans": [
                    {
                        "start": 450,
                        "end": 472,
                        "text": "(Gong and Zhang, 2016)",
                        "ref_id": "BIBREF16"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data Streams & Downstream Datasets",
                "sec_num": "2.2"
            },
            {
                "text": "We consider three key aspects for evaluating the utility of the language models {f t } that are continuously updated over the data stream D 1..T , also illustrated in Figure 2 : 1) knowledge retention and transfer over the pretraining corpora seen earlier; 2) adaptation to the latest data domain, and 3) temporal generalization when training and evaluation data are from different time steps.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 167,
                        "end": 175,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Evaluation Protocol",
                "sec_num": "2.3"
            },
            {
                "text": "Knowledge Retention. A key utility of continual language model pretraining is to obtain a single model applicable to all domains. We focus on the evaluation of the ability with the domainincremental paper stream, because for the tweet stream, the practical need of performance over outdated data is limited. Knowledge retention is measured with the downstream task performance from earlier or the current domains that the pretrained model has visited. More formally, for each pretrained model checkpoint in {f i }, we fine-tune f i over downstream tasks {S t } where t \u2264 i and evaluate the corresponding test set performance. It is important that the models do not suffer from catastrophic forgetting (Robins, 1995) , i.e., significantly reduced helpfulness when f i is fine-tuned for downstream tasks S t from earlier domains with t < i.",
                "cite_spans": [
                    {
                        "start": 701,
                        "end": 715,
                        "text": "(Robins, 1995)",
                        "ref_id": "BIBREF45"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Protocol",
                "sec_num": "2.3"
            },
            {
                "text": "Adaption to Latest Data Domain. In certain scenarios, performance of downstream models over the latest data domain should be emphasized. For example, classifiers in the tweet domain are usually trained and evaluated with up-to-date data for practical deployment. Formally, we focus on the downstream task performance of models fine-tuned from the final pretrained model checkpoint f T , where the downstream tasks S T are also from the latest domain. To succeed in these metrics, it is crucial for the model to transfer knowledge from earlier domains to the latest domain.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Protocol",
                "sec_num": "2.3"
            },
            {
                "text": "Temporal Generalization Ability. We consider another practical fine-tuning scenario in the tweet stream where the model is trained on outdated data and evaluated on the latest data (Rijhwani and Preotiuc-Pietro, 2020; Huang and Paul, 2018) , referred to as the temporal generalization ability. Formally, we fine-tune the final pretrained model checkpoint f T over the training set of downstream tasks S t from an earlier time step (t < T ), and evaluate on the test set of the downstream tasks S T from the latest time step T .",
                "cite_spans": [
                    {
                        "start": 181,
                        "end": 217,
                        "text": "(Rijhwani and Preotiuc-Pietro, 2020;",
                        "ref_id": "BIBREF44"
                    },
                    {
                        "start": 218,
                        "end": 239,
                        "text": "Huang and Paul, 2018)",
                        "ref_id": "BIBREF23"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Evaluation Protocol",
                "sec_num": "2.3"
            },
            {
                "text": "Lifelong language model pretraining introduces novel challenges because of the large training sets and more comprehensive evaluation protocols compared to classification tasks. We establish several strong baselines, and evaluate the performance of continual learning algorithms from different categories spanning over model-expansion, memorybased, and distillation-based approaches, We illustrate the approaches in Figure 3 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 415,
                        "end": 423,
                        "text": "Figure 3",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Methods",
                "sec_num": "3"
            },
            {
                "text": "We consider several simple baselines which continual learning algorithms will be compared against. RoBERTa-base (f 0 ) corresponds to not pretraining on any of the domain-specific corpora.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Simple Baselines",
                "sec_num": "3.1"
            },
            {
                "text": "By separately pretraining f 0 on each corpus",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Simple Baselines",
                "sec_num": "3.1"
            },
            {
                "text": "D 1 , D 2 , ...D T , we obtain T Task-Specific pretrained models.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Simple Baselines",
                "sec_num": "3.1"
            },
            {
                "text": "We also pretrain f 0 sequentially over D 1..T , which we refer to as sequential pretraining. While it allows knowledge transfer between domains compared to domain-specific models, without any continual learning algorithms, sequential pretraining is prone to catastrophic forgetting (Robins, 1995) . Finally, we randomly shuffle corpora from all domains D 1..T before pretraining, noted as Multi-Task Learning (MTL). MTL corresponds to an offline training paradigm that models new corpora by re-training over all corpora seen before. The drawback is that it requires storing full data from earlier domains, and that it can be extremely costly to repetitively retrain over earlier data if new data keeps emerging.",
                "cite_spans": [
                    {
                        "start": 282,
                        "end": 296,
                        "text": "(Robins, 1995)",
                        "ref_id": "BIBREF45"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Simple Baselines",
                "sec_num": "3.1"
            },
            {
                "text": "! \"#$ ! \"#% ! \" ! !&% Layer Layer + 1 Replay memory ' ( \" Memory Stream Replay memory \"#% { ( , ' } \" \"#% \" \u210e \"#% \u210e \" Logit Distillation Rep. Distillation \"#% \"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Distillation + CL Adapters",
                "sec_num": null
            },
            {
                "text": "Contrast & SEED DistillationS imilarity matrix ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Distillation + CL Adapters",
                "sec_num": null
            },
            {
                "text": "Regularization-based Methods We first introduce model-expansion based approaches, which add small trainable modules (e.g., multi-layer perceptron) to the model per new domain while keeping other parts of the model frozen. The Adapter approach is a representative approach that learns a set of \"adapter\" layers g t = {g k t } K k=1 for each domain D t and each of the K transformer layers (Houlsby et al., 2019) . We also experiment with a simple Layer Expansion approach, which learns separate top two layers of the transformer and the prediction head for each domain. We also involve a regularization-based continual learning baseline, online EWC (Schwarz et al., 2018) , which directly penalize change of model parameters.",
                "cite_spans": [
                    {
                        "start": 388,
                        "end": 410,
                        "text": "(Houlsby et al., 2019)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 648,
                        "end": 670,
                        "text": "(Schwarz et al., 2018)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Model-expansion and",
                "sec_num": "3.2"
            },
            {
                "text": "We also experiment with Experience Replay (ER) (Chaudhry et al., 2019) , which alleviates forgetting by storing a subset of earlier examples and periodically re-training (replaying) over them. We maintain a fixed-size memory M (100k examples by default) and populate the memory M each time pretraining on a domain D t finishes with examples in the current domain. We ensure M always contains a balanced sample of examples from all seen domains D 1..t . We replay a mini-batch of examples from the memory every 10 training steps.",
                "cite_spans": [
                    {
                        "start": 47,
                        "end": 70,
                        "text": "(Chaudhry et al., 2019)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Memory Replay Methods",
                "sec_num": "3.3"
            },
            {
                "text": "While knowledge distillation (KD) (Hinton et al., 2015) techniques have been studied intensively for pretrained language models (Sun et al., 2019) , applying them to continual learning has been underexplored outside image classification tasks (Li and Hoiem, 2018; Rebuffi et al., 2017; Hou et al., 2018) . Distillation based CL approaches store one previ-ous model checkpoint of the model (noted as f t\u22121 ) and regularize the differences between f t\u22121 and the current model f t . We adapt several existing knowledge distillation techniques to PTLMs and utilize them for continual learning. We note, while individual distillation techniques are not original, their adaptation to CL algorithms can be novel.",
                "cite_spans": [
                    {
                        "start": 34,
                        "end": 55,
                        "text": "(Hinton et al., 2015)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 128,
                        "end": 146,
                        "text": "(Sun et al., 2019)",
                        "ref_id": "BIBREF51"
                    },
                    {
                        "start": 243,
                        "end": 263,
                        "text": "(Li and Hoiem, 2018;",
                        "ref_id": "BIBREF30"
                    },
                    {
                        "start": 264,
                        "end": 285,
                        "text": "Rebuffi et al., 2017;",
                        "ref_id": "BIBREF43"
                    },
                    {
                        "start": 286,
                        "end": 303,
                        "text": "Hou et al., 2018)",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Distillation-based CL Methods",
                "sec_num": "3.4"
            },
            {
                "text": "We perform distillation with examples from the current domain D t and a replay memory M (similar to ER). Despite the potential gap between D t and the training data of f t\u22121 , the approach allows utilizing more data for distillation. Formally, each time the model receives a mini-batch of stream examples x s or a draws mini-batch of memory examples x m from M (both noted as x), we collect certain outputs of the model (e.g., output logits or intermediate representations) with f t\u22121 and f t . We compute a distillation loss KD (x, f t\u22121 , f t ) that penalizes the differences between the model outputs, and jointly optimize it with the masked language modeling loss MLM . The final objective is written as = MLM + \u03b1 KD , where \u03b1 is a hyperparameter to weight the distillation loss.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Distillation-based CL Methods",
                "sec_num": "3.4"
            },
            {
                "text": "Logit Distillation. In logit distillation (Hinton et al., 2015) , we collect the output logits of f t and f t\u22121 , noted as y t and y t\u22121 respectively. The distillation loss is computed as D KL (y t , y t\u22121 ), where D KL is the Kullback-Leibler divergence function.",
                "cite_spans": [
                    {
                        "start": 42,
                        "end": 63,
                        "text": "(Hinton et al., 2015)",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Distillation-based CL Methods",
                "sec_num": "3.4"
            },
            {
                "text": "Representation Distillation. We also consider minimizing the representational deviation of sentences between previous and current models (Sun et al., 2019; Jiao et al., 2020) . We extract the representation of each word of two models, noted as h 1:N t\u22121 and h 1:N t , before the masked language modeling prediction head, where N is the length of the sentence. Then, we compute MSE loss ||h 1:N t\u22121 \u2212 h 1:N t || 2 2 as the distillation loss. Contrastive Distillation. In addition to output logits and hidden representations, we further look into representational similarity within a batch of examples as additional knowledge to distill. The approach is adapted from (Cha et al., 2021) , which is originally studied for supervised image classification tasks. We briefly introduce the adapted algorithm and leave the details in Appendix E. During continual pretraining, in addition to the language model pretraining objective, we add an unsupervised contrastive learning objective, namely the SimCSE (Gao et al., 2021) objective to encourage sentence representations to reflect semantic simi-larities between sentences. Then, we compute the intra-batch representational similarity matrices of sentence representations (i.e. between each pair of examples in the mini-batch) with f t\u22121 and f t , noted as B t\u22121 and B t , and minimize the cross entropy",
                "cite_spans": [
                    {
                        "start": 137,
                        "end": 155,
                        "text": "(Sun et al., 2019;",
                        "ref_id": "BIBREF51"
                    },
                    {
                        "start": 156,
                        "end": 174,
                        "text": "Jiao et al., 2020)",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 665,
                        "end": 683,
                        "text": "(Cha et al., 2021)",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 997,
                        "end": 1015,
                        "text": "(Gao et al., 2021)",
                        "ref_id": "BIBREF15"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Distillation-based CL Methods",
                "sec_num": "3.4"
            },
            {
                "text": "loss distill = \u2212 1 N N i=1 N j=1 B t\u22121 ij log B t ij",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Distillation-based CL Methods",
                "sec_num": "3.4"
            },
            {
                "text": "Self-Supervised Distillation (SEED). SEED distillation proposed by (Fang et al., 2021 ) has a similar spirit as the contrastive distillation. The only difference is that it distills representational similarity between the batch and a large set of other examples. We leave the details of the algorithm in Appendix E. We further combine SEED Distillationwith logit distillation and refer to the approach as SEED-Logit Distillation.",
                "cite_spans": [
                    {
                        "start": 67,
                        "end": 85,
                        "text": "(Fang et al., 2021",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Distillation-based CL Methods",
                "sec_num": "3.4"
            },
            {
                "text": "We summarize our findings over the created data streams. We ask whether lifelong pretraining and continual learning algorthms are effective base on our evaluation protocol proposed in Sec. 2.3.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Results",
                "sec_num": "4"
            },
            {
                "text": "We use the RoBERTa-base model (Liu et al., 2019b) , initialized with RoBERTa-base weights throughout the experiments. We set the maximal sequence length to 128 and an effective training batch size of 2,048. On the research paper stream, models are trained for 8k steps in the first domain and 4k steps in the subsequent domains. On the Tweet stream, we train the models for 4k steps in each domain. These correspond to less than a single pass of data in each domain. See Appendix A for detailed setups.",
                "cite_spans": [
                    {
                        "start": 30,
                        "end": 49,
                        "text": "(Liu et al., 2019b)",
                        "ref_id": "BIBREF32"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiment Settings",
                "sec_num": "4.1"
            },
            {
                "text": "As we introduced in Sec. 2.2, in the domain incremental research paper stream, we expect a model f t to perform well on all downstream tasks S 1..t from domains D 1..t . In Table 2 , we report the performance of models on all downstream tasks S 1..T fine-tuned from the final pretraining checkpoint, f T . We visualize more complete change of downstream task performance over different time steps of pretraining (i.e.,, f 1 , f 2 , f 3 , f 4 ) in Fig. 4 . We also report the log perplexity of masked language modeling (MLM) in Table 2 as additional information. With these results, we address the research questions below. Figure 4 , we see the performance of Sequential Pretraining on Chemprot and RCT (from D 1 ) drops significantly from t = 1 to 4. The results imply lifelong pretraining allows later domains to benefit from knowledge transfer from earlier domains, but the performance on earlier domains is limited because of forgetting.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 173,
                        "end": 180,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 447,
                        "end": 453,
                        "text": "Fig. 4",
                        "ref_id": "FIGREF1"
                    },
                    {
                        "start": 527,
                        "end": 534,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 623,
                        "end": 631,
                        "text": "Figure 4",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Domain Incremental Data Stream",
                "sec_num": "4.2"
            },
            {
                "text": "Does continual learning algorithms help retain knowledge in sequential pretraining? Next, we compare different kinds of CL algorithms and investigate the effect of CL algorithms in alleviating forgetting and improving knowledge transfer. Table 2 shows that Online-EWC slightly improves MLM perplexity compared to Sequential PT, but brings no improvement to the fine-tuning performance. We hypothesize that regularization directly in the parameter space as in Online-EWC is not effective when the parameter space is very high dimensional. Layer Expansion approach), likely because a great portion of the model is kept frozen. In contrast, the memory-replay based approach (ER) allows training the full parameters of the model and has been shown to be highly effective in continual learning of classification tasks (Wang et al., 2019; Chaudhry et al., 2019) . However, we surprisingly find that ER could hardly improve over Sequential Pretraining except D 1 . A similar pattern can be found in the MLM perplexity. We hypothesize that the positive effect of example replay has diminished because of the overfitting to the memory examples. Table 3 summarizes the effect of tuning hyperpameters in ER. When we reduce the frequency of replay (from every 10 steps to 100 steps), the MLM performance improves, which implies reduced overfitting; however, the performance of downstream task performance does not improve. When we increase the size of the memory |M | from 100k to 10M , the MLM perplexity also improves; still, there are still no improvements in downstream tasks. It may imply ER itself is not an effective approach for continual pretraining.",
                "cite_spans": [
                    {
                        "start": 813,
                        "end": 832,
                        "text": "(Wang et al., 2019;",
                        "ref_id": "BIBREF53"
                    },
                    {
                        "start": 833,
                        "end": 855,
                        "text": "Chaudhry et al., 2019)",
                        "ref_id": "BIBREF8"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 238,
                        "end": 245,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 1136,
                        "end": 1143,
                        "text": "Table 3",
                        "ref_id": "TABREF4"
                    }
                ],
                "eq_spans": [],
                "section": "Domain Incremental Data Stream",
                "sec_num": "4.2"
            },
            {
                "text": "Unlike ER, distillation approaches utilize richer information such as output logits or representation similarity to preserve past knowledge. We find either Logit KD or SEED-Logit KD to be most effective depending on the task, while Models are fine-tuned from checkpoints of lifelong pretrained LMs at different time steps t. For Chemprot and RCT-Sample from D1, we use t \u2208 {1, 2, 3, 4}; while for ACL-ARC and SciERC from D2, t \u2208 {2, 3, 4}. Methods achieving the best performance at the end of training (t = 4) is highlighted.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Domain Incremental Data Stream",
                "sec_num": "4.2"
            },
            {
                "text": "D 1 , D 2 at least by 1.0%. However, performance on D 3 , D 4 , which come later in the data stream, does not improve over Sequential Pretraining, possibly because the distillation loss term makes the model rigid in obtaining new knowledge.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Domain Incremental Data Stream",
                "sec_num": "4.2"
            },
            {
                "text": "What is the gap between lifelong pretraining and multi-task learning across all the domains? Multi-Task Learning refers to the offline training paradigm, which retrain PTLMs over all corpora (D 1..t ) each time a new corpus D t becomes available. We examine whether lifelong pretraining is comparable to multi-task pretraining in terms of performance. From Table 2 and Figure 4 , we see Sequential Pretraining in general underperforms MTL except for the final domain. However, certain CL approaches, such as Logit-Distillation, could improve over MTL on all downstream tasks from the first and the second domain. We speculate the reason is that continual learning naturally provides a curriculum (Xu et al., 2020; Shi et al., 2015) to models where each individual task is easier to learn. The results have a positive implication that lifelong pretraining is not only more computationally efficient and requires less storage of past data, but may also improve the performance of pretraining.",
                "cite_spans": [
                    {
                        "start": 696,
                        "end": 713,
                        "text": "(Xu et al., 2020;",
                        "ref_id": "BIBREF55"
                    },
                    {
                        "start": 714,
                        "end": 731,
                        "text": "Shi et al., 2015)",
                        "ref_id": "BIBREF49"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 357,
                        "end": 364,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 369,
                        "end": 377,
                        "text": "Figure 4",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Domain Incremental Data Stream",
                "sec_num": "4.2"
            },
            {
                "text": "Does lifelong pretraining make models more data efficient? In Table 5 , we further examine the performance of final pretrained models under different amounts of training examples. We include full results in Appendix B. We find in general, performance improvements are more significant in the low-resource setup.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 62,
                        "end": 69,
                        "text": "Table 5",
                        "ref_id": "TABREF11"
                    }
                ],
                "eq_spans": [],
                "section": "Domain Incremental Data Stream",
                "sec_num": "4.2"
            },
            {
                "text": "Computational Costs. We analyze computational costs of different CL algorithms and present additional experiments with controlled computational costs. We find additional computational cost is necessary for performance improvement of distillation-based CL. However, it is not possible to trade performance simply by investing more computation budget with arbitrary CL algorithms. We leave detailed discussions in Appendix F.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Domain Incremental Data Stream",
                "sec_num": "4.2"
            },
            {
                "text": "We conduct analysis on pretraining PTLM on chronologically-ordered tweet corpora, to understand whether lifelong pretraining helps adaptation to the latest data and improves temporal generalization ability. The results are summarized in Table 4 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 237,
                        "end": 244,
                        "text": "Table 4",
                        "ref_id": "TABREF7"
                    }
                ],
                "eq_spans": [],
                "section": "Temporal Data Stream",
                "sec_num": "4.3"
            },
            {
                "text": "Will LMs be outdated? We compare the performance of Task-Specific (2014) to the Task-Specific models pretrained on the year of downstream datasets (noted as Task-Specific (Latest)) and notice consistent improvements in downstream tasks in 2018 and 2020 (first two columns in Table 4 ). Sequential Pretraining could also outperform the Task-Specific (2014) model. It verifies that language models may get outdated, which can be addressed by task-specific or lifelong pretraining over the latest corpora. later data (D 3 , D 4 ) can be improved over Task-Specific models when continual learning algorithms are applied. From the first two columns of Table 4 , we see Logit-KD and SEED-KD improve Hashtag prediction score over data of years 2018 and 2020. SEED-Logit KD further improves prediction F1 on Emoji prediction. Note that these findings are in contrast to the research paper stream, where CL algorithms do not improve performance in the latest domain D 4 . The reason can be the higher similarity between domains in the tweet corpora making the knowledge transfer easier, which is further discussed in Appendix H.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 275,
                        "end": 282,
                        "text": "Table 4",
                        "ref_id": "TABREF7"
                    },
                    {
                        "start": 647,
                        "end": 654,
                        "text": "Table 4",
                        "ref_id": "TABREF7"
                    }
                ],
                "eq_spans": [],
                "section": "Temporal Data Stream",
                "sec_num": "4.3"
            },
            {
                "text": "Does lifelong pretraining improve temporal generalization? Temporal generalization evaluates downstream performance over latest test data when fine-tuned over outdated training data. We show lifelong pretraining brings clear improvement to temporal generalization. From Table 4 , we see even Sequential Pretraining could improve over the model pretrained merely on the year 2020 data (Task-Specific (2020)) consistently. We find performance further improves with CL algorithms applied. SEED-Logit-KD performs best in general on crossyear hashtag prediction tasks. In crossyear emoji prediction, we find Contrast-KD and SEED-KD perform best. We also find that SEED-Logit-KD could slightly outperform Logit-KD. (Lazaridou et al., 2021) while specifically focusing on factual knowledge (Dhingra et al., 2021; Jang et al., 2021) .",
                "cite_spans": [
                    {
                        "start": 709,
                        "end": 733,
                        "text": "(Lazaridou et al., 2021)",
                        "ref_id": null
                    },
                    {
                        "start": 783,
                        "end": 805,
                        "text": "(Dhingra et al., 2021;",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 806,
                        "end": 824,
                        "text": "Jang et al., 2021)",
                        "ref_id": "BIBREF25"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 270,
                        "end": 277,
                        "text": "Table 4",
                        "ref_id": "TABREF7"
                    }
                ],
                "eq_spans": [],
                "section": "Temporal Data Stream",
                "sec_num": "4.3"
            },
            {
                "text": "Continual Learning Algorithms in NLP. Continual learning in NLP has mainly been studied for classification tasks. An effective approach is to utilize a number of stored past examples (de Masson d'Autume et al., 2019; , or pseudo examples (e.g., the ones generated with a PTLM Kanwatchara et al., 2021) ). Recent extensions of the algorithm (Chuang et al., 2020) perform knowledge distillation with generated pseudo examples. Other lines of works focus on regularization over the sentence representations (Wang et al., 2019; Liu et al., 2019a) or directly merging models in the parameter space (Matena and Raffel, 2021) . Model expansion-based approaches (Liu et al., 2019a; Pfeiffer et al., 2021) , including learning domain specific expert models , are also actively studied.",
                "cite_spans": [
                    {
                        "start": 187,
                        "end": 216,
                        "text": "Masson d'Autume et al., 2019;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 276,
                        "end": 301,
                        "text": "Kanwatchara et al., 2021)",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 340,
                        "end": 361,
                        "text": "(Chuang et al., 2020)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 504,
                        "end": 523,
                        "text": "(Wang et al., 2019;",
                        "ref_id": "BIBREF53"
                    },
                    {
                        "start": 524,
                        "end": 542,
                        "text": "Liu et al., 2019a)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 593,
                        "end": 618,
                        "text": "(Matena and Raffel, 2021)",
                        "ref_id": "BIBREF37"
                    },
                    {
                        "start": 654,
                        "end": 673,
                        "text": "(Liu et al., 2019a;",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 674,
                        "end": 696,
                        "text": "Pfeiffer et al., 2021)",
                        "ref_id": "BIBREF42"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Temporal Data Stream",
                "sec_num": "4.3"
            },
            {
                "text": "In this paper, we formulated the lifelong language model pretraining problem and constructed two data streams associated with downstream datasets. We evaluated knowledge retention, adaptation to the latest data, and temporal generalization ability of continually pretrained language models. Our experiments show distillation-based approaches being most effective in these evaluation setups. A limitation of the work is that it has not been fully addressed whether there exists a variant of distillation-based CL approach that consistently outperforms Logit-KD. Based on the current observation, we conclude the performance of different KD approaches for CL is highly task-dependent. It asks for more future works into continual learning algorithms within the proposed problem setup. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "6"
            },
            {
                "text": "We use a linearly decreasing learning rate initialized with 5e-4 on the research paper stream and 3e-4 on the tweet stream. On the research paper stream, we train the model for 8,000 steps in the first task, and 4,000 steps in the subsequent tasks. On the tweet stream, we train the model for 8,000 steps in all tasks. We hold out 128,000 sentences from each corpus to evaluate MLM performance.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Detailed Experiment Settings",
                "sec_num": null
            },
            {
                "text": "As the size of pretraining corpora is large, during training, each training example is visited only once. We use the masked language modeling perplexity over held-out validation sets of the pretraining corpora as the metrics for hyperparameter tuning. Common hyperparameters such as learning rate and batch sizes are tuned with Task-specific models with the first task. Hyperparameters that are specific to continual learning algorithms, such as the scale of the distillation loss, is tuned using the first two domains in the stream according to the MLM performance over validation sets. The weight of the distillation term \u03b1 is set as 1.0 for logit distillation and 0.1 for other distillation algorithms. By default, we replay or perform distillation with a mini-batch of examples from the replay memory every 10 training steps in ER and Distillationbased CL approaches. We use the huggingface transformers library https://github.com/ huggingface/transformers for implementation.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Detailed Experiment Settings",
                "sec_num": null
            },
            {
                "text": "B Low-Resource Fine-Tuning Figure 6 summarizes the performance of fine-tuned models from the final model checkpoint (t = 4) Emoji prediction, fine-tuned from the pre-trained model in the final time step.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 27,
                        "end": 35,
                        "text": "Figure 6",
                        "ref_id": "FIGREF3"
                    }
                ],
                "eq_spans": [],
                "section": "A Detailed Experiment Settings",
                "sec_num": null
            },
            {
                "text": "using different amount of downstream training examples. We see on Chemprot and SciERC, the benefit of Sequential Pretraining over RoBERTa-base is more significant in low-resource fine-tuning setups. Whenever Seqential Pretraining outperforms RoBERTa-base, we notice Logit-KD could further improve over Sequential Pretraining.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Detailed Experiment Settings",
                "sec_num": null
            },
            {
                "text": "Tables 5 and 6 summarize full results over the Tweet stream. Compared to the table 4 in the main text, we add downstream performance over data from years 2014 and 2016 (D 1 , D 2 ), and temporal generalization from year 2014 to 2020 (D 1 \u2192 D 4 ).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C Full Results over the Tweet Stream",
                "sec_num": null
            },
            {
                "text": "The research paper stream consists of full text of 6.6M, 12.1M, 7.8M, and 7.5M research papers from the S2ORC dataset. We evaluate downstream fine-tuning performance on two in-domain datasets for each research area: Chemprot relation exaction dataset (Vindahl, 2016) and RCT abstract sentence role labeling dataset (Dernoncourt and Lee, 2017) for the bio-medical domain; ACL-ARC citation intent classification dataset (Jurgens et al., 2018) and SciERC relation extraction dataset (Luan et al., 2018) for the Q to cache examples from the current domain D t . Given a mini-batch of training examples x, it computes cosine similarity between each pair of examples within the batch x and Q with f t\u22121 and f t , resulting in two similarity matrices B t\u22121 , P y \u2208 R |B|\u00d7|Q| . Similar to the contrastive distillation, the distillation loss is the cross-entropy between two similarity matrices B t\u22121 and B t computed in the same way as Eq. 2.",
                "cite_spans": [
                    {
                        "start": 251,
                        "end": 266,
                        "text": "(Vindahl, 2016)",
                        "ref_id": "BIBREF52"
                    },
                    {
                        "start": 315,
                        "end": 342,
                        "text": "(Dernoncourt and Lee, 2017)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 418,
                        "end": 440,
                        "text": "(Jurgens et al., 2018)",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 480,
                        "end": 499,
                        "text": "(Luan et al., 2018)",
                        "ref_id": "BIBREF34"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "D Dataset Details",
                "sec_num": null
            },
            {
                "text": "Computational cost is a crucial matter for online continual learning systems. In this section, we analyze the computational costs of continual learning algorithms and perform controlled experiments of computational costs. We quantify computational costs with the total number of forward (C f ) and backward (C b ) computations (C = C f +C b ) over the PTLMs, which is easy to control; in practice, we find the wall clock time of training was approximately linear to C. We summarize the number of forward and backward passes and the wall clock time of training in Table 7 . In the visit of b batches from the training stream, Sequential PT performs b forward and backward passes respectively over the PTLM, resulting in C = 2b. Experience replay further replays 1 batch of examples every r steps over the training stream, which results in C = (2 + 2/k)b. In our main experiments, r is set to 10 (Sec. 3.3). Logit-Distill and Rep-Distill require one additional forward pass over a frozen PTLM to compute the target of distillation, resulting in C = (3 + 3/k)b. Distillation algorithms that perform contrastive learning with SimCSE (i.e. SEED-Distill and SEED-Logit-Distill) additionally require one forward and backward pass using the same batch of examples with different dropout masks. Therefore, for SEED-Logit-Distill, C = (5 + 5/k)b.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 563,
                        "end": 570,
                        "text": "Table 7",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "F Analysis and Controlled Experiments of Computational Costs",
                "sec_num": null
            },
            {
                "text": "To control the number of forward and backward passes, we present approaches to compensate the lower computation costs compared to Distillation algorithms and one approach to shrink the computational cost of distillation algorithms: (1) for Sequential PT, we train the models for 1.2 times more steps so that C = 2.4b, noted as Sequential PT b =1.2b ; (2) for ER, we increase the replay frequency k to 5 from the default setup 10, so that C = 2.4b. We also decrease the cost of Logit-KD and SEED-Logit-KD by reducing the frequency of distillation from every 1 batch to every r =10 steps, while still replaying and distilling knowledge over 1 batch of memory examples every 10 training steps. This results in C f = (1 + 2/k + 1/k )b and C b = (1 + 1/k)b, where C = 2.4b when both r and r are 10. The approach is referred to as Sparse Logit-KD. Finally, for SEED-Logit-KD, we remove the SimCSE loss from training and perform sparse distillation similar to Sparse-Logit-KD, which also results in C = 2.4b.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "F Analysis and Controlled Experiments of Computational Costs",
                "sec_num": null
            },
            {
                "text": "The performance of the models is presented in Table 8 . We notice that at the end of pretraining, increasing the number of training steps in Sequential PT by 1.2 times does not lead to performance boost on the latest domain (D 4 ), while the performance over tasks from earlier domains (Chemprot, ACL-ARC, SciERC) slightly dropped, possibly due to increased forgetting. For ER, we notice replaying only slightly more frequently (ER k=5 ) than the default setup (k=10) greatly increased the perplexity of MLM, implying significantly increased overfitting to the memory; while the performance differences of downstream tasks compared to the default ER is mixed. When we decrease the replay frequency of distillation, the performance on Logit-KD and SEED-KD also decreased and does not outperform ER.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 46,
                        "end": 53,
                        "text": "Table 8",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "F Analysis and Controlled Experiments of Computational Costs",
                "sec_num": null
            },
            {
                "text": "The results show additional computation costs can be necessary for continual learning algorithms such as Logit-KD and SEED-Logit-KD. However, the results also show that there is no simple tradeoff between computational cost and performance. We have seen that it is not always beneficial to increase the number of training steps over the emerging data, as it increases forgetting in earlier domains. Similarly, increasing the frequency of replay may lead to significant overfitting to the replay memory. Investigating into more effective continual learning algorithms, despite increased computation costs, allows us to obtain performance improvement that cannot be simply traded with more computation with arbitrary continual learning algorithms. We leave more thorough studies into this topic as future work.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "F Analysis and Controlled Experiments of Computational Costs",
                "sec_num": null
            },
            {
                "text": "In this section, we present an additional set of experiments on BERT-base (Devlin et al., 2019) model, which is originally pretrained with Wikipedia articles before 2019, with Tweets only after 2019. The research paper stream, we find some domains to be more similar than others. For example, Biomedical (D 1 ) and Material Science domains (D 3 ) have larger similarity in their vocabulary distributions, which explains general downstream performance increase on D 1 after the model is pretrained on D 3 (Fig. 4 (a,b) ). The differences in vocabulary distribution explain inconsistency in results between two data streams, specifically, whether lifelong pretraining improves downstream model performance on the latest domain, as we mentioned in Sec. 4.3. Other than this, our main findings, such as the effect of distillation-based CL algorithms on reducing forgetting, are consistent over two datasets with such significant differences in their changes of vocabulary distribution. We believe it implies the conclusions in this paper should be reliable in diverse data streams.",
                "cite_spans": [
                    {
                        "start": 74,
                        "end": 95,
                        "text": "(Devlin et al., 2019)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 504,
                        "end": 517,
                        "text": "(Fig. 4 (a,b)",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "G Experiments with BERT on Tweet Stream After 2019",
                "sec_num": null
            },
            {
                "text": "We would like to note that, in practice, continually pretrained models over real-world data streams would require identification and removal of biased contents from pretraining corpora, which may affect the prediction of downstream models. As PTLMs are continuously updated, the bias in earlier pretraining may have a profound negative impact. In future works, it is preferable to develop algorithms to \"forget\" certain biased knowledge from language models. We further note that any data released in this paper, especially the tweet stream, should only be used for research purposes.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "I Ethic Risks",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": " (Augenstein et al., 2017) . We report micro-averaged F1 on Chemprot, RCT, MNER datasets following the evaluation metrics in the original work, and report macroaveraged F1 on all other datasets. We use the official data splits for all datasets except for RCT, where we employ a low-resource training setup following Gururangan et al. (2020) .The pretraining corpora for the tweet stream consist of 25M tweets in each year. For downstream tasks, we use a separate set of 1M tweets from each year to construct multi-label hashtag prediction (Gong and Zhang, 2016) datasets and singlelabel emoji prediction datasets (Barbieri et al., 2018) . We replace user names to special tokens. For Hashtag prediction, the label space consists of tweets containing 200 most frequent hashtags in each year. We independently sample 500 tweets per label (hashtag) as training, validation and test sets, which results 10k examples in each of the data splits. For emoji prediction, we construct 20way single-label emoji prediction datasets for each year following Barbieri et al. (2018) with the 1M held out tweets. We sample 5,000 tweets per emoji in each split, resulting in balanced datasets of the same size as the hashtag prediction datasets.",
                "cite_spans": [
                    {
                        "start": 1,
                        "end": 26,
                        "text": "(Augenstein et al., 2017)",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 316,
                        "end": 340,
                        "text": "Gururangan et al. (2020)",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 539,
                        "end": 561,
                        "text": "(Gong and Zhang, 2016)",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 613,
                        "end": 636,
                        "text": "(Barbieri et al., 2018)",
                        "ref_id": null
                    },
                    {
                        "start": 1044,
                        "end": 1066,
                        "text": "Barbieri et al. (2018)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "annex",
                "sec_num": null
            },
            {
                "text": "During continual pretraining, in addition to the language model pretraining objective, we add a unsupervised contrastive learning objective, namely the SimCSE (Gao et al., 2021) objective, so that the similarity in the sentence representation better reflects the semantic similarity in the sentence. We use the l 2 -normalized representation of the startof-sequence token at the final layer as the sentence representation, noted as h. Then, we distill the intra-batch representational similarity from the previous model f t\u22121 to the current model f t . Given a mini-batch of N examples x, we compute the representational dot-product similarity matrix between normalized sentence representations h between each pair of examples with f t\u22121 and f t , noted as B t\u22121 and B t , where each element B ij is,where \u03c4 is a temperature hyperparameter. We specify a temperature \u03c4 t = 0.05 for the teacher model f t\u22121 and a temperature \u03c4 s for the student model f t = 0.01. We compute the cross-entropy between B t\u22121 and B t as the distillation loss,SEED distillation proposed by (Fang et al., 2021) has a similar spirit as the contrastive distillation with differences in the examples used for computing similarity matrices computes. The algorithm distills representational similarity between the batch and a large set of other examples, maintained in an example queue Q. As the number of target examples K can be much larger than the batch size, it allows distillation of richer information by regularizing similarities. During pretraining, the method maintains a fixed-size queue (1 + 1/k)b (1 + 1/k)b (2 + 2/k)b 2.2b 4.2 \u00d7 10 4 sec.",
                "cite_spans": [
                    {
                        "start": 159,
                        "end": 177,
                        "text": "(Gao et al., 2021)",
                        "ref_id": "BIBREF15"
                    },
                    {
                        "start": 1067,
                        "end": 1086,
                        "text": "(Fang et al., 2021)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "E Details of Continual Learning Algorithms E.1 Contrastive Distillation",
                "sec_num": null
            },
            {
                "text": "(2 + 2/k)b (1 + 1/k)b (3 + 3/k)b 3.3b 6.9 \u00d7 10 4 sec.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Logit-Distill",
                "sec_num": null
            },
            {
                "text": "(3 + 3/k)b (2 + 2/k)b (5 + 5/k)b 5.5b 9.7 \u00d7 10 4 sec. training corpora D 1..4 consist of tweets from the first half of 2019, the second half of 2019, the first half of 2020, and the second half of 2020 respectively. We accordingly construct hashtag prediction and cross-year hashtag prediction datasets. The performance of downstream tasks fine-tuned from the final pretrained model is presented in Table 9 . We see Sequential PT clearly outperforms BERTbase which is not continually pretrained, and that Logit-KD generally improves hashtag prediction performance compared to Sequential PT except on the first half of 2019. We hypothesize the small temporal gap between D 1..4 makes improvements less significant than our main experiment setup. We present temporal generalization performance in cross-year hashtag prediction tasks in Table 10 . Similarly, Logit-KD improves over Sequential PT in two out of three cross-year hashtag prediction setups.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 399,
                        "end": 406,
                        "text": "Table 9",
                        "ref_id": null
                    },
                    {
                        "start": 834,
                        "end": 842,
                        "text": "Table 10",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "SEED-Logit-Distill",
                "sec_num": null
            },
            {
                "text": "In this section, we provide further analysis about the created research paper stream and the tweet stream. We measure cosine distances d v of vocabulary distributions between each pair of different domains (D 1..4 ) and summarize the results in Figure 7 . The results indicate that the Tweet stream has a magnitude smaller vocabulary distribution gap between domains, which is in the scale of 1e \u22125 , compared to the research paper stream, which is in the scale of 1e \u22122 . On the Tweet stream, we see the differences of vocabulary distributions align with the temporal gap between domains. On the",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 245,
                        "end": 253,
                        "text": "Figure 7",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "H Analysis of Data Streams",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "An empirical investigation towards efficient multi-domain language model pre-training",
                "authors": [
                    {
                        "first": "Kristjan",
                        "middle": [],
                        "last": "Arumae",
                        "suffix": ""
                    },
                    {
                        "first": "Qing",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Parminder",
                        "middle": [],
                        "last": "Bhatia",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "4854--4864",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.emnlp-main.394"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Kristjan Arumae, Qing Sun, and Parminder Bhatia. 2020. An empirical investigation towards effi- cient multi-domain language model pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 4854-4864, Online. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "ScienceIE -extracting keyphrases and relations from scientific publications",
                "authors": [
                    {
                        "first": "Isabelle",
                        "middle": [],
                        "last": "Augenstein",
                        "suffix": ""
                    },
                    {
                        "first": "Mrinal",
                        "middle": [],
                        "last": "Das",
                        "suffix": ""
                    },
                    {
                        "first": "Sebastian",
                        "middle": [],
                        "last": "Riedel",
                        "suffix": ""
                    },
                    {
                        "first": "Lakshmi",
                        "middle": [],
                        "last": "Vikraman",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Mccallum",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017)",
                "volume": "10",
                "issue": "",
                "pages": "546--555",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/S17-2091"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Isabelle Augenstein, Mrinal Das, Sebastian Riedel, Lakshmi Vikraman, and Andrew McCallum. 2017. SemEval 2017 task 10: ScienceIE -extracting keyphrases and relations from scientific publica- tions. In Proceedings of the 11th International Workshop on Semantic Evaluation (SemEval-2017), pages 546-555, Vancouver, Canada. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Covid-19 tweets analysis through transformer language models",
                "authors": [],
                "year": 2021,
                "venue": "ArXiv",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Abdul Hameed Azeemi and Adeel Waheed. 2021. Covid-19 tweets analysis through transformer lan- guage models. ArXiv, abs/2103.00199.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "SemEval 2018 task 2: Multilingual emoji prediction",
                "authors": [
                    {
                        "first": "Francesco",
                        "middle": [],
                        "last": "Ronzano",
                        "suffix": ""
                    },
                    {
                        "first": "Luis",
                        "middle": [],
                        "last": "Espinosa-Anke",
                        "suffix": ""
                    },
                    {
                        "first": "Miguel",
                        "middle": [],
                        "last": "Ballesteros",
                        "suffix": ""
                    },
                    {
                        "first": "Valerio",
                        "middle": [],
                        "last": "Basile",
                        "suffix": ""
                    },
                    {
                        "first": "Viviana",
                        "middle": [],
                        "last": "Patti",
                        "suffix": ""
                    },
                    {
                        "first": "Horacio",
                        "middle": [],
                        "last": "Saggion",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of The 12th International Workshop on Semantic Evaluation",
                "volume": "",
                "issue": "",
                "pages": "24--33",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/S18-1003"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Francesco Ronzano, Luis Espinosa-Anke, Miguel Ballesteros, Valerio Basile, Viviana Patti, and Horacio Saggion. 2018. SemEval 2018 task 2: Multilingual emoji prediction. In Proceedings of The 12th International Workshop on Semantic Evaluation, pages 24-33, New Orleans, Louisiana. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "SciB-ERT: A pretrained language model for scientific text",
                "authors": [
                    {
                        "first": "Iz",
                        "middle": [],
                        "last": "Beltagy",
                        "suffix": ""
                    },
                    {
                        "first": "Kyle",
                        "middle": [],
                        "last": "Lo",
                        "suffix": ""
                    },
                    {
                        "first": "Arman",
                        "middle": [],
                        "last": "Cohan",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
                "volume": "",
                "issue": "",
                "pages": "3615--3620",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D19-1371"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Iz Beltagy, Kyle Lo, and Arman Cohan. 2019. SciB- ERT: A pretrained language model for scientific text. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Lan- guage Processing (EMNLP-IJCNLP), pages 3615- 3620, Hong Kong, China. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Mc-Candlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learners",
                "authors": [
                    {
                        "first": "Tom",
                        "middle": [
                            "B"
                        ],
                        "last": "Brown",
                        "suffix": ""
                    },
                    {
                        "first": "Benjamin",
                        "middle": [],
                        "last": "Mann",
                        "suffix": ""
                    },
                    {
                        "first": "Nick",
                        "middle": [],
                        "last": "Ryder",
                        "suffix": ""
                    },
                    {
                        "first": "Melanie",
                        "middle": [],
                        "last": "Subbiah",
                        "suffix": ""
                    },
                    {
                        "first": "Jared",
                        "middle": [],
                        "last": "Kaplan",
                        "suffix": ""
                    },
                    {
                        "first": "Prafulla",
                        "middle": [],
                        "last": "Dhariwal",
                        "suffix": ""
                    },
                    {
                        "first": "Arvind",
                        "middle": [],
                        "last": "Neelakantan",
                        "suffix": ""
                    },
                    {
                        "first": "Pranav",
                        "middle": [],
                        "last": "Shyam",
                        "suffix": ""
                    },
                    {
                        "first": "Girish",
                        "middle": [],
                        "last": "Sastry",
                        "suffix": ""
                    },
                    {
                        "first": "Amanda",
                        "middle": [],
                        "last": "Askell",
                        "suffix": ""
                    },
                    {
                        "first": "Sandhini",
                        "middle": [],
                        "last": "Agarwal",
                        "suffix": ""
                    },
                    {
                        "first": "Ariel",
                        "middle": [],
                        "last": "Herbert-Voss",
                        "suffix": ""
                    },
                    {
                        "first": "Gretchen",
                        "middle": [],
                        "last": "Krueger",
                        "suffix": ""
                    },
                    {
                        "first": "Tom",
                        "middle": [],
                        "last": "Henighan",
                        "suffix": ""
                    },
                    {
                        "first": "Rewon",
                        "middle": [],
                        "last": "Child",
                        "suffix": ""
                    },
                    {
                        "first": "Aditya",
                        "middle": [],
                        "last": "Ramesh",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [
                            "M"
                        ],
                        "last": "Ziegler",
                        "suffix": ""
                    },
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Clemens",
                        "middle": [],
                        "last": "Winter",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher",
                        "middle": [],
                        "last": "Hesse",
                        "suffix": ""
                    },
                    {
                        "first": "Mark",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Eric",
                        "middle": [],
                        "last": "Sigler",
                        "suffix": ""
                    },
                    {
                        "first": "Mateusz",
                        "middle": [],
                        "last": "Litwin",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam Mc- Candlish, Alec Radford, Ilya Sutskever, and Dario Amodei. 2020. Language models are few-shot learn- ers. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Informa- tion Processing Systems 2020, NeurIPS 2020, De- cember 6-12, 2020, virtual.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "Co2l: Contrastive continual learning. ArXiv",
                "authors": [
                    {
                        "first": "Hyuntak",
                        "middle": [],
                        "last": "Cha",
                        "suffix": ""
                    },
                    {
                        "first": "Jaeho",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Jinwoo",
                        "middle": [],
                        "last": "Shin",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Hyuntak Cha, Jaeho Lee, and Jinwoo Shin. 2021. Co2l: Contrastive continual learning. ArXiv, abs/2106.14413.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "On tiny episodic memories in continual learning",
                "authors": [
                    {
                        "first": "Arslan",
                        "middle": [],
                        "last": "Chaudhry",
                        "suffix": ""
                    },
                    {
                        "first": "Marcus",
                        "middle": [],
                        "last": "Rohrbach",
                        "suffix": ""
                    },
                    {
                        "first": "Mohamed",
                        "middle": [],
                        "last": "Elhoseiny",
                        "suffix": ""
                    },
                    {
                        "first": "Thalaiyasingam",
                        "middle": [],
                        "last": "Ajanthan",
                        "suffix": ""
                    },
                    {
                        "first": "K",
                        "middle": [],
                        "last": "Puneet",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Dokania",
                        "suffix": ""
                    },
                    {
                        "first": "H",
                        "middle": [
                            "S"
                        ],
                        "last": "Philip",
                        "suffix": ""
                    },
                    {
                        "first": "Marc'aurelio",
                        "middle": [],
                        "last": "Torr",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ranzato",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1902.10486"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Arslan Chaudhry, Marcus Rohrbach, Mohamed Elho- seiny, Thalaiyasingam Ajanthan, Puneet K Dokania, Philip HS Torr, and Marc'Aurelio Ranzato. 2019. On tiny episodic memories in continual learning. arXiv preprint arXiv:1902.10486.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Lifelong language knowledge distillation",
                "authors": [
                    {
                        "first": "Yung-Sung",
                        "middle": [],
                        "last": "Chuang",
                        "suffix": ""
                    },
                    {
                        "first": "Shang-Yu",
                        "middle": [],
                        "last": "Su",
                        "suffix": ""
                    },
                    {
                        "first": "Yun-Nung",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "2914--2924",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.emnlp-main.233"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Yung-Sung Chuang, Shang-Yu Su, and Yun-Nung Chen. 2020. Lifelong language knowledge distil- lation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Process- ing (EMNLP), pages 2914-2924, Online. Associa- tion for Computational Linguistics.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Episodic memory in lifelong language learning",
                "authors": [
                    {
                        "first": "Sebastian",
                        "middle": [],
                        "last": "Cyprien De Masson D'autume",
                        "suffix": ""
                    },
                    {
                        "first": "Lingpeng",
                        "middle": [],
                        "last": "Ruder",
                        "suffix": ""
                    },
                    {
                        "first": "Dani",
                        "middle": [],
                        "last": "Kong",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Yogatama",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Advances in Neural Information Processing Systems 32: Annual Conference on Neural Information Processing Systems",
                "volume": "",
                "issue": "",
                "pages": "13122--13131",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Cyprien de Masson d'Autume, Sebastian Ruder, Ling- peng Kong, and Dani Yogatama. 2019. Episodic memory in lifelong language learning. In Advances in Neural Information Processing Systems 32: An- nual Conference on Neural Information Processing Systems 2019, NeurIPS 2019, December 8-14, 2019, Vancouver, BC, Canada, pages 13122-13131.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "PubMed 200k RCT: a dataset for sequential sentence classification in medical abstracts",
                "authors": [
                    {
                        "first": "Franck",
                        "middle": [],
                        "last": "Dernoncourt",
                        "suffix": ""
                    },
                    {
                        "first": "Ji",
                        "middle": [
                            "Young"
                        ],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of the Eighth International Joint Conference on Natural Language Processing",
                "volume": "2",
                "issue": "",
                "pages": "308--313",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Franck Dernoncourt and Ji Young Lee. 2017. PubMed 200k RCT: a dataset for sequential sentence clas- sification in medical abstracts. In Proceedings of the Eighth International Joint Conference on Natu- ral Language Processing (Volume 2: Short Papers), pages 308-313, Taipei, Taiwan. Asian Federation of Natural Language Processing.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "BERT: Pre-training of deep bidirectional transformers for language understanding",
                "authors": [
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Devlin",
                        "suffix": ""
                    },
                    {
                        "first": "Ming-Wei",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "Kenton",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    },
                    {
                        "first": "Kristina",
                        "middle": [],
                        "last": "Toutanova",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "1",
                "issue": "",
                "pages": "4171--4186",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/N19-1423"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of deep bidirectional transformers for language under- standing. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Time-aware language models as temporal knowledge bases",
                "authors": [
                    {
                        "first": "Bhuwan",
                        "middle": [],
                        "last": "Dhingra",
                        "suffix": ""
                    },
                    {
                        "first": "Jeremy",
                        "middle": [
                            "R"
                        ],
                        "last": "Cole",
                        "suffix": ""
                    },
                    {
                        "first": "Julian",
                        "middle": [],
                        "last": "Martin Eisenschlos",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Gillick",
                        "suffix": ""
                    },
                    {
                        "first": "Jacob",
                        "middle": [],
                        "last": "Eisenstein",
                        "suffix": ""
                    },
                    {
                        "first": "William",
                        "middle": [
                            "W"
                        ],
                        "last": "Cohen",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "ArXiv",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bhuwan Dhingra, Jeremy R. Cole, Julian Martin Eisen- schlos, D. Gillick, Jacob Eisenstein, and William W. Cohen. 2021. Time-aware language models as tem- poral knowledge bases. ArXiv, abs/2106.15110.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Seed: Self-supervised distillation for visual representation",
                "authors": [
                    {
                        "first": "Zhiyuan",
                        "middle": [],
                        "last": "Fang",
                        "suffix": ""
                    },
                    {
                        "first": "Jianfeng",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Lijuan",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Yezhou",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Zicheng",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "ArXiv",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhiyuan Fang, Jianfeng Wang, Lijuan Wang, L. Zhang, Yezhou Yang, and Zicheng Liu. 2021. Seed: Self-supervised distillation for visual representation. ArXiv, abs/2101.04731.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Simcse: Simple contrastive learning of sentence embeddings",
                "authors": [
                    {
                        "first": "Tianyu",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    },
                    {
                        "first": "Xingcheng",
                        "middle": [],
                        "last": "Yao",
                        "suffix": ""
                    },
                    {
                        "first": "Danqi",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "ArXiv",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. Simcse: Simple contrastive learning of sentence em- beddings. ArXiv, abs/2104.08821.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Hashtag recommendation using attention-based convolutional neural network",
                "authors": [
                    {
                        "first": "Yuyun",
                        "middle": [],
                        "last": "Gong",
                        "suffix": ""
                    },
                    {
                        "first": "Qi",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence, IJCAI 2016",
                "volume": "",
                "issue": "",
                "pages": "2782--2788",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yuyun Gong and Qi Zhang. 2016. Hashtag recom- mendation using attention-based convolutional neu- ral network. In Proceedings of the Twenty-Fifth International Joint Conference on Artificial Intelli- gence, IJCAI 2016, New York, NY, USA, 9-15 July 2016, pages 2782-2788. IJCAI/AAAI Press.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Demix layers: Disentangling domains for modular language modeling",
                "authors": [
                    {
                        "first": "Suchin",
                        "middle": [],
                        "last": "Gururangan",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Lewis",
                        "suffix": ""
                    },
                    {
                        "first": "Ari",
                        "middle": [],
                        "last": "Holtzman",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Smith",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "ArXiv",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Suchin Gururangan, Michael Lewis, Ari Holtzman, Noah A. Smith, and Luke Zettlemoyer. 2021. Demix layers: Disentangling domains for modular lan- guage modeling. ArXiv, abs/2108.05036.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Don't stop pretraining: Adapt language models to domains and tasks",
                "authors": [
                    {
                        "first": "Ana",
                        "middle": [],
                        "last": "Suchin Gururangan",
                        "suffix": ""
                    },
                    {
                        "first": "Swabha",
                        "middle": [],
                        "last": "Marasovi\u0107",
                        "suffix": ""
                    },
                    {
                        "first": "Kyle",
                        "middle": [],
                        "last": "Swayamdipta",
                        "suffix": ""
                    },
                    {
                        "first": "Iz",
                        "middle": [],
                        "last": "Lo",
                        "suffix": ""
                    },
                    {
                        "first": "Doug",
                        "middle": [],
                        "last": "Beltagy",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Downey",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Smith",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
                "volume": "9",
                "issue": "",
                "pages": "8342--8360",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.acl-main.740"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Suchin Gururangan, Ana Marasovi\u0107, Swabha Swayamdipta, Kyle Lo, Iz Beltagy, Doug Downey, and Noah A. Smith. 2020. Don't stop pretraining: Adapt language models to domains and tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 9 8342-8360, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Distilling the knowledge in a neural network",
                "authors": [
                    {
                        "first": "Geoffrey",
                        "middle": [
                            "E"
                        ],
                        "last": "Hinton",
                        "suffix": ""
                    },
                    {
                        "first": "Oriol",
                        "middle": [],
                        "last": "Vinyals",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Dean",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "ArXiv",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Geoffrey E. Hinton, Oriol Vinyals, and J. Dean. 2015. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Dynamic language models for continuously evolving content",
                "authors": [
                    {
                        "first": "Tao",
                        "middle": [],
                        "last": "Spurthi Amba Hombaiah",
                        "suffix": ""
                    },
                    {
                        "first": "Mingyang",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Marc-Alexander",
                        "middle": [],
                        "last": "Bendersky",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Najork",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Spurthi Amba Hombaiah, Tao Chen, Mingyang Zhang, Michael Bendersky, and Marc-Alexander Najork. 2021. Dynamic language models for continuously evolving content. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Lifelong learning via progressive distillation and retrospection",
                "authors": [
                    {
                        "first": "Saihui",
                        "middle": [],
                        "last": "Hou",
                        "suffix": ""
                    },
                    {
                        "first": "Xinyu",
                        "middle": [],
                        "last": "Pan",
                        "suffix": ""
                    },
                    {
                        "first": "Chen",
                        "middle": [
                            "Change"
                        ],
                        "last": "Loy",
                        "suffix": ""
                    },
                    {
                        "first": "Zilei",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Dahua",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "ECCV",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. 2018. Lifelong learning via progres- sive distillation and retrospection. In ECCV.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Parameter-efficient transfer learning for NLP",
                "authors": [
                    {
                        "first": "Neil",
                        "middle": [],
                        "last": "Houlsby",
                        "suffix": ""
                    },
                    {
                        "first": "Andrei",
                        "middle": [],
                        "last": "Giurgiu",
                        "suffix": ""
                    },
                    {
                        "first": "Stanislaw",
                        "middle": [],
                        "last": "Jastrzebski",
                        "suffix": ""
                    },
                    {
                        "first": "Bruna",
                        "middle": [],
                        "last": "Morrone",
                        "suffix": ""
                    },
                    {
                        "first": "Quentin",
                        "middle": [],
                        "last": "De Laroussilhe",
                        "suffix": ""
                    },
                    {
                        "first": "Andrea",
                        "middle": [],
                        "last": "Gesmundo",
                        "suffix": ""
                    },
                    {
                        "first": "Mona",
                        "middle": [],
                        "last": "Attariyan",
                        "suffix": ""
                    },
                    {
                        "first": "Sylvain",
                        "middle": [],
                        "last": "Gelly",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 36th International Conference on Machine Learning, ICML 2019",
                "volume": "97",
                "issue": "",
                "pages": "2790--2799",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. 2019. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Confer- ence on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, volume 97 of Proceedings of Machine Learning Research, pages 2790-2799. PMLR.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Examining temporality in document classification",
                "authors": [
                    {
                        "first": "Xiaolei",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [
                            "J"
                        ],
                        "last": "Paul",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics",
                "volume": "2",
                "issue": "",
                "pages": "694--699",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/P18-2110"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Xiaolei Huang and Michael J. Paul. 2018. Examining temporality in document classification. In Proceed- ings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Pa- pers), pages 694-699, Melbourne, Australia. Asso- ciation for Computational Linguistics.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Continual learning for text classification with information disentanglement based regularization",
                "authors": [
                    {
                        "first": "Yufan",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "Yanzhe",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Jiaao",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Xuezhi",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Diyi",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "",
                "issue": "",
                "pages": "2736--2746",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2021.naacl-main.218"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Yufan Huang, Yanzhe Zhang, Jiaao Chen, Xuezhi Wang, and Diyi Yang. 2021. Continual learning for text classification with information disentanglement based regularization. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2736-2746, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Towards continual knowledge learning of language models",
                "authors": [
                    {
                        "first": "Joel",
                        "middle": [],
                        "last": "Jang",
                        "suffix": ""
                    },
                    {
                        "first": "Seonghyeon",
                        "middle": [],
                        "last": "Ye",
                        "suffix": ""
                    },
                    {
                        "first": "Sohee",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Joongbo",
                        "middle": [],
                        "last": "Shin",
                        "suffix": ""
                    },
                    {
                        "first": "Janghoon",
                        "middle": [],
                        "last": "Han",
                        "suffix": ""
                    },
                    {
                        "first": "Gyeonghun",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "Stanley",
                        "middle": [
                            "Jungkyu"
                        ],
                        "last": "Choi",
                        "suffix": ""
                    },
                    {
                        "first": "Minjoon",
                        "middle": [],
                        "last": "Seo",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2110.03215"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Joel Jang, Seonghyeon Ye, Sohee Yang, Joongbo Shin, Janghoon Han, Gyeonghun Kim, Stanley Jungkyu Choi, and Minjoon Seo. 2021. Towards contin- ual knowledge learning of language models. arXiv preprint arXiv:2110.03215.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "TinyBERT: Distilling BERT for natural language understanding",
                "authors": [
                    {
                        "first": "Xiaoqi",
                        "middle": [],
                        "last": "Jiao",
                        "suffix": ""
                    },
                    {
                        "first": "Yichun",
                        "middle": [],
                        "last": "Yin",
                        "suffix": ""
                    },
                    {
                        "first": "Lifeng",
                        "middle": [],
                        "last": "Shang",
                        "suffix": ""
                    },
                    {
                        "first": "Xin",
                        "middle": [],
                        "last": "Jiang",
                        "suffix": ""
                    },
                    {
                        "first": "Xiao",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Linlin",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Fang",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Qun",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Findings of the Association for Computational Linguistics: EMNLP 2020",
                "volume": "",
                "issue": "",
                "pages": "4163--4174",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.findings-emnlp.372"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Xiaoqi Jiao, Yichun Yin, Lifeng Shang, Xin Jiang, Xiao Chen, Linlin Li, Fang Wang, and Qun Liu. 2020. TinyBERT: Distilling BERT for natural lan- guage understanding. In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 4163-4174, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Measuring the evolution of a scientific field through citation frames",
                "authors": [
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Jurgens",
                        "suffix": ""
                    },
                    {
                        "first": "Srijan",
                        "middle": [],
                        "last": "Kumar",
                        "suffix": ""
                    },
                    {
                        "first": "Raine",
                        "middle": [],
                        "last": "Hoover",
                        "suffix": ""
                    },
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Mc-Farland",
                        "suffix": ""
                    },
                    {
                        "first": "Dan",
                        "middle": [],
                        "last": "Jurafsky",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Transactions of the Association for Computational Linguistics",
                "volume": "6",
                "issue": "",
                "pages": "391--406",
                "other_ids": {
                    "DOI": [
                        "10.1162/tacl_a_00028"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "David Jurgens, Srijan Kumar, Raine Hoover, Dan Mc- Farland, and Dan Jurafsky. 2018. Measuring the evo- lution of a scientific field through citation frames. Transactions of the Association for Computational Linguistics, 6:391-406.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Rational LAMOL: A rationale-based lifelong learning framework",
                "authors": [
                    {
                        "first": "Kasidis",
                        "middle": [],
                        "last": "Kanwatchara",
                        "suffix": ""
                    },
                    {
                        "first": "Thanapapas",
                        "middle": [],
                        "last": "Horsuwan",
                        "suffix": ""
                    },
                    {
                        "first": "Piyawat",
                        "middle": [],
                        "last": "Lertvittayakumjorn",
                        "suffix": ""
                    },
                    {
                        "first": "Boonserm",
                        "middle": [],
                        "last": "Kijsirikul",
                        "suffix": ""
                    },
                    {
                        "first": "Peerapon",
                        "middle": [],
                        "last": "Vateekul",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing",
                "volume": "1",
                "issue": "",
                "pages": "2942--2953",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2021.acl-long.229"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Kasidis Kanwatchara, Thanapapas Horsuwan, Piyawat Lertvittayakumjorn, Boonserm Kijsirikul, and Peer- apon Vateekul. 2021. Rational LAMOL: A rationale-based lifelong learning framework. In Pro- ceedings of the 59th Annual Meeting of the Associa- tion for Computational Linguistics and the 11th In- ternational Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 2942- 2953, Online. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Tom\u00e1s Kocisk\u00fd, Susannah Young, and P. Blunsom. 2021. Assessing temporal generalization in neural language models. NeurIPS",
                "authors": [
                    {
                        "first": "Angeliki",
                        "middle": [],
                        "last": "Lazaridou",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Kuncoro",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Gribovskaya",
                        "suffix": ""
                    },
                    {
                        "first": "Devang",
                        "middle": [],
                        "last": "Agrawal",
                        "suffix": ""
                    },
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Liska",
                        "suffix": ""
                    },
                    {
                        "first": "Tayfun",
                        "middle": [],
                        "last": "Terzi",
                        "suffix": ""
                    },
                    {
                        "first": "Mai",
                        "middle": [],
                        "last": "Gimenez",
                        "suffix": ""
                    },
                    {
                        "first": "Cyprien",
                        "middle": [],
                        "last": "De Masson D'autume",
                        "suffix": ""
                    },
                    {
                        "first": "Sebastian",
                        "middle": [],
                        "last": "Ruder",
                        "suffix": ""
                    },
                    {
                        "first": "Dani",
                        "middle": [],
                        "last": "Yogatama",
                        "suffix": ""
                    },
                    {
                        "first": "Kris",
                        "middle": [],
                        "last": "Cao",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Angeliki Lazaridou, A. Kuncoro, E. Gribovskaya, De- vang Agrawal, Adam Liska, Tayfun Terzi, Mai Gimenez, Cyprien de Masson d'Autume, Sebastian Ruder, Dani Yogatama, Kris Cao, Tom\u00e1s Kocisk\u00fd, Susannah Young, and P. Blunsom. 2021. Assessing temporal generalization in neural language models. NeurIPS.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Learning without forgetting",
                "authors": [
                    {
                        "first": "Zhizhong",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Derek",
                        "middle": [],
                        "last": "Hoiem",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "IEEE Transactions on Pattern Analysis and Machine Intelligence",
                "volume": "40",
                "issue": "",
                "pages": "2935--2947",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhizhong Li and Derek Hoiem. 2018. Learning with- out forgetting. IEEE Transactions on Pattern Analy- sis and Machine Intelligence, 40:2935-2947.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Continual learning for sentence representations using conceptors",
                "authors": [
                    {
                        "first": "Tianlin",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Lyle",
                        "middle": [],
                        "last": "Ungar",
                        "suffix": ""
                    },
                    {
                        "first": "Jo\u00e3o",
                        "middle": [],
                        "last": "Sedoc",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "1",
                "issue": "",
                "pages": "3274--3279",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/N19-1331"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Tianlin Liu, Lyle Ungar, and Jo\u00e3o Sedoc. 2019a. Con- tinual learning for sentence representations using conceptors. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers), pages 3274-3279, Minneapolis, Minnesota. Associ- ation for Computational Linguistics.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Roberta: A robustly optimized bert pretraining approach",
                "authors": [
                    {
                        "first": "Yinhan",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Myle",
                        "middle": [],
                        "last": "Ott",
                        "suffix": ""
                    },
                    {
                        "first": "Naman",
                        "middle": [],
                        "last": "Goyal",
                        "suffix": ""
                    },
                    {
                        "first": "Jingfei",
                        "middle": [],
                        "last": "Du",
                        "suffix": ""
                    },
                    {
                        "first": "Mandar",
                        "middle": [],
                        "last": "Joshi",
                        "suffix": ""
                    },
                    {
                        "first": "Danqi",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Omer",
                        "middle": [],
                        "last": "Levy",
                        "suffix": ""
                    },
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Lewis",
                        "suffix": ""
                    },
                    {
                        "first": "Luke",
                        "middle": [],
                        "last": "Zettlemoyer",
                        "suffix": ""
                    },
                    {
                        "first": "Veselin",
                        "middle": [],
                        "last": "Stoyanov",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Man- dar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019b. Roberta: A robustly optimized bert pretraining ap- proach. ArXiv, abs/1907.11692.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "S2ORC: The semantic scholar open research corpus",
                "authors": [
                    {
                        "first": "Kyle",
                        "middle": [],
                        "last": "Lo",
                        "suffix": ""
                    },
                    {
                        "first": "Lucy",
                        "middle": [
                            "Lu"
                        ],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Mark",
                        "middle": [],
                        "last": "Neumann",
                        "suffix": ""
                    },
                    {
                        "first": "Rodney",
                        "middle": [],
                        "last": "Kinney",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Weld",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "4969--4983",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.acl-main.447"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kin- ney, and Daniel Weld. 2020. S2ORC: The semantic scholar open research corpus. In Proceedings of the 58th Annual Meeting of the Association for Compu- tational Linguistics, pages 4969-4983, Online. As- sociation for Computational Linguistics.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Multi-task identification of entities, relations, and coreference for scientific knowledge graph construction",
                "authors": [
                    {
                        "first": "Yi",
                        "middle": [],
                        "last": "Luan",
                        "suffix": ""
                    },
                    {
                        "first": "Luheng",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Mari",
                        "middle": [],
                        "last": "Ostendorf",
                        "suffix": ""
                    },
                    {
                        "first": "Hannaneh",
                        "middle": [],
                        "last": "Hajishirzi",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
                "volume": "",
                "issue": "",
                "pages": "3219--3232",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D18-1360"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Yi Luan, Luheng He, Mari Ostendorf, and Hannaneh Hajishirzi. 2018. Multi-task identification of enti- ties, relations, and coreference for scientific knowl- edge graph construction. In Proceedings of the 2018 Conference on Empirical Methods in Natural Lan- guage Processing, pages 3219-3232, Brussels, Bel- gium. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF35": {
                "ref_id": "b35",
                "title": "Time waits for no one! analysis and challenges of temporal misalignment",
                "authors": [
                    {
                        "first": "Kelvin",
                        "middle": [],
                        "last": "Luu",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Khashabi",
                        "suffix": ""
                    },
                    {
                        "first": "Suchin",
                        "middle": [],
                        "last": "Gururangan",
                        "suffix": ""
                    },
                    {
                        "first": "Karishma",
                        "middle": [],
                        "last": "Mandyam",
                        "suffix": ""
                    },
                    {
                        "first": "Noah",
                        "middle": [
                            "A"
                        ],
                        "last": "Smith",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kelvin Luu, Daniel Khashabi, Suchin Gururangan, Kar- ishma Mandyam, and Noah A. Smith. 2021. Time waits for no one! analysis and challenges of tempo- ral misalignment.",
                "links": null
            },
            "BIBREF36": {
                "ref_id": "b36",
                "title": "Multidomain pretrained language models for green NLP",
                "authors": [
                    {
                        "first": "Antonis",
                        "middle": [],
                        "last": "Maronikolakis",
                        "suffix": ""
                    },
                    {
                        "first": "Hinrich",
                        "middle": [],
                        "last": "Sch\u00fctze",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Proceedings of the Second Workshop on Domain Adaptation for NLP",
                "volume": "",
                "issue": "",
                "pages": "1--8",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Antonis Maronikolakis and Hinrich Sch\u00fctze. 2021. Multidomain pretrained language models for green NLP. In Proceedings of the Second Workshop on Domain Adaptation for NLP, pages 1-8, Kyiv, Ukraine. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF37": {
                "ref_id": "b37",
                "title": "Merging models with fisher-weighted averaging",
                "authors": [
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Matena",
                        "suffix": ""
                    },
                    {
                        "first": "Colin",
                        "middle": [],
                        "last": "Raffel",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "ArXiv",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Michael Matena and Colin Raffel. 2021. Merging models with fisher-weighted averaging. ArXiv, abs/2111.09832.",
                "links": null
            },
            "BIBREF38": {
                "ref_id": "b38",
                "title": "Covid-twitter-bert: A natural language processing model to analyse covid-19 content on twitter",
                "authors": [
                    {
                        "first": "Martin",
                        "middle": [],
                        "last": "M\u00fcller",
                        "suffix": ""
                    },
                    {
                        "first": "Marcel",
                        "middle": [],
                        "last": "Salath\u00e9",
                        "suffix": ""
                    },
                    {
                        "first": "Per",
                        "middle": [
                            "Egil"
                        ],
                        "last": "Kummervold",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Martin M\u00fcller, Marcel Salath\u00e9, and Per Egil Kummer- vold. 2020. Covid-twitter-bert: A natural language processing model to analyse covid-19 content on twitter. ArXiv, abs/2005.07503.",
                "links": null
            },
            "BIBREF39": {
                "ref_id": "b39",
                "title": "The materials science procedural text corpus: Annotating materials synthesis procedures with shallow semantic structures",
                "authors": [
                    {
                        "first": "Zachary",
                        "middle": [],
                        "last": "Sheshera Mysore",
                        "suffix": ""
                    },
                    {
                        "first": "Edward",
                        "middle": [],
                        "last": "Jensen",
                        "suffix": ""
                    },
                    {
                        "first": "Kevin",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "Haw-Shiuan",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "Emma",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "Jeffrey",
                        "middle": [],
                        "last": "Strubell",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Flanigan",
                        "suffix": ""
                    },
                    {
                        "first": "Elsa",
                        "middle": [],
                        "last": "Mccallum",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Olivetti",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 13th Linguistic Annotation Workshop",
                "volume": "",
                "issue": "",
                "pages": "56--64",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/W19-4007"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Sheshera Mysore, Zachary Jensen, Edward Kim, Kevin Huang, Haw-Shiuan Chang, Emma Strubell, Jef- frey Flanigan, Andrew McCallum, and Elsa Olivetti. 2019. The materials science procedural text corpus: Annotating materials synthesis procedures with shal- low semantic structures. In Proceedings of the 13th Linguistic Annotation Workshop, pages 56-64, Flo- rence, Italy. Association for Computational Linguis- tics.",
                "links": null
            },
            "BIBREF40": {
                "ref_id": "b40",
                "title": "BERTweet: A pre-trained language model for English tweets",
                "authors": [
                    {
                        "first": "Thanh",
                        "middle": [],
                        "last": "Dat Quoc Nguyen",
                        "suffix": ""
                    },
                    {
                        "first": "Anh",
                        "middle": [
                            "Tuan"
                        ],
                        "last": "Vu",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Nguyen",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
                "volume": "",
                "issue": "",
                "pages": "9--14",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.emnlp-demos.2"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Dat Quoc Nguyen, Thanh Vu, and Anh Tuan Nguyen. 2020. BERTweet: A pre-trained language model for English tweets. In Proceedings of the 2020 Conference on Empirical Methods in Natural Lan- guage Processing: System Demonstrations, pages 9- 14, Online. Association for Computational Linguis- tics.",
                "links": null
            },
            "BIBREF41": {
                "ref_id": "b41",
                "title": "Data-driven materials research enabled by natural language processing and information extraction",
                "authors": [
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Elsa",
                        "suffix": ""
                    },
                    {
                        "first": "Jacqueline",
                        "middle": [
                            "M"
                        ],
                        "last": "Olivetti",
                        "suffix": ""
                    },
                    {
                        "first": "Edward",
                        "middle": [],
                        "last": "Cole",
                        "suffix": ""
                    },
                    {
                        "first": "Olga",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "Gerbrand",
                        "middle": [],
                        "last": "Kononova",
                        "suffix": ""
                    },
                    {
                        "first": "Thomas",
                        "middle": [],
                        "last": "Ceder",
                        "suffix": ""
                    },
                    {
                        "first": "Anna",
                        "middle": [
                            "M"
                        ],
                        "last": "Yong-Jin Han",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Hiszpanski",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Applied Physics Reviews",
                "volume": "7",
                "issue": "4",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Elsa A Olivetti, Jacqueline M Cole, Edward Kim, Olga Kononova, Gerbrand Ceder, Thomas Yong-Jin Han, and Anna M Hiszpanski. 2020. Data-driven mate- rials research enabled by natural language process- ing and information extraction. Applied Physics Re- views, 7(4):041317.",
                "links": null
            },
            "BIBREF42": {
                "ref_id": "b42",
                "title": "AdapterFusion: Non-destructive task composition for transfer learning",
                "authors": [
                    {
                        "first": "Jonas",
                        "middle": [],
                        "last": "Pfeiffer",
                        "suffix": ""
                    },
                    {
                        "first": "Aishwarya",
                        "middle": [],
                        "last": "Kamath",
                        "suffix": ""
                    },
                    {
                        "first": "Andreas",
                        "middle": [],
                        "last": "R\u00fcckl\u00e9",
                        "suffix": ""
                    },
                    {
                        "first": "Kyunghyun",
                        "middle": [],
                        "last": "Cho",
                        "suffix": ""
                    },
                    {
                        "first": "Iryna",
                        "middle": [],
                        "last": "Gurevych",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume",
                "volume": "",
                "issue": "",
                "pages": "487--503",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jonas Pfeiffer, Aishwarya Kamath, Andreas R\u00fcckl\u00e9, Kyunghyun Cho, and Iryna Gurevych. 2021. AdapterFusion: Non-destructive task composition for transfer learning. In Proceedings of the 16th Conference of the European Chapter of the Associ- ation for Computational Linguistics: Main Volume, pages 487-503, Online. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF43": {
                "ref_id": "b43",
                "title": "icarl: Incremental classifier and representation learning",
                "authors": [
                    {
                        "first": "Alexander",
                        "middle": [],
                        "last": "Sylvestre-Alvise Rebuffi",
                        "suffix": ""
                    },
                    {
                        "first": "Georg",
                        "middle": [],
                        "last": "Kolesnikov",
                        "suffix": ""
                    },
                    {
                        "first": "Christoph",
                        "middle": [
                            "H"
                        ],
                        "last": "Sperl",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Lampert",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "2017 IEEE Conference on Computer Vision and Pattern Recognition",
                "volume": "",
                "issue": "",
                "pages": "5533--5542",
                "other_ids": {
                    "DOI": [
                        "10.1109/CVPR.2017.587"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H. Lampert. 2017. icarl: Incremental classifier and representation learning. In 2017 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2017, Honolulu, HI, USA, July 21-26, 2017, pages 5533-5542. IEEE Computer Society.",
                "links": null
            },
            "BIBREF44": {
                "ref_id": "b44",
                "title": "Temporally-informed analysis of named entity recognition",
                "authors": [
                    {
                        "first": "Shruti",
                        "middle": [],
                        "last": "Rijhwani",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Preotiuc-Pietro",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "7605--7617",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.acl-main.680"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Shruti Rijhwani and Daniel Preotiuc-Pietro. 2020. Temporally-informed analysis of named entity recognition. In Proceedings of the 58th Annual Meeting of the Association for Computational Lin- guistics, pages 7605-7617, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF45": {
                "ref_id": "b45",
                "title": "Catastrophic forgetting, rehearsal and pseudorehearsal",
                "authors": [
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Anthony",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Robins",
                        "suffix": ""
                    }
                ],
                "year": 1995,
                "venue": "Connect. Sci",
                "volume": "7",
                "issue": "",
                "pages": "123--146",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Anthony V. Robins. 1995. Catastrophic forgetting, re- hearsal and pseudorehearsal. Connect. Sci., 7:123- 146.",
                "links": null
            },
            "BIBREF46": {
                "ref_id": "b46",
                "title": "Temporal adaptation of bert and performance on downstream document classification: Insights from social media",
                "authors": [
                    {
                        "first": "Paul",
                        "middle": [],
                        "last": "R\u00f6ttger",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Pierrehumbert",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Paul R\u00f6ttger and J. Pierrehumbert. 2021. Temporal adaptation of bert and performance on downstream document classification: Insights from social media. Findings of EMNLP.",
                "links": null
            },
            "BIBREF48": {
                "ref_id": "b48",
                "title": "Progress & compress: A scalable framework for continual learning",
                "authors": [
                    {
                        "first": "Yee Whye",
                        "middle": [],
                        "last": "Teh",
                        "suffix": ""
                    },
                    {
                        "first": "Razvan",
                        "middle": [],
                        "last": "Pascanu",
                        "suffix": ""
                    },
                    {
                        "first": "Raia",
                        "middle": [],
                        "last": "Hadsell",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "Proceedings of the 35th International Conference on Machine Learning",
                "volume": "80",
                "issue": "",
                "pages": "4535--4544",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yee Whye Teh, Razvan Pascanu, and Raia Hadsell. 2018. Progress & compress: A scalable framework for continual learning. In Proceedings of the 35th International Conference on Machine Learning, ICML 2018, Stockholmsm\u00e4ssan, Stockholm, Swe- den, July 10-15, 2018, volume 80 of Proceedings of Machine Learning Research, pages 4535-4544. PMLR.",
                "links": null
            },
            "BIBREF49": {
                "ref_id": "b49",
                "title": "Recurrent neural network language model adaptation with curriculum learning",
                "authors": [
                    {
                        "first": "Yangyang",
                        "middle": [],
                        "last": "Shi",
                        "suffix": ""
                    },
                    {
                        "first": "Martha",
                        "middle": [],
                        "last": "Larson",
                        "suffix": ""
                    },
                    {
                        "first": "Catholijn",
                        "middle": [
                            "M"
                        ],
                        "last": "Jonker",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Comput. Speech Lang",
                "volume": "33",
                "issue": "",
                "pages": "136--154",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yangyang Shi, Martha Larson, and Catholijn M. Jonker. 2015. Recurrent neural network language model adaptation with curriculum learning. Comput. Speech Lang., 33:136-154.",
                "links": null
            },
            "BIBREF50": {
                "ref_id": "b50",
                "title": "LAMOL: language modeling for lifelong language learning",
                "authors": [
                    {
                        "first": "Fan-Keng",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Cheng-Hao",
                        "middle": [],
                        "last": "Ho",
                        "suffix": ""
                    },
                    {
                        "first": "Hung-Yi",
                        "middle": [],
                        "last": "Lee",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "8th International Conference on Learning Representations",
                "volume": "2020",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Fan-Keng Sun, Cheng-Hao Ho, and Hung-Yi Lee. 2020. LAMOL: language modeling for lifelong lan- guage learning. In 8th International Conference on Learning Representations, ICLR 2020, Addis Ababa, Ethiopia, April 26-30, 2020. OpenReview.net.",
                "links": null
            },
            "BIBREF51": {
                "ref_id": "b51",
                "title": "Patient knowledge distillation for BERT model compression",
                "authors": [
                    {
                        "first": "Siqi",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Yu",
                        "middle": [],
                        "last": "Cheng",
                        "suffix": ""
                    },
                    {
                        "first": "Zhe",
                        "middle": [],
                        "last": "Gan",
                        "suffix": ""
                    },
                    {
                        "first": "Jingjing",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP)",
                "volume": "",
                "issue": "",
                "pages": "4323--4332",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/D19-1441"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Siqi Sun, Yu Cheng, Zhe Gan, and Jingjing Liu. 2019. Patient knowledge distillation for BERT model com- pression. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natu- ral Language Processing (EMNLP-IJCNLP), pages 4323-4332, Hong Kong, China. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF52": {
                "ref_id": "b52",
                "title": "Chemprot-3.0: a global chemical biology diseases mapping",
                "authors": [
                    {
                        "first": "Jens",
                        "middle": [],
                        "last": "Vindahl",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jens Vindahl. 2016. Chemprot-3.0: a global chemical biology diseases mapping.",
                "links": null
            },
            "BIBREF53": {
                "ref_id": "b53",
                "title": "Sentence embedding alignment for lifelong relation extraction",
                "authors": [
                    {
                        "first": "Hong",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Wenhan",
                        "middle": [],
                        "last": "Xiong",
                        "suffix": ""
                    },
                    {
                        "first": "Mo",
                        "middle": [],
                        "last": "Yu",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaoxiao",
                        "middle": [],
                        "last": "Guo",
                        "suffix": ""
                    },
                    {
                        "first": "Shiyu",
                        "middle": [],
                        "last": "Chang",
                        "suffix": ""
                    },
                    {
                        "first": "William",
                        "middle": [
                            "Yang"
                        ],
                        "last": "Wang",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies",
                "volume": "1",
                "issue": "",
                "pages": "796--806",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/N19-1086"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Hong Wang, Wenhan Xiong, Mo Yu, Xiaoxiao Guo, Shiyu Chang, and William Yang Wang. 2019. Sen- tence embedding alignment for lifelong relation ex- traction. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, Volume 1 (Long and Short Papers), pages 796-806, Minneapolis, Minnesota. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF54": {
                "ref_id": "b54",
                "title": "Efficient meta lifelonglearning with limited memory",
                "authors": [
                    {
                        "first": "Zirui",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Barnabas",
                        "middle": [],
                        "last": "Sanket Vaibhav Mehta",
                        "suffix": ""
                    },
                    {
                        "first": "Jaime",
                        "middle": [],
                        "last": "Poczos",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Carbonell",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "535--548",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.emnlp-main.39"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Zirui Wang, Sanket Vaibhav Mehta, Barnabas Poczos, and Jaime Carbonell. 2020. Efficient meta lifelong- learning with limited memory. In Proceedings of the 2020 Conference on Empirical Methods in Natu- ral Language Processing (EMNLP), pages 535-548, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF55": {
                "ref_id": "b55",
                "title": "Curriculum learning for natural language understanding",
                "authors": [
                    {
                        "first": "Benfeng",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Licheng",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Zhendong",
                        "middle": [],
                        "last": "Mao",
                        "suffix": ""
                    },
                    {
                        "first": "Quan",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Hongtao",
                        "middle": [],
                        "last": "Xie",
                        "suffix": ""
                    },
                    {
                        "first": "Yongdong",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "6095--6104",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2020.acl-main.542"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Benfeng Xu, Licheng Zhang, Zhendong Mao, Quan Wang, Hongtao Xie, and Yongdong Zhang. 2020. Curriculum learning for natural language under- standing. In Proceedings of the 58th Annual Meet- ing of the Association for Computational Linguistics, pages 6095-6104, Online. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF56": {
                "ref_id": "b56",
                "title": "Adapt-and-distill: Developing small, fast and effective pretrained language models for domains",
                "authors": [
                    {
                        "first": "Yunzhi",
                        "middle": [],
                        "last": "Yao",
                        "suffix": ""
                    },
                    {
                        "first": "Shaohan",
                        "middle": [],
                        "last": "Huang",
                        "suffix": ""
                    },
                    {
                        "first": "Wenhui",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Dong",
                        "suffix": ""
                    },
                    {
                        "first": "Furu",
                        "middle": [],
                        "last": "Wei",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
                "volume": "",
                "issue": "",
                "pages": "460--470",
                "other_ids": {
                    "DOI": [
                        "10.18653/v1/2021.findings-acl.40"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Yunzhi Yao, Shaohan Huang, Wenhui Wang, Li Dong, and Furu Wei. 2021. Adapt-and-distill: Developing small, fast and effective pretrained language mod- els for domains. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 460-470, Online. Association for Computa- tional Linguistics.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "uris": null,
                "type_str": "figure",
                "num": null,
                "text": "Comparison of adapter, memory replay, and distillation-based continual learning algorithms. Details of the methods are introduced in Sec. 3."
            },
            "FIGREF1": {
                "uris": null,
                "type_str": "figure",
                "num": null,
                "text": "Performance evolution of downstream models."
            },
            "FIGREF2": {
                "uris": null,
                "type_str": "figure",
                "num": null,
                "text": "Performance of downstream models with various number of training examples, exemplified with SciERC. The models are fine-tuned from the final pretrained model (f4)."
            },
            "FIGREF3": {
                "uris": null,
                "type_str": "figure",
                "num": null,
                "text": "Performance of downstream models with various number of training examples. The models are fine-tuned from the final pretrained model (f4)."
            },
            "FIGREF4": {
                "uris": null,
                "type_str": "figure",
                "num": null,
                "text": "Cosine distance of vocabulary distributions between each pair of datasets in two data streams."
            },
            "TABREF0": {
                "content": "<table/>",
                "num": null,
                "text": "Summary of downstream datasets relevant to each domain in the research paper stream.",
                "html": null,
                "type_str": "table"
            },
            "TABREF1": {
                "content": "<table><tr><td>Task</td><td colspan=\"2\">D1 -Biomedical</td><td>D2 -Computer Science</td><td colspan=\"2\">D3 -Materials Science</td><td>D4 -Physics</td></tr><tr><td>Dataset</td><td colspan=\"3\">Chemprot RCT-Sample MLM ACL-ARC SciERC MLM</td><td>MNER</td><td>Synthesis MLM Keyphrase Hyponym MLM</td></tr><tr><td colspan=\"6\">Roberta-base 1.993 Rep-KD 82.03\u00b10.7 78.07\u00b10.7 82.34\u00b10.3 79.59\u00b10.5 1.684 71.17\u00b12.5 78.78\u00b11.1 1.810 84.13\u00b10.3 92.02\u00b10.8 1.585 65.96\u00b11.6 73.93\u00b15.5 1.389</td></tr><tr><td>Contrast-KD</td><td>82.29\u00b10.5</td><td>79.92\u00b10.4</td><td colspan=\"3\">1.722 71.15\u00b11.1 80.49\u00b11.6 1.856 83.26\u00b10.4 92.62\u00b10.7 1.612 65.95\u00b11.7 72.26\u00b13.1 1.428</td></tr><tr><td>SEED-KD</td><td>82.78\u00b10.3</td><td>80.38\u00b10.4</td><td colspan=\"3\">1.720 69.98\u00b12.4 81.61\u00b10.7 1.829 82.99\u00b10.4 92.35\u00b10.7 1.609 65.35\u00b11.0 74.79\u00b14.1 1.401</td></tr><tr><td>SEED-Logit-KD</td><td>83.72\u00b10.4</td><td>81.05\u00b10.2</td><td colspan=\"3\">1.391 69.90\u00b14.5 83.03\u00b10.6 1.703 83.28\u00b10.5 92.87\u00b11.0 1.428 65.96\u00b11.5 71.92\u00b15.5 1.460</td></tr><tr><td>Task-Specific LM</td><td>83.74\u00b10.3</td><td>81.10\u00b10.5</td><td colspan=\"3\">1.210 72.20\u00b12.6 81.24\u00b11.7 1.629 84.02\u00b10.2 91.56\u00b10.4 1.418 65.95\u00b11.1 69.43\u00b14.5 1.426</td></tr><tr><td>MTL</td><td>82.91\u00b11.6</td><td>80.67\u00b10.4</td><td colspan=\"3\">1.289 69.46\u00b11.8 81.12\u00b10.8 1.616 83.92\u00b10.3 92.66\u00b10.6 1.355 65.37\u00b11.6 73.31\u00b15.2 1.418</td></tr></table>",
                "num": null,
                "text": "64.32\u00b12.8 79.07\u00b11.6 2.153 83.15\u00b10.3 91.25\u00b10.6 2.117 66.21\u00b11.0 67.59\u00b14.5 2.278 Sequential Pretraining 82.09\u00b10.5 79.60\u00b10.5 1.654 72.73\u00b12.9 81.43\u00b10.8 1.807 83.99\u00b10.3 92.10\u00b11.0 1.590 67.57\u00b11.0 74.68\u00b14.4 1.381 ER 82.73\u00b10.3 79.98\u00b10.3 1.737 72.50\u00b11.0 81.64\u00b11.1 1.857 83.99\u00b10.4 92.65\u00b10.4 1.621 66.11\u00b11.1 72.82\u00b14.3 1.391 Online EWC 81.83\u00b10.2 78.84\u00b10.5 1.655 71.81\u00b12.6 80.79\u00b10.5 1.803 83.43\u00b10.4 91.89\u00b10.5 1.571 66.70\u00b10.6 72.98\u00b16.0 1.388 Adapter 83.30\u00b10.4 80.41\u00b10.4 1.417 69.32\u00b13.5 80.22\u00b11.5 1.633 83.91\u00b10.3 91.69\u00b10.6 1.522 66.23\u00b11.4 69.65\u00b14.5 1.554 Layer Expansion 83.74\u00b10.3 81.10\u00b10.5 1.210 65.17\u00b12.9 79.35\u00b10.8 1.756 82.48\u00b10.4 92.33\u00b11.0 1.389 65.70\u00b11.1 73.34\u00b13.7 1.534 Logit-KD 83.39\u00b10.4 81.21\u00b10.1 1.392 73.70\u00b13.4 81.92\u00b10.8 1.699 83.96\u00b10.3 92.20\u00b11.0 1.425 64.75\u00b11.1 71.29\u00b13.6 1.460",
                "html": null,
                "type_str": "table"
            },
            "TABREF2": {
                "content": "<table/>",
                "num": null,
                "text": "Results on the Research Paper stream. We report log perplexity of MLM and the performance of downstream models fine-tuned from the final checkpoint of the pretrained model (t = 4). Performance of the best performing CL algorithm is marked bold.",
                "html": null,
                "type_str": "table"
            },
            "TABREF4": {
                "content": "<table/>",
                "num": null,
                "text": "",
                "html": null,
                "type_str": "table"
            },
            "TABREF7": {
                "content": "<table><tr><td>We show fine-tuning performance over years 2018 and 2020 (D3, D4) and the Temporal generalization from 2014 or 2016 to 2020 data (D1 \u2192 D4, D2 \u2192 D4) on Twitter Hashtag and Emoji predic-tion datasets. Models are fine-tuned from the final pre-trained model fT . Full results are included in Appendix C.</td></tr></table>",
                "num": null,
                "text": "Results on temporal data stream.",
                "html": null,
                "type_str": "table"
            },
            "TABREF10": {
                "content": "<table><tr><td>Task</td><td>2014</td><td>2016</td><td>2018</td><td>2020</td></tr><tr><td/><td colspan=\"2\">Hashtag Prediction</td><td/><td/></tr><tr><td>RoBERTa-base</td><td colspan=\"2\">56.65 Emoji Prediction</td><td/><td/></tr><tr><td>RoBERTa-base Sequential PT ER Adapter Logit-KD Rep-KD Contrast-KD SEED-KD SEED-Logit-KD Task-Specific (2014)</td><td colspan=\"4\">28.73 \u00b10.2 26.86 \u00b10.2 25.71 \u00b10.1 24.42 \u00b10.2 32.69 \u00b10.2 30.55 \u00b10.3 29.30 \u00b10.1 27.69 \u00b10.1 32.88 \u00b10.2 30.52 \u00b10.2 29.50 \u00b10.1 27.75 \u00b10.1 32.15 \u00b10.2 29.85 \u00b10.0 28.72 \u00b10.0 26.80 \u00b10.3 33.08 \u00b10.3 30.88 \u00b10.1 29.77 \u00b10.1 27.80 \u00b10.1 32.71 \u00b10.2 30.51 \u00b10.2 29.45 \u00b10.1 27.27 \u00b10.2 32.90 \u00b10.1 31.01 \u00b10.1 29.48 \u00b10.2 27.72 \u00b10.3 32.91 \u00b10.1 30.84 \u00b10.3 30.12 \u00b10.1 27.66 \u00b10.1 33.28 \u00b10.1 31.17 \u00b10.1 29.98 \u00b10.1 27.84 \u00b10.2 33.37</td></tr></table>",
                "num": null,
                "text": "\u00b10.6 45.50 \u00b12.1 48.08 \u00b11.0 56.42 \u00b10.2 Sequential PT 59.00 \u00b10.1 54.28 \u00b10.3 56.79 \u00b10.5 59.85 \u00b10.4 ER 59.00 \u00b10.1 54.90 \u00b10.2 56.93 \u00b10.1 59.56 \u00b11.7 Adapter 58.76 \u00b10.7 52.55 \u00b11.5 54.34 \u00b11.7 59.01 \u00b11.0 Logit-KD 60.93 \u00b10.5 55.96 \u00b10.2 58.21 \u00b10.5 60.52 \u00b10.2 Rep-KD 60.47 \u00b10.1 51.77 \u00b12.6 55.79 \u00b11.4 59.80 \u00b10.2 Contrast-KD 60.72 \u00b10.6 55.85 \u00b10.0 57.94 \u00b10.4 59.54 \u00b10.3 SEED-KD 58.82 \u00b10.4 54.55 \u00b10.5 56.87 \u00b10.2 59.71 \u00b10.2 SEED-Logit-KD 61.28 \u00b10.2 55.59 \u00b10.5 57.75 \u00b10.4 60.74 \u00b10.6 Task-Specific (2014) 61.62 \u00b10.3 55.38 \u00b10.6 56.16 \u00b10.6 59.59 \u00b10.3 Task-Specific (Latest) 59.91 \u00b10.3 55.47 \u00b11.0 56.61 \u00b10.4 59.87 \u00b10.6 MTL 60.51 \u00b10.3 55.16 \u00b11.6 57.89 \u00b10.4 59.95 \u00b10.3 \u00b10.2 30.54 \u00b10.3 28.94 \u00b10.0 26.98 \u00b10.2 Task-Specific (Latest) 32.31 \u00b10.0 29.83 \u00b10.5 29.06 \u00b10.2 27.19 \u00b10.1 MTL 32.78 \u00b10.1 30.54 \u00b10.0 29.52 \u00b10.2 27.47 \u00b10.0",
                "html": null,
                "type_str": "table"
            },
            "TABREF11": {
                "content": "<table/>",
                "num": null,
                "text": "Full performance on Twitter Hashtag prediction and",
                "html": null,
                "type_str": "table"
            }
        }
    }
}