Base_JSON/prefixC/json/computel/2022.computel-1.5.json

{
    "paper_id": "2022",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T13:12:12.858299Z"
    },
    "title": "One Wug, Two Wug+s: Transformer Inflection Models Hallucinate Affixes",
    "authors": [
        {
            "first": "Farhan",
            "middle": [],
            "last": "Samir",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of British Columbia",
                "location": {}
            },
            "email": "[email protected]"
        },
        {
            "first": "Miikka",
            "middle": [],
            "last": "Silfverberg",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "University of British Columbia",
                "location": {}
            },
            "email": ""
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "Data augmentation strategies are increasingly important in NLP pipelines for low-resourced and endangered languages, and in neural morphological inflection, augmentation by so called data hallucination is a popular technique. This paper presents a detailed analysis of inflection models trained with and without data hallucination for the low-resourced Canadian Indigenous language Gitksan. Our analysis reveals evidence for a concatenative inductive bias in augmented models-in contrast to models trained without hallucination, they strongly prefer affixing inflection patterns over suppletive ones. We find that preference for affixation in general improves inflection performance in \"wug test\" like settings, where the model is asked to inflect lexemes missing from the training set. However, data hallucination dramatically reduces prediction accuracy for reduplicative forms due to a misanalysis of reduplication as affixation. While the overall impact of data hallucination for unseen lexemes remains positive, our findings call for greater qualitative analysis and more varied evaluation conditions in testing automatic inflection systems. Our results indicate that further innovations in data augmentation for computational morphology are desirable.",
    "pdf_parse": {
        "paper_id": "2022",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "Data augmentation strategies are increasingly important in NLP pipelines for low-resourced and endangered languages, and in neural morphological inflection, augmentation by so called data hallucination is a popular technique. This paper presents a detailed analysis of inflection models trained with and without data hallucination for the low-resourced Canadian Indigenous language Gitksan. Our analysis reveals evidence for a concatenative inductive bias in augmented models-in contrast to models trained without hallucination, they strongly prefer affixing inflection patterns over suppletive ones. We find that preference for affixation in general improves inflection performance in \"wug test\" like settings, where the model is asked to inflect lexemes missing from the training set. However, data hallucination dramatically reduces prediction accuracy for reduplicative forms due to a misanalysis of reduplication as affixation. While the overall impact of data hallucination for unseen lexemes remains positive, our findings call for greater qualitative analysis and more varied evaluation conditions in testing automatic inflection systems. Our results indicate that further innovations in data augmentation for computational morphology are desirable.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Data augmentation strategies, for instance, backtranslation (Sennrich et al., 2016) and mixed sample data augmentation (Zhang et al., 2018; Guo et al., 2020) , are increasingly important components of NLP pipelines (Feng et al., 2021) . These strategies often form the cornerstone of modern NLP models for lower-resourced and endangered languages and dialects in particular (e.g., Kumar et al., 2021; Hauer et al., 2020; Zhao et al., 2020; Ryan and Hulden, 2020) , where models can otherwise badly overfit due to the paucity of training data. The same model trained without hallucination exhibits no preference.",
                "cite_spans": [
                    {
                        "start": 60,
                        "end": 83,
                        "text": "(Sennrich et al., 2016)",
                        "ref_id": "BIBREF20"
                    },
                    {
                        "start": 119,
                        "end": 139,
                        "text": "(Zhang et al., 2018;",
                        "ref_id": null
                    },
                    {
                        "start": 140,
                        "end": 157,
                        "text": "Guo et al., 2020)",
                        "ref_id": "BIBREF8"
                    },
                    {
                        "start": 215,
                        "end": 234,
                        "text": "(Feng et al., 2021)",
                        "ref_id": null
                    },
                    {
                        "start": 381,
                        "end": 400,
                        "text": "Kumar et al., 2021;",
                        "ref_id": null
                    },
                    {
                        "start": 401,
                        "end": 420,
                        "text": "Hauer et al., 2020;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 421,
                        "end": 439,
                        "text": "Zhao et al., 2020;",
                        "ref_id": null
                    },
                    {
                        "start": 440,
                        "end": 462,
                        "text": "Ryan and Hulden, 2020)",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Consider the task of low-resource morphological inflection: high-capacity neural models trained without data augmentation are prone to collapsing at test time, achieving as little as 0% accuracy (Silfverberg et al., 2017) . Conversely, those very same models trained on artificially augmented data can generalize respectably. Unfortunately, there is little research on understanding why these augmentation strategies work. We know little about the changes they cause in the model -are they simply a form of weight regularization? Do they alleviate class imbalance? Or do they provide a task-specific inductive bias?",
                "cite_spans": [
                    {
                        "start": 195,
                        "end": 221,
                        "text": "(Silfverberg et al., 2017)",
                        "ref_id": "BIBREF23"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In this paper, we investigate the data hallucination strategy, a relatively commonplace strategy (Anastasopoulos and Neubig, 2019; Silfverberg et al., 2017) for increasing the size of small morphological datasets. We conduct our study in the context of developing a Paradigm Cell-Filling (PCFP; Ackerman et al., 2009; Silfverberg and Hulden, 2018) system for the Gitksan language -a critically endangered language with an estimated 300-850 speakers (Dunlop et al., 2018) -that can be used for applications such as developing pedagogical noun and verb conjugation exercises and further computer-assisted language learning applications.",
                "cite_spans": [
                    {
                        "start": 97,
                        "end": 130,
                        "text": "(Anastasopoulos and Neubig, 2019;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 131,
                        "end": 156,
                        "text": "Silfverberg et al., 2017)",
                        "ref_id": "BIBREF23"
                    },
                    {
                        "start": 295,
                        "end": 317,
                        "text": "Ackerman et al., 2009;",
                        "ref_id": "BIBREF0"
                    },
                    {
                        "start": 318,
                        "end": 347,
                        "text": "Silfverberg and Hulden, 2018)",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 449,
                        "end": 470,
                        "text": "(Dunlop et al., 2018)",
                        "ref_id": "BIBREF4"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Given a partial inflectional paradigm with n filled slots and a number of empty slots, the task is to complete the paradigm by predicting all the missing slots from the given ones. Following previous work on PCFP (Silfverberg and Hulden, 2018; Liu and Hulden, 2020) , we leverage morphological reinflection models to complete PCFP. Specifically, we employ the one-source model of Liu and Hulden (2020) : We use each of the n given forms in turn to predict the form in an empty target slot, giving n output forms (see Fig. 1 , where n = 2). We then select one of the output forms as our prediction for the empty slot: We pick the predictions that the model makes with the highest confidence, a decision strategy we denote MAX. 1 Given the relatively small size of our paradigm dataset, further described in Section 2, we investigate whether data hallucination is an effective strategy for mitigating overfitting. In accordance with recent results (Liu and Hulden, 2021) , we find that data hallucination improves performance in \"wug test\" (Berko, 1958) like conditions: where no inflectional variant of a lexeme was witnessed during training. Surprisingly, however, we also find that data hallucination significantly worsens performance for lexemes which were partially observed during training; that is at least one of the inflectional variants of the lexeme was present in the training data.",
                "cite_spans": [
                    {
                        "start": 213,
                        "end": 243,
                        "text": "(Silfverberg and Hulden, 2018;",
                        "ref_id": "BIBREF22"
                    },
                    {
                        "start": 244,
                        "end": 265,
                        "text": "Liu and Hulden, 2020)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 380,
                        "end": 401,
                        "text": "Liu and Hulden (2020)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 726,
                        "end": 727,
                        "text": "1",
                        "ref_id": null
                    },
                    {
                        "start": 946,
                        "end": 968,
                        "text": "(Liu and Hulden, 2021)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 1038,
                        "end": 1051,
                        "text": "(Berko, 1958)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 517,
                        "end": 523,
                        "text": "Fig. 1",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "These findings motivated a controlled error analysis of our PCFP system to discover why data hallucination generalizes to the unobserved test setting but seemingly slashes performance in the observed test setting. This analysis yields two major insights. First, we find that the model trained without hallucination is \"often right for the wrong reason\" (Mc-Coy et al., 2019) : our error analysis reveals that a unaugmented Transformer model exhibits undesirable memorization to a significant degree, even when incorporating recently prescribed parameter settings for inflection (Wu et al., 2021; Liu and Hulden, 2020) . This allows the model to memorize lexeme-specific inflection patterns, rather than",
                "cite_spans": [
                    {
                        "start": 353,
                        "end": 374,
                        "text": "(Mc-Coy et al., 2019)",
                        "ref_id": null
                    },
                    {
                        "start": 578,
                        "end": 595,
                        "text": "(Wu et al., 2021;",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 596,
                        "end": 617,
                        "text": "Liu and Hulden, 2020)",
                        "ref_id": "BIBREF13"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "ROOT we / wa ROOT-1PL.II wa'm ROOT-3.II wet / wat learning the morphophonological structure of the language. That is, we find that the model trained without hallucination relies on a brittle memorization strategy. Second, we find evidence that data hallucination introduces an inductive bias towards concatenative morphology: where inflection is accomplished by appending affixes to a word stem. We find that the MAX strategy combined with data hallucination selects a simpler transformation: In Fig. 1 , the augmented model prefers the simple transformation of appending diit to hap to predict the target hapdiit over the unpredictable transformation from tk'esxw to hapdiit. Conversely, the model trained without hallucination exhibits no strong preference over either transformation. Since concatenative morphology is the dominant inflection process in Gitksan, this inductive bias serves the hallucination model well in inflecting unfamiliar lexemes during testing.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 496,
                        "end": 502,
                        "text": "Fig. 1",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "MSD Form",
                "sec_num": null
            },
            {
                "text": "Data hallucination, however, can be damaging depending on the morphophonological phenomena at hand. We find, for instance, that it dramatically reduces performance in inflections involving reduplication, a transformation that requires copying of phonological material rather than a simple concatenation (Haspelmath and Sims, 2013) . While the overall effect of data augmentation in inflection has been reported as overwhelmingly positive (e.g., Lane and Bird, 2020; Anastasopoulos and Neubig, 2019; Liu and Hulden, 2021) , our detailed analysis reveals that it carries both benefits and drawbacks and should therefore be applied with caution. Furthermore, our findings call for greater qualitative analysis and more varied evaluation conditions in testing automatic inflection systems.",
                "cite_spans": [
                    {
                        "start": 303,
                        "end": 330,
                        "text": "(Haspelmath and Sims, 2013)",
                        "ref_id": "BIBREF9"
                    },
                    {
                        "start": 445,
                        "end": 465,
                        "text": "Lane and Bird, 2020;",
                        "ref_id": "BIBREF12"
                    },
                    {
                        "start": 466,
                        "end": 498,
                        "text": "Anastasopoulos and Neubig, 2019;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 499,
                        "end": 520,
                        "text": "Liu and Hulden, 2021)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "MSD Form",
                "sec_num": null
            },
            {
                "text": "Our dataset comprises paradigms that were programatically extracted from an interlinear-glossed dataset of 18, 000 tokens (Forbes et al., 2017) . De-tails of the gloss to paradigm conversion procedure can be found in Appendix B. The interlinear glosses were collected during still-active language documentation efforts with Gitksan speakers.",
                "cite_spans": [
                    {
                        "start": 122,
                        "end": 143,
                        "text": "(Forbes et al., 2017)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data",
                "sec_num": "2"
            },
            {
                "text": "The Gitksan-speaking community recognizes two dialects: Eastern (Upriver) and Western (Downriver), and our dataset comprises forms from both dialects. Although the two dialects are largely mutually intelligible, some lexical and phonological differences manifest, with the most prominent being a vowel shift. Consider the Gitksan translation for the word \"name\" in Table 1 . The dialectal variation manifests as several entries for a given morphosyntactic description (henceforth MSD) in the paradigm: we (Western) vs. wa (Eastern).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 365,
                        "end": 372,
                        "text": "Table 1",
                        "ref_id": "TABREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Data",
                "sec_num": "2"
            },
            {
                "text": "Instead of attempting to model one-to-many (MSD to form) paradigms, we adhere to the simplifying constraint that each paradigm have a single realization per morphosyntactic description. In order to convert a one-to-many paradigm to a one-to-one paradigm, we aim to select a single form for each MSD so that, taken together, the inflected forms are maximally similar to each other. In the partial paradigm for for Table 1 , the inflection from wa to wa'm is a simpler transformation than we to wa'm, making it simpler for a neural inflection model to acquire generalizable inflection rules. Thus, in Table 1 we would select a one-to-one paradigm with the forms wa, wa'm, and wat.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 413,
                        "end": 420,
                        "text": "Table 1",
                        "ref_id": "TABREF0"
                    },
                    {
                        "start": 599,
                        "end": 606,
                        "text": "Table 1",
                        "ref_id": "TABREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Data",
                "sec_num": "2"
            },
            {
                "text": "To obtain maximally similar inflected forms, we apply the following algorithm to a one-to-many paradigm. First, we generate all possible one-toone realizations of the paradigm. For instance, for Table 1 one paradigm could comprise the MSD-toform mappings: ROOT:wa, ROOT:-1PL.II:wa'm, ROOT-3.II:wet; there would be four possible oneto-one paradigms in total. Next, given a candidate one-to-one paradigm, we construct a fullyconnected graph where each inflectional form is a vertex and every (undirected) edge is weighted by the Levenshtein distance. We then compute the weight of the minimum spanning-tree of the graph. Finally, we return the one-to-one paradigm that has the minimum-spanning tree with the lowest weight. 2 We divide the resulting paradigms into four disjoint subsets. (1) A dataset for training a morpho-logical reinflection model \u03a0 train that will be used for the PCFP task; (2) A test set containing partial paradigms \u03a0 obs so that some of the lexemes' inflectional variants were seen during training while the other inflectional variants are used only for testing; A validation set \u03a0 dev constructed in the same manner as \u03a0 obs ; (4) A test set simulating the conditions of a \"wug test\" (Liu and Hulden, 2021; Berko, 1958) containing complete paradigms (\u03a0 wug ) so that none of the lexemes' inflectional variants were observed during training.",
                "cite_spans": [
                    {
                        "start": 721,
                        "end": 722,
                        "text": "2",
                        "ref_id": null
                    },
                    {
                        "start": 1207,
                        "end": 1229,
                        "text": "(Liu and Hulden, 2021;",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 1230,
                        "end": 1242,
                        "text": "Berko, 1958)",
                        "ref_id": "BIBREF2"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 195,
                        "end": 202,
                        "text": "Table 1",
                        "ref_id": "TABREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Data",
                "sec_num": "2"
            },
            {
                "text": "In order to train or evaluate a reinflection system for PCFP, we first need all the paradigms to have at least two entries. This is necessary since a reinflection datapoint is of the form src_form:src_msd;tgt_form:tgt_msd. Thus, our first step is to drop all paradigms that only have a single entry, providing us with 459 paradigms. Next, we randomly sample paradigms (without replacement) and add them to \u03a0 wug until \u03a0 wug contains 10% of the 1303 forms in our dataset. 3 This procedure guarantees that no forms in paradigms belonging to \u03a0 wug are ever observed during training.",
                "cite_spans": [
                    {
                        "start": 471,
                        "end": 472,
                        "text": "3",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data",
                "sec_num": "2"
            },
            {
                "text": "For the remaining paradigms \u03c0, we split them into two disjoint sets: \u03c0 train and \u03c0 hold\u2212out . The forms in \u03c0 train are added to the training set \u03a0 train . The forms in \u03c0 hold\u2212out are added either to the development set \u03a0 dev or partially observed test set \u03a0 obs . This way, the model is allowed to observe some of the forms belonging to the (partial) paradigms in \u03a0 dev and \u03a0 obs during training. However, it is guaranteed not to have observed the particular forms in \u03a0 dev and \u03a0 obs during training. 4 More concretely, for a paradigm of size n, between 2 and n \u2212 1 forms (inclusive) are placed into train and the remaining forms are all placed into test (or all placed into dev). We obtain the following number of inflectional variants in each disjoint subset: |\u03a0 train | = 927, |\u03a0 dev | = 124, |\u03a0 obs | = 125, |\u03a0 wug | = 131. In the next section, we describe our procedure for employing these four sets of (partial) paradigms for training and evaluating a PCFP system.",
                "cite_spans": [
                    {
                        "start": 501,
                        "end": 502,
                        "text": "4",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data",
                "sec_num": "2"
            },
            {
                "text": "Having split our paradigm dataset into the desired disjoint subsets \u03a0 train , \u03a0 obs , \u03a0 dev , \u03a0 wug , we can train Transformers in morphological reinflection that can, in turn, be used for the PCFP task. 5 Training. We form reinflection training pairs by using the given forms in each paradigm in \u03a0 train . Concretely, for every \u03c0 \u2208 \u03a0 train , we take the cross product of the entries in \u03c0 and learn to reinflect each given form in the paradigm to another form in the same paradigm as demonstrated in Fig. 2 . 6 Counting reinflection datapoints over all paradigms, we obtain 1365 datapoints in the training set for the reinflection system.",
                "cite_spans": [
                    {
                        "start": 204,
                        "end": 205,
                        "text": "5",
                        "ref_id": null
                    },
                    {
                        "start": 509,
                        "end": 510,
                        "text": "6",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 500,
                        "end": 506,
                        "text": "Fig. 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Experiments and Results",
                "sec_num": "3"
            },
            {
                "text": "We train two Transformer models. First, we train a \"standard\" Transformer model on the aforementioned 1365 datapoints using the parameter settings described in Wu et al. (2021) and Liu and Hulden (2020) ; see Appendix A. Next, we train a second \"augmented\" Transformer model, using the same hyperparameter settings, on the original 1365 datapoints in addition to 10, 000 datapoints hallucinated from the original training dataset. We obtain the hallucination method, number of hallucinated examples (10, 000), and implementation from Anastasopoulos and Neubig (2019) . Evaluation. We evaluate the models both on paradigms describing lexemes whose inflections were partially observed (\u03a0 obs ) and lexemes that are entirely unfamiliar (\u03a0 wug ). Since most of our paradigms are very sparse, containing only contain a few forms, we do a leave-one-out style evaluation procedure where, for every target form in either \u03a0 wug or \u03a0 obs that belongs to paradigm \u03c0, we predict it using every other form that belongs to the same paradigm \u03c0. 7 This gives us |\u03c0| \u2212 1 predictions for a target form, where |\u03c0| is the total number filled slots in the paradigm.",
                "cite_spans": [
                    {
                        "start": 160,
                        "end": 176,
                        "text": "Wu et al. (2021)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 181,
                        "end": 202,
                        "text": "Liu and Hulden (2020)",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 534,
                        "end": 566,
                        "text": "Anastasopoulos and Neubig (2019)",
                        "ref_id": "BIBREF1"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments and Results",
                "sec_num": "3"
            },
            {
                "text": "Finally, we use the MAX strategy to select the form that was predicted with the highest likelihood averaged across output characters. We consider a paradigm \u03c0 as correctly predicted if all forms for the paradigm that are present in \u03a0 obs or \u03a0 wug were correctly predicted. Results and Discussion. We make a number of observations regarding the results in Fig. 3 . First, we observe that there is a significant reduction in performance for the unfamiliar lexemes (\u03a0 wug ) relative to the familiar lexemes (\u03a0 obs ) -replicating observations made in the context of the SIGMOR-PHON shared tasks (Goldman et al., 2021; Cotterell et al., 2017; Liu and Hulden, 2021) . We find that the augmented model reduces the deficit to 10%. That hallucination improves performance on unfamiliar lexemes has been previously observed (Liu and Hulden, 2021) .",
                "cite_spans": [
                    {
                        "start": 591,
                        "end": 613,
                        "text": "(Goldman et al., 2021;",
                        "ref_id": null
                    },
                    {
                        "start": 614,
                        "end": 637,
                        "text": "Cotterell et al., 2017;",
                        "ref_id": "BIBREF3"
                    },
                    {
                        "start": 638,
                        "end": 659,
                        "text": "Liu and Hulden, 2021)",
                        "ref_id": "BIBREF14"
                    },
                    {
                        "start": 814,
                        "end": 836,
                        "text": "(Liu and Hulden, 2021)",
                        "ref_id": "BIBREF14"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 355,
                        "end": 361,
                        "text": "Fig. 3",
                        "ref_id": "FIGREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Experiments and Results",
                "sec_num": "3"
            },
            {
                "text": "We also find, however, that hallucination worsens performance on familiar lexemes. In both cases, the aggregate accuracy scores glean little insights into these surprising results. Why does accuracy drop by nearly 50% for the non-hallucination model across the two testing conditions? How does hallucination improve performance on unfamiliar lexemes? And why does hallucination reduce performance on familiar lexeme paradigms? To understand these differences in performance between the two models and testing conditions, we turn to an analysis of the errors.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiments and Results",
                "sec_num": "3"
            },
            {
                "text": "To reveal insights into the behaviour of the two Transformer models, we look into the case of Gitksan pluralization, which is instantiated as suppletion or reduplication depending on the lexeme, enabling us to investigate whether either Transformer can learn two disparate inflectional strategies. This error analysis enables us to systematically characterize the effects hallucination has on the Transformer model in inflection, demonstrating that the effects can be both beneficial and adverse. Unaugmented Transformers memorize inflection patterns. We begin by analyzing the models' behaviour on suppletive forms; Gitksan uses suppletion as a productive strategy for pluralization. For instance, the stem for singular forms for \"laugh\" is tk'esxw, but the plural stem is hap. The transformation from a singular form to a suppletive plural form is unpredictable (ts'ehlx \u2192 hapdiit); the model must instead rely on other plural source forms (e.g., hap \u2192 hapdiit). Even if the model is unable to produce the correct suppletive plural inflection, it should be able to perform the simpler task of placing higher confidence in the prediction from the plural source form (hap) over the singular source form (ts'ehlx). Failing to exhibit this preference would indicate that the model is simply niye'et",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Error analysis",
                "sec_num": "4"
            },
            {
                "text": "n i y e ' e t Figure 2 : From a paradigm in the training data spanning three forms, we can generate six reinflection training examples. Forms are tokenized into individual characters. Further, we distinguish tags for the input form from tags for the output form. memorizing target inflectional forms, rather than trying to acquire the morphophonological structure of the language.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 14,
                        "end": 22,
                        "text": "Figure 2",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "3.II niye'e ROOT niye'e'y 1SG.II n i y e ' e t IN(3.II) OUT(ROOT) n i y e ' e n i y e ' e t IN(3.II) OUT(1SG.II) n i y e ' e ' y n i y e ' e IN(ROOT) OUT(3.II) n i y e ' e t n i y e ' e IN(ROOT) OUT(1SG.II) n i y e ' e ' y n i y e ' e ' y IN(1SG.II) OUT(ROOT) n i y e ' e n i y e ' e ' y IN(1SG.II) OUT(3.II)",
                "sec_num": null
            },
            {
                "text": "Concretely, we acquire all of the 95 suppletive plurals in either \u03a0 wug or \u03a0 obs . We then follow our leave-one-out procedure, where every other form in the same paradigm \u03c0 as the target suppletive plural form is used as a source to try to predict the target form. Instead of evaluating whether the target form was correctly predicted, we test whether the model assigns higher likelihoods to the reinflection examples where the source form is also a suppletive plural (hap) over examples where the source form is singular (tk'esxw).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "3.II niye'e ROOT niye'e'y 1SG.II n i y e ' e t IN(3.II) OUT(ROOT) n i y e ' e n i y e ' e t IN(3.II) OUT(1SG.II) n i y e ' e ' y n i y e ' e IN(ROOT) OUT(3.II) n i y e ' e t n i y e ' e IN(ROOT) OUT(1SG.II) n i y e ' e ' y n i y e ' e ' y IN(1SG.II) OUT(ROOT) n i y e ' e n i y e ' e ' y IN(1SG.II) OUT(3.II)",
                "sec_num": null
            },
            {
                "text": "This analysis can be interpreted as a binary classification task when we hold the target suppletive form (hapdiit) fixed. The task is then to classify the source suppletive plural forms as positive instances and the source singular forms as negative instances. We can then use standard binary classification metrics to quantify performance. We use weighted Average Precision (Murphy, 2012) , where the weight is the total number of suppletive forms in the paradigm \u03c0. We use the Average Precision implementation from scikit-learn (Pedregosa et al., 2011). 8 We find that the augmented model performs significantly better in this task, achieving a weighted Average Precision of .89 while the unaugmented model achieves .52. This analysis provides evidence that the unaugmented model is memorizing the target suppletive plural form (hapdiit), rather than attending to and copying the suppletive plural stem (hap) and concatenating the appropriate affix (\"diit\"). This result can explain, in part, the substantial drop in performance of the unaugmented model from \u03a0 obs to \u03a0 wug : memorization is unlikely to generalize well for inflecting unfamiliar lexemes. Further, it can explain the stronger performance of the hallucination model in predicting forms in \u03a0 wug : this inductive bias towards concatenative morphology can generalize well to unfamiliar lexemes given the prevalence of concatenative morphology in the Gitksan dataset.",
                "cite_spans": [
                    {
                        "start": 530,
                        "end": 557,
                        "text": "(Pedregosa et al., 2011). 8",
                        "ref_id": null
                    }
                ],
                "ref_spans": [
                    {
                        "start": 375,
                        "end": 389,
                        "text": "(Murphy, 2012)",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "3.II niye'e ROOT niye'e'y 1SG.II n i y e ' e t IN(3.II) OUT(ROOT) n i y e ' e n i y e ' e t IN(3.II) OUT(1SG.II) n i y e ' e ' y n i y e ' e IN(ROOT) OUT(3.II) n i y e ' e t n i y e ' e IN(ROOT) OUT(1SG.II) n i y e ' e ' y n i y e ' e ' y IN(1SG.II) OUT(ROOT) n i y e ' e n i y e ' e ' y IN(1SG.II) OUT(3.II)",
                "sec_num": null
            },
            {
                "text": "Augmented Transformers struggle with nonconcatenative morphology.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "3.II niye'e ROOT niye'e'y 1SG.II n i y e ' e t IN(3.II) OUT(ROOT) n i y e ' e n i y e ' e t IN(3.II) OUT(1SG.II) n i y e ' e ' y n i y e ' e IN(ROOT) OUT(3.II) n i y e ' e t n i y e ' e IN(ROOT) OUT(1SG.II) n i y e ' e ' y n i y e ' e ' y IN(1SG.II) OUT(ROOT) n i y e ' e n i y e ' e ' y IN(1SG.II) OUT(3.II)",
                "sec_num": null
            },
            {
                "text": "Our Gitksan paradigm dataset comprises more than just concatenative morphology, however. Another pluralization strategy in Gitksan, albeit rarer, is reduplication, where number is indicated by copying a part of the word stem. For example, wat (\"name\") and hu-wat (\"name+PL\"). The copied stem segment frequently undergoes further phonological alternations in the case of partial reduplication (as opposed to full reduplication; Haspelmath and Sims, 2013) . While reduplication bears superficial resemblance to affixation, it cannot be analyzed as a concatenation of a stem and affix.",
                "cite_spans": [
                    {
                        "start": 427,
                        "end": 453,
                        "text": "Haspelmath and Sims, 2013)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "3.II niye'e ROOT niye'e'y 1SG.II n i y e ' e t IN(3.II) OUT(ROOT) n i y e ' e n i y e ' e t IN(3.II) OUT(1SG.II) n i y e ' e ' y n i y e ' e IN(ROOT) OUT(3.II) n i y e ' e t n i y e ' e IN(ROOT) OUT(1SG.II) n i y e ' e ' y n i y e ' e ' y IN(1SG.II) OUT(ROOT) n i y e ' e n i y e ' e ' y IN(1SG.II) OUT(3.II)",
                "sec_num": null
            },
            {
                "text": "This resemblance, however, is sufficient to confuse prominent data hallucination techniques (Anastasopoulos and Neubig, 2019; Silfverberg et al., 2017) . Consider the Gitksan word dew (\"freeze\") which is pluralized using full reduplication: dewdew. The hallucinated form of this data-point would have random characters substituted for the stem: e.g., txu -> dewtxu. Clearly, this hallucinated datapoint does not preserve the reduplicative structure. Unfortunately, the hallucination strategy could impair the model's ability to perform reduplication, given that the number of examples of reduplication would become smaller relative to the size of the complete dataset.",
                "cite_spans": [
                    {
                        "start": 92,
                        "end": 125,
                        "text": "(Anastasopoulos and Neubig, 2019;",
                        "ref_id": "BIBREF1"
                    },
                    {
                        "start": 126,
                        "end": 151,
                        "text": "Silfverberg et al., 2017)",
                        "ref_id": "BIBREF23"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "3.II niye'e ROOT niye'e'y 1SG.II n i y e ' e t IN(3.II) OUT(ROOT) n i y e ' e n i y e ' e t IN(3.II) OUT(1SG.II) n i y e ' e ' y n i y e ' e IN(ROOT) OUT(3.II) n i y e ' e t n i y e ' e IN(ROOT) OUT(1SG.II) n i y e ' e ' y n i y e ' e ' y IN(1SG.II) OUT(ROOT) n i y e ' e n i y e ' e ' y IN(1SG.II) OUT(3.II)",
                "sec_num": null
            },
            {
                "text": "Indeed, we find strong evidence that the hallucination model is unable to perform reduplication. We find that the standard model is able to predict the 12 instances of reduplication in \u03a0 wug and \u03a0 obs with .92 accuracy, while the hallucination model slashes this proficiency to a mere .25. Our analysis emphasizes the need for data-augmentation techniques that preserve reduplicative structure, given the phenomenon's typologically robust prevalence (Haspelmath and Sims, 2013) .",
                "cite_spans": [
                    {
                        "start": 450,
                        "end": 477,
                        "text": "(Haspelmath and Sims, 2013)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "3.II niye'e ROOT niye'e'y 1SG.II n i y e ' e t IN(3.II) OUT(ROOT) n i y e ' e n i y e ' e t IN(3.II) OUT(1SG.II) n i y e ' e ' y n i y e ' e IN(ROOT) OUT(3.II) n i y e ' e t n i y e ' e IN(ROOT) OUT(1SG.II) n i y e ' e ' y n i y e ' e ' y IN(1SG.II) OUT(ROOT) n i y e ' e n i y e ' e ' y IN(1SG.II) OUT(3.II)",
                "sec_num": null
            },
            {
                "text": "Reduplication is pronounced in the Gitksan dataset and causes problems for current data hallucination methods. However, it is by no means the only phenomenon where data hallucination can generate incorrect inflection patterns. Consider the example of lenition in our paradigm dataset where the final consonant undergoes voicing between vowels: ayook + 3.II -> ayook+'m -> ayooga'm. Hallucination identifies ayoo as the stem here due to the k/g alternation. If a hallucinated stem ending in a consonant like dap is used, we get an example dapk -> dapga'm, where k is no longer surrounded by vowels but is still voiced when the a'm affix attaches, contrary to the morphophonology of Gitksan. Thus, it is possible that hallucination's inability to preserve morphological phenomena like reduplication and lenition explain the drop in performance on the observed paradigms. 9 Approaches that try to perform data hallucination incorporating the target language's structure have been explored (Lane and Bird, 2020) , but it's unclear how to generalize this method without expert knowledge of the target language.",
                "cite_spans": [
                    {
                        "start": 986,
                        "end": 1007,
                        "text": "(Lane and Bird, 2020)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "3.II niye'e ROOT niye'e'y 1SG.II n i y e ' e t IN(3.II) OUT(ROOT) n i y e ' e n i y e ' e t IN(3.II) OUT(1SG.II) n i y e ' e ' y n i y e ' e IN(ROOT) OUT(3.II) n i y e ' e t n i y e ' e IN(ROOT) OUT(1SG.II) n i y e ' e ' y n i y e ' e ' y IN(1SG.II) OUT(ROOT) n i y e ' e n i y e ' e ' y IN(1SG.II) OUT(3.II)",
                "sec_num": null
            },
            {
                "text": "In this paper, we explore the effect of data hallucination on the Gitksan language that is currently underserved in NLP. Given the low amount of training data for the model, inflection models are likely to encounter many unfamiliar lexemes during test time. Thus, it is important to assess the model's ability to make adequate morphological generalizations for such lexemes. To this end, we tested the model's ability to generalize for lexemes on a cline of familiarity from familiar (\u03a0 obs ) to unfamiliar (\u03a0 wug Section 2).",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General Discussion",
                "sec_num": "5"
            },
            {
                "text": "Under these disparate conditions, we find that a data-augmented model and a standard model exhibit drastically different behaviours. We found that the standard model, a Transformer model trained under recommended parameter settings (Wu et al., 2021) , memorizes inflection patterns to a significant degree (Section 3 and Section 4). At the same time, we find that data hallucination alleviates the need for memorization significantly, generalizing well to unfamiliar lexemes (Section 3) with an inductive bias towards concatenative morphology (Section 4). Data hallucination, however, is not universally beneficial: we find it reduces the model's capacity to recognize common morphophonological phenomena (Section 4), limiting the performance improvements it can bring.",
                "cite_spans": [
                    {
                        "start": 232,
                        "end": 249,
                        "text": "(Wu et al., 2021)",
                        "ref_id": "BIBREF24"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General Discussion",
                "sec_num": "5"
            },
            {
                "text": "Although our study was conducted on a single language, we note that our characterization of data hallucination could be informative for languages other than Gitksan. As Section 4 demonstrates, data hallucination can encourage the model to apply voicing in incorrect contexts. Such effects are not limited to Gitksan. In English, data hallucination could give rise to erroneously conditioned allomorphy: for instance, hallucination can generate a synthetic past tense inflection example mar -> mard from a gold standard training example such as like -> liked. The desired hallucinated past tense form is of course mared. Overall, our work suggests common data augmentation strategies for NLP like data hallucination merit closer inspection and that further innovations in data augmentation for computational morphology are desirable. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "General Discussion",
                "sec_num": "5"
            },
            {
                "text": "We train all models using the implementation of Transformer in the fairseq package (Ott et al., 2019) . Both the encoder and decoder have 4 layers with 4 attention heads, an embedding size of 256 and hidden layer size of 1024. We train with the Adam optimizer starting of the learning rate at 0.001. We chose the batch size (400) and maximum updates (5000) based on the highest accuracy on the development data. Our model setting resembles the work of Wu et al. (2021) who found that a relatively large batch size is beneficial for morphological inflection. Prediction is performed with the best checkpoint model, according to validation accuracy score, and a beam of width 5.",
                "cite_spans": [
                    {
                        "start": 83,
                        "end": 101,
                        "text": "(Ott et al., 2019)",
                        "ref_id": "BIBREF17"
                    },
                    {
                        "start": 452,
                        "end": 468,
                        "text": "Wu et al. (2021)",
                        "ref_id": "BIBREF24"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "A Transformer training details",
                "sec_num": null
            },
            {
                "text": "We perform all experiments on a database of Gitksan inflection tables. In total, there are 1055 inflection tables containing 2125 inflected forms. An interlinear-glossed corpus of Gitksan narratives Forbes et al. (2017) forms the basis of our database. The Gitksan corpus is glossed at the root level meaning that word forms are broken down into roots, derivational morphemes and inflectional morphemes. This level of description is too finegrained for our purposes and we, therefore, combine roots and potential derivational material into word stems. The inflected forms for each noun and verb stem are gathered into inflection tables. In total, there are 33 possible inflected forms and each inflection table will contain a subset of these forms. And example table is shown in Appendix C.",
                "cite_spans": [
                    {
                        "start": 199,
                        "end": 219,
                        "text": "Forbes et al. (2017)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "B Database of Gitksan Inflection Tables",
                "sec_num": null
            },
            {
                "text": "Note that other decision strategies such as randomly selecting an output form or taking the majority vote are also possible. These alternative strategies consistently underperform MAX, so we exclude them from the main text.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Note that the resulting paradigms are not necessarily free of dialectal variation. For instance, a paradigm where only the Western dialect form was observed for the ROOT and the Eastern dialect was observed for ROOT-3.II would still contain forms from both dialects.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "Strictly speaking, it will containly slightly more than 10%, since the last sampled paradigm may have more forms than the desired amount.4 More specifically, it has never seen the MSD:form pairs occurring in the training set.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "All code and results for this paper are available at: <anonymized for review>.6 Note that this means that we filter out identity pairs.7 We also predict from forms that belong to the training set if forms from paradigm \u03c0 were included in the training set, but we only evaluate performance on the forms in \u03a0wug and \u03a0 obs .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "https://scikit-learn.org/stable/",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "modules/generated/sklearn.metrics. average_precision_score.html",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            },
            {
                "text": "It could also explain why we don't see a greater increase in performance on the \u03a0wug test set with the augmented model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "A Gitksan inflection table for jok ('to dwell') generated from IGT and displayed in TSV format. Each row in the table contains five cells: (1) a morphosyntactic description, (2) an English translation, (3) a gloss with an English lemma, (3) a canonical segmented output form, (4) the surfce word form, and (5) a gloss with a Gitksan lemma. Many cells in the table are empty since they were unattested in the IGT data.ROOT dwell jok jok jok ROOT-SX dwell-SX jok-it jogat jok-SX ROOT-SX dwell-SX jok-it jogot jok-SX ROOT-PL _ _ _ _ ROOT-3PL _ _ _ _ ROOT-ATTR _ _ _ _ ROOT-3.II dwell-3.II jok-t jokt jok-3.II ROOT-PL-SX PL~dwell-SX CVC~jok-it jaxjogat PL~jok-SX ROOT-PL-SX PL~dwell-SX CVC~jok-it jaxjogot PL~jok-SX ROOT-1SG.II dwell-1SG.II jok-'y jogo'y jok-1SG.II ROOT-2SG.II _ _ _ _ ROOT-2PL.II _ _ _ _ ROOT-3PL.II dwell-3PL.II jok-diit jokdiit jok-3PL.II ROOT-1PL.II dwell-1PL.II jok-'m jogo'm jok-1PL.II ROOT-PL-3PL _ _ _ _ ROOT-TR-3.II _ _ _ _ ROOT-PL-3.II PL~dwell-3.II CVC~jok-t jaxjokt PL~jok-3.II",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "C Sample inflection table",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Parts and wholes: Implicative patterns in inflectional paradigms",
                "authors": [
                    {
                        "first": "Farrell",
                        "middle": [],
                        "last": "Ackerman",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "James",
                        "suffix": ""
                    },
                    {
                        "first": "Robert",
                        "middle": [],
                        "last": "Blevins",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Malouf",
                        "suffix": ""
                    }
                ],
                "year": 2009,
                "venue": "Analogy in grammar: Form and acquisition",
                "volume": "",
                "issue": "",
                "pages": "54--82",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Farrell Ackerman, James P Blevins, and Robert Mal- ouf. 2009. Parts and wholes: Implicative patterns in inflectional paradigms. Analogy in grammar: Form and acquisition, pages 54-82.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "Pushing the limits of low-resource morphological inflection",
                "authors": [
                    {
                        "first": "Antonios",
                        "middle": [],
                        "last": "Anastasopoulos",
                        "suffix": ""
                    },
                    {
                        "first": "Graham",
                        "middle": [],
                        "last": "Neubig",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1908.05838"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Antonios Anastasopoulos and Graham Neubig. 2019. Pushing the limits of low-resource morphological inflection. arXiv preprint arXiv:1908.05838.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "The child's learning of english morphology",
                "authors": [
                    {
                        "first": "Jean",
                        "middle": [],
                        "last": "Berko",
                        "suffix": ""
                    }
                ],
                "year": 1958,
                "venue": "",
                "volume": "14",
                "issue": "",
                "pages": "150--177",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jean Berko. 1958. The child's learning of english mor- phology. WORD, 14:150-177.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Conllsigmorphon 2017 shared task: Universal morphological reinflection in 52 languages",
                "authors": [
                    {
                        "first": "Ryan",
                        "middle": [],
                        "last": "Cotterell",
                        "suffix": ""
                    },
                    {
                        "first": "Christo",
                        "middle": [],
                        "last": "Kirov",
                        "suffix": ""
                    },
                    {
                        "first": "John",
                        "middle": [],
                        "last": "Sylak-Glassman",
                        "suffix": ""
                    },
                    {
                        "first": "G\u00e9raldine",
                        "middle": [],
                        "last": "Walther",
                        "suffix": ""
                    },
                    {
                        "first": "Ekaterina",
                        "middle": [],
                        "last": "Vylomova",
                        "suffix": ""
                    },
                    {
                        "first": "Patrick",
                        "middle": [],
                        "last": "Xia",
                        "suffix": ""
                    },
                    {
                        "first": "Manaal",
                        "middle": [],
                        "last": "Faruqui",
                        "suffix": ""
                    },
                    {
                        "first": "Sandra",
                        "middle": [],
                        "last": "K\u00fcbler",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Yarowsky",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "CoNLL Shared Task",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ryan Cotterell, Christo Kirov, John Sylak-Glassman, G\u00e9raldine Walther, Ekaterina Vylomova, Patrick Xia, Manaal Faruqui, Sandra K\u00fcbler, David Yarowsky, Jason Eisner, and Mans Hulden. 2017. Conll- sigmorphon 2017 shared task: Universal morpholog- ical reinflection in 52 languages. In CoNLL Shared Task.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "Report on the status of BC First Nations languages. Report of the First People's Cultural Council",
                "authors": [
                    {
                        "first": "Britt",
                        "middle": [],
                        "last": "Dunlop",
                        "suffix": ""
                    },
                    {
                        "first": "Suzanne",
                        "middle": [],
                        "last": "Gessner",
                        "suffix": ""
                    },
                    {
                        "first": "Tracey",
                        "middle": [],
                        "last": "Herbert",
                        "suffix": ""
                    },
                    {
                        "first": "Aliana",
                        "middle": [],
                        "last": "Parker",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Britt Dunlop, Suzanne Gessner, Tracey Herbert, and Aliana Parker. 2018. Report on the status of BC First Nations languages. Report of the First People's Cultural Council. Retrieved March 24, 2019.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Teruko Mitamura, and Eduard Hovy. 2021. A survey of data augmentation approaches for NLP",
                "authors": [
                    {
                        "first": "Y",
                        "middle": [],
                        "last": "Steven",
                        "suffix": ""
                    },
                    {
                        "first": "Varun",
                        "middle": [],
                        "last": "Feng",
                        "suffix": ""
                    },
                    {
                        "first": "Jason",
                        "middle": [],
                        "last": "Gangal",
                        "suffix": ""
                    },
                    {
                        "first": "Sarath",
                        "middle": [],
                        "last": "Wei",
                        "suffix": ""
                    },
                    {
                        "first": "Soroush",
                        "middle": [],
                        "last": "Chandar",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Vosoughi",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021",
                "volume": "",
                "issue": "",
                "pages": "968--988",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Steven Y. Feng, Varun Gangal, Jason Wei, Sarath Chan- dar, Soroush Vosoughi, Teruko Mitamura, and Ed- uard Hovy. 2021. A survey of data augmentation approaches for NLP. In Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 968-988, Online. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Three Gitksan texts",
                "authors": [
                    {
                        "first": "Clarissa",
                        "middle": [],
                        "last": "Forbes",
                        "suffix": ""
                    },
                    {
                        "first": "Henry",
                        "middle": [],
                        "last": "Davis",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Schwan",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Ubc Gitksan Research",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Laboratory",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Papers for the 52nd International Conference on Salish and Neighbouring Languages",
                "volume": "",
                "issue": "",
                "pages": "47--89",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Clarissa Forbes, Henry Davis, Michael Schwan, and the UBC Gitksan Research Laboratory. 2017. Three Gitksan texts. In Papers for the 52nd International Conference on Salish and Neighbouring Languages, pages 47-89. UBC Working Papers in Linguistics.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "2021. (un) solving morphological inflection: Lemma overlap artificially inflates models' performance",
                "authors": [
                    {
                        "first": "Omer",
                        "middle": [],
                        "last": "Goldman",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Guriel",
                        "suffix": ""
                    },
                    {
                        "first": "Reut",
                        "middle": [],
                        "last": "Tsarfaty",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2108.05682"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Omer Goldman, David Guriel, and Reut Tsarfaty. 2021. (un) solving morphological inflection: Lemma over- lap artificially inflates models' performance. arXiv preprint arXiv:2108.05682.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Sequence-level mixed sample data augmentation",
                "authors": [
                    {
                        "first": "Demi",
                        "middle": [],
                        "last": "Guo",
                        "suffix": ""
                    },
                    {
                        "first": "Yoon",
                        "middle": [],
                        "last": "Kim",
                        "suffix": ""
                    },
                    {
                        "first": "Alexander",
                        "middle": [],
                        "last": "Rush",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
                "volume": "",
                "issue": "",
                "pages": "5547--5552",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Demi Guo, Yoon Kim, and Alexander Rush. 2020. Sequence-level mixed sample data augmentation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 5547-5552, Online. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Understanding morphology",
                "authors": [
                    {
                        "first": "Martin",
                        "middle": [],
                        "last": "Haspelmath",
                        "suffix": ""
                    },
                    {
                        "first": "Andrea",
                        "middle": [],
                        "last": "Sims",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Martin Haspelmath and Andrea Sims. 2013. Under- standing morphology. Routledge.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Lowresource G2P and P2G conversion with synthetic training data",
                "authors": [
                    {
                        "first": "Bradley",
                        "middle": [],
                        "last": "Hauer",
                        "suffix": ""
                    },
                    {
                        "first": "Ahmad",
                        "middle": [],
                        "last": "Amir",
                        "suffix": ""
                    },
                    {
                        "first": "Yixing",
                        "middle": [],
                        "last": "Habibi",
                        "suffix": ""
                    },
                    {
                        "first": "Arnob",
                        "middle": [],
                        "last": "Luan",
                        "suffix": ""
                    },
                    {
                        "first": "Grzegorz",
                        "middle": [],
                        "last": "Mallik",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Kondrak",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 17th SIGMOR-PHON Workshop on Computational Research in Phonetics, Phonology, and Morphology",
                "volume": "",
                "issue": "",
                "pages": "117--122",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Bradley Hauer, Amir Ahmad Habibi, Yixing Luan, Arnob Mallik, and Grzegorz Kondrak. 2020. Low- resource G2P and P2G conversion with synthetic training data. In Proceedings of the 17th SIGMOR- PHON Workshop on Computational Research in Pho- netics, Phonology, and Morphology, pages 117-122, Online. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Shuly Wintner, and Yulia Tsvetkov. 2021. Machine translation into low-resource language varieties",
                "authors": [
                    {
                        "first": "Sachin",
                        "middle": [],
                        "last": "Kumar",
                        "suffix": ""
                    },
                    {
                        "first": "Antonios",
                        "middle": [],
                        "last": "Anastasopoulos",
                        "suffix": ""
                    }
                ],
                "year": null,
                "venue": "ACL/IJCNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sachin Kumar, Antonios Anastasopoulos, Shuly Wint- ner, and Yulia Tsvetkov. 2021. Machine trans- lation into low-resource language varieties. In ACL/IJCNLP.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Bootstrapping techniques for polysynthetic morphological analysis",
                "authors": [
                    {
                        "first": "William",
                        "middle": [],
                        "last": "Lane",
                        "suffix": ""
                    },
                    {
                        "first": "Steven",
                        "middle": [],
                        "last": "Bird",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2005.00956"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "William Lane and Steven Bird. 2020. Bootstrapping techniques for polysynthetic morphological analysis. arXiv preprint arXiv:2005.00956.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Analogy models for neural word inflection",
                "authors": [
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Mans",
                        "middle": [],
                        "last": "Hulden",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "COLING",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "L. Liu and Mans Hulden. 2020. Analogy models for neural word inflection. In COLING.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Can a transformer pass the wug test? tuning copying bias in neural morphological inflection models",
                "authors": [
                    {
                        "first": "Ling",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Mans",
                        "middle": [],
                        "last": "Hulden",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:2104.06483"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Ling Liu and Mans Hulden. 2021. Can a transformer pass the wug test? tuning copying bias in neural morphological inflection models. arXiv preprint arXiv:2104.06483.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Right for the wrong reasons: Diagnosing syntactic heuristics in natural language inference",
                "authors": [
                    {
                        "first": "Tom",
                        "middle": [],
                        "last": "Mccoy",
                        "suffix": ""
                    },
                    {
                        "first": "Ellie",
                        "middle": [],
                        "last": "Pavlick",
                        "suffix": ""
                    },
                    {
                        "first": "Tal",
                        "middle": [],
                        "last": "Linzen",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics",
                "volume": "",
                "issue": "",
                "pages": "3428--3448",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Tom McCoy, Ellie Pavlick, and Tal Linzen. 2019. Right for the wrong reasons: Diagnosing syntactic heuris- tics in natural language inference. In Proceedings of the 57th Annual Meeting of the Association for Com- putational Linguistics, pages 3428-3448, Florence, Italy. Association for Computational Linguistics.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Machine learning: a probabilistic perspective",
                "authors": [
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Kevin",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Murphy",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kevin P Murphy. 2012. Machine learning: a probabilis- tic perspective. MIT press.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "fairseq: A fast, extensible toolkit for sequence modeling",
                "authors": [
                    {
                        "first": "Myle",
                        "middle": [],
                        "last": "Ott",
                        "suffix": ""
                    },
                    {
                        "first": "Sergey",
                        "middle": [],
                        "last": "Edunov",
                        "suffix": ""
                    },
                    {
                        "first": "Alexei",
                        "middle": [],
                        "last": "Baevski",
                        "suffix": ""
                    },
                    {
                        "first": "Angela",
                        "middle": [],
                        "last": "Fan",
                        "suffix": ""
                    },
                    {
                        "first": "Sam",
                        "middle": [],
                        "last": "Gross",
                        "suffix": ""
                    },
                    {
                        "first": "Nathan",
                        "middle": [],
                        "last": "Ng",
                        "suffix": ""
                    },
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Grangier",
                        "suffix": ""
                    },
                    {
                        "first": "Michael",
                        "middle": [],
                        "last": "Auli",
                        "suffix": ""
                    }
                ],
                "year": 2019,
                "venue": "NAACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Myle Ott, Sergey Edunov, Alexei Baevski, Angela Fan, Sam Gross, Nathan Ng, David Grangier, and Michael Auli. 2019. fairseq: A fast, extensible toolkit for sequence modeling. In NAACL.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Scikit-learn: Machine learning in Python",
                "authors": [
                    {
                        "first": "F",
                        "middle": [],
                        "last": "Pedregosa",
                        "suffix": ""
                    },
                    {
                        "first": "G",
                        "middle": [],
                        "last": "Varoquaux",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Gramfort",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Michel",
                        "suffix": ""
                    },
                    {
                        "first": "B",
                        "middle": [],
                        "last": "Thirion",
                        "suffix": ""
                    },
                    {
                        "first": "O",
                        "middle": [],
                        "last": "Grisel",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Blondel",
                        "suffix": ""
                    },
                    {
                        "first": "P",
                        "middle": [],
                        "last": "Prettenhofer",
                        "suffix": ""
                    },
                    {
                        "first": "R",
                        "middle": [],
                        "last": "Weiss",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Dubourg",
                        "suffix": ""
                    },
                    {
                        "first": "J",
                        "middle": [],
                        "last": "Vanderplas",
                        "suffix": ""
                    },
                    {
                        "first": "A",
                        "middle": [],
                        "last": "Passos",
                        "suffix": ""
                    },
                    {
                        "first": "D",
                        "middle": [],
                        "last": "Cournapeau",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Brucher",
                        "suffix": ""
                    },
                    {
                        "first": "M",
                        "middle": [],
                        "last": "Perrot",
                        "suffix": ""
                    },
                    {
                        "first": "E",
                        "middle": [],
                        "last": "Duchesnay",
                        "suffix": ""
                    }
                ],
                "year": 2011,
                "venue": "Journal of Machine Learning Research",
                "volume": "12",
                "issue": "",
                "pages": "2825--2830",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duch- esnay. 2011. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12:2825-2830.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Data augmentation for transformer-based G2P",
                "authors": [
                    {
                        "first": "Zach",
                        "middle": [],
                        "last": "Ryan",
                        "suffix": ""
                    },
                    {
                        "first": "Mans",
                        "middle": [],
                        "last": "Hulden",
                        "suffix": ""
                    }
                ],
                "year": 2020,
                "venue": "Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology",
                "volume": "",
                "issue": "",
                "pages": "184--188",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zach Ryan and Mans Hulden. 2020. Data augmentation for transformer-based G2P. In Proceedings of the 17th SIGMORPHON Workshop on Computational Research in Phonetics, Phonology, and Morphology, pages 184-188, Online. Association for Computa- tional Linguistics.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Improving neural machine translation models with monolingual data",
                "authors": [
                    {
                        "first": "Rico",
                        "middle": [],
                        "last": "Sennrich",
                        "suffix": ""
                    },
                    {
                        "first": "Barry",
                        "middle": [],
                        "last": "Haddow",
                        "suffix": ""
                    },
                    {
                        "first": "Alexandra",
                        "middle": [],
                        "last": "Birch",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of the 54th",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016. Improving neural machine translation models with monolingual data. In Proceedings of the 54th",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Annual Meeting of the Association for Computational Linguistics",
                "authors": [],
                "year": null,
                "venue": "",
                "volume": "1",
                "issue": "",
                "pages": "86--96",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 86-96, Berlin, Germany. Association for Computational Lin- guistics.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "An encoder-decoder approach to the paradigm cell filling problem",
                "authors": [
                    {
                        "first": "Miikka",
                        "middle": [],
                        "last": "Silfverberg",
                        "suffix": ""
                    },
                    {
                        "first": "Mans",
                        "middle": [],
                        "last": "Hulden",
                        "suffix": ""
                    }
                ],
                "year": 2018,
                "venue": "EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Miikka Silfverberg and Mans Hulden. 2018. An encoder-decoder approach to the paradigm cell filling problem. In EMNLP.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Data augmentation for morphological reinflection",
                "authors": [
                    {
                        "first": "Miikka",
                        "middle": [],
                        "last": "Silfverberg",
                        "suffix": ""
                    },
                    {
                        "first": "Adam",
                        "middle": [],
                        "last": "Wiemerslage",
                        "suffix": ""
                    },
                    {
                        "first": "L",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Lingshuang Jack",
                        "middle": [],
                        "last": "Mao",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Miikka Silfverberg, Adam Wiemerslage, L. Liu, and Lingshuang Jack Mao. 2017. Data augmentation for morphological reinflection. In CoNLL.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Applying the transformer to character-level transduction",
                "authors": [
                    {
                        "first": "Shijie",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Ryan",
                        "middle": [],
                        "last": "Cotterell",
                        "suffix": ""
                    },
                    {
                        "first": "Mans",
                        "middle": [],
                        "last": "Hulden",
                        "suffix": ""
                    }
                ],
                "year": 2021,
                "venue": "EACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shijie Wu, Ryan Cotterell, and Mans Hulden. 2021. Ap- plying the transformer to character-level transduction. In EACL.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF1": {
                "num": null,
                "uris": null,
                "type_str": "figure",
                "text": "Predicting a plural inflection for a lexeme using two possible source forms (singular stem and plural stem). (a) A Transformer model trained with data hallucination prefers the plural form as the source (depicted by a thicker arrow, representing model confidence). (b)"
            },
            "FIGREF2": {
                "num": null,
                "uris": null,
                "type_str": "figure",
                "text": "Performances of the augmented and standard models using the MAX decision strategy on \u03a0 obs and \u03a0 wug test sets."
            },
            "TABREF0": {
                "type_str": "table",
                "content": "<table><tr><td>: A partial paradigm for the word meaning \"name\" in Gitksan. The paradigm has two entries (ROOT and ROOT-3.II) that each have two dialectal variants attested in the data. Four different one-to-one (MSD to Form) realizations of the paradigm are possible.</td></tr></table>",
                "text": "",
                "num": null,
                "html": null
            },
            "TABREF1": {
                "type_str": "table",
                "content": "<table/>",
                "text": "Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. 2018. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations. Xingyuan Zhao, Satoru Ozaki, Antonios Anastasopoulos, Graham Neubig, and Lori Levin. 2020. Automatic interlinear glossing for under-resourced languages leveraging translations. In Proceedings of the 28th International Conference on Computational Linguistics, pages 5397-5408, Barcelona, Spain (Online). International Committee on Computational Linguistics.",
                "num": null,
                "html": null
            }
        }
    }
}