| "I'm happy to share an experimental dataset and model for clustering and classifying sentences from academic texts based on their rhetorical functions (RF), i.e the communicative function they serve (summarizing results, expressing limitations etc.) | |
| This builds upon work by Kenichi Iwatsuki, Akiko Aizawa and others who first proposed the use of Academic Phrasebank (AP) by the University of Manchester as a extensive source of RF that can be found in academic texts. https://iwa2ki.com/FE/#SDU2021 | |
| My work builds on theirs by building up a new dataset of rhetorical function labelled sentences from scratch, through the use of a performant LLM and carefully structured prompts. These prompts utilized the existing RF examples from the AP along with 819 article titles as seed phrases. | |
| These titles were chosen to include multiple articles from each of the 211 scientific fields included in the second level of OpenAlexs Concept Hierarchy. This serves as a diverse set of data covering a variety of topics while providing >800 examples for each RF, which I term SciRhetoriSet. | |
| For producing RF contextualized embeddings I chose nomic-ai/nomic-embed-text-v1.5 as a fast sota model for these tasks. Evaluation was structured as a retrieval task, where the goal was to retrieve the correct sentence function as the highest ranked item by cosine similarity. | |
| With this performance, I'm excited about the potential for downstream applications for this model including for information extraction, summarisation, topic Modelling and knowledge gap analysis. | |
| For the dataset, model and demo visit: | |
| https://huggingface.co/collections/KaiserML/rhetoribert-679413accfbfe827f16cb63f | |
| Meanwhile, I intend to shortly release a proper writeup and later intend to improve upon the diversity of seed titles to ensure that the model can be used across the scientific landscape and refine the term list used. With this I welcome any suggestions for improvements! Thanks for your time." | |
| Current science communication has a number of drawbacks and bottlenecks which have been subject of discussion lately: Among others, the rising number of published articles makes it nearly impossible to get a full overview of the state of the art in a certain field, or reproducibility is hampered by fixed-length, document-based publications which normally cannot cover all details of a research work. Recently, several initiatives have proposed knowledge graphs (KG) for organising scientific information as a solution to many of the current issues. The focus of these proposals is, however, usually restricted to very specific use cases. In this paper, we aim to transcend this limited perspective and present a comprehensive analysis of requirements for an Open Research Knowledge Graph (ORKG) by (a) collecting and reviewing daily core tasks of a scientist, (b) establishing their consequential requirements for a KG-based system, (c) identifying overlaps and specificities, and their coverage in current solutions. As a result, we map necessary and desirable requirements for successful KG-based science communication, derive implications, and outline possible solutions. | |
| "We introduce a new language representation model called BERT, which stands for Bidirectional Encoder Representations from Transformers. Unlike recent language representation models, BERT is designed to pre-train deep bidirectional representations from unlabeled text by jointly conditioning on both left and right context in all layers. As a result, the pre-trained BERT model can be fine-tuned with just one additional output layer to create state-of-the-art models for a wide range of tasks, such as question answering and language inference, without substantial task-specific architecture modifications. | |
| BERT is conceptually simple and empirically powerful. It obtains new state-of-the-art results on eleven natural language processing tasks, including pushing the GLUE score to 80.5% (7.7% point absolute improvement), MultiNLI accuracy to 86.7% (4.6% absolute improvement), SQuAD v1.1 question answering Test F1 to 93.2 (1.5 point absolute improvement) and SQuAD v2.0 Test F1 to 83.1 (5.1 point absolute improvement)." | |