UIUC_BioNLP at SemEval-2021 Task 11: A Cascade of Neural Models for Structuring Scholarly NLP Contributions

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 377–386 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 377 UIUC BioNLP at SemEval-2021 Task 11: A Cascade of Neural Models for Structuring Scholarly NLP Contributions Haoyang Liu, Janina Sarol, and Halil Kilicoglu School of Information Sciences, University of Illinois at Urbana-Champaign {hl57, mjsarol, halil}@illinois.edu Abstract We propose a cascade of neural models that performs sentence classification, phrase recog- nition, and triple extraction to automatically structure the scholarly contributions of NLP publications in English. To identify the most important contribution sentences in a paper, we used a BERT-based classifier with posi- tional features (Subtask 1). A BERT-CRF model was used to recognize and character- ize relevant phrases in contribution sentences (Subtask 2). We categorized the triples into several types based on whether and how their elements were expressed in text, and addressed each type using separate BERT-based classi- fiers as well as rules (Subtask 3). Our sys- tem was officially ranked second in Phase 1 evaluation and first in both parts of Phase 2 evaluation. After fixing a submission error in Phase 1, our approach yielded the best re- sults overall. In this paper, in addition to a system description, we also provide fur- ther analysis of our results, highlighting its strengths and limitations. We make our code publicly available at https://github.com/ Liu-Hy/nlp-contrib-graph. 1 Introduction With the deluge of scientific publications in recent years, keeping pace with the literature and manag- ing information overload have become increasingly challenging for researchers. There is a growing need for tools that can automatically extract and structure semantic information from scientific pub- lications to facilitate advanced approaches to infor- mation access and knowledge curation (Shen et al., 2018). The field of natural language processing (NLP) has witnessed an enormous growth in recent years with advances in deep learning, and there are increasing efforts in developing methods to extract scholarly knowledge from NLP pub- lications (QasemiZadeh and Schumann, 2016; D’Souza and Auer, 2020b). One such effort is NLPCONTRIBUTIONS, an annotation scheme for describing the scholarly contributions in NLP pub- lications and a corpus annotated using this anno- tation scheme (D’Souza and Auer, 2020b). This corpus has been proposed for training and test- ing of machine reading models, whose output can be integrated with the Open Research Knowledge Graph framework (ORKG) (Jaradeh et al., 2019). ORKG formalizes the research contributions of a scholarly publication as a knowledge graph, which can further be linked to other publications via the graph. The goal of the NLPContributionGraph (NCG) shared task (D’Souza et al., 2021) is to fa- cilitate the development of machine reading models that can extract ORKG-compatible scholarly con- tribution information from NLP publications. The shared task consists of three subtasks (see D’Souza et al. (2021) for a more detailed description): • Subtask 1: Identification of contribution sen- tences from NLP publications • Subtask 2: Recognition of scientific terms and relations in contribution sentences • Subtask 3: Extraction and classification of triples that pair …

Content truncated to 500 words