JBNU-CCLab at SemEval-2022 Task 12: Machine Reading Comprehension and Span Pair Classification for Linking Mathematical Symbols to Their Descriptions

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1679 - 1686 July 14-15, 2022 ©2022 Association for Computational Linguistics JBNU-CCLab at SemEval-2022 Task 12: Machine Reading Comprehension and Span Pair Classification for Linking Mathematical Symbols to Their Descriptions Sung-Min Lee and Seung-Hoon Na Division of Computer Science and Engineering, Jeonbuk National University, South Korea {cap1232,nash}@jbnu.ac.kr Abstract This paper describes our system in the SemEval-2022 Task 12: ‘linking mathematical symbols to their descriptions’, achieving first on the leaderboard for all the subtasks compris- ing named entity extraction (NER) and rela- tion extraction (RE). Our system is a two-stage pipeline model based on SciBERT that detects symbols, descriptions, and their relationships in scientific documents. The system consists of 1) machine reading comprehension(MRC)-based NER model, where each entity type is repre- sented as a question and its entity mention span is extracted as an answer using an MRC model, and 2) span pair classification for RE, where two entity mentions and their type markers are encoded into span representations that are then fed to a Softmax classifier. In addition, we de- ploy a rule-based symbol tokenizer to improve the detection of the exact boundary of symbol entities. Regularization and ensemble methods are further explored to improve the RE model. 1 1 Introduction Mathematical symbols and descriptions appear in various forms across document section boundaries without explicit markups, and mathematical sym- bols appear in the form of long texts. Thus, link- ing mathematical symbols and their descriptions is challenging. SemEval 2022 task 12: ’linking mathematical symbols to their descriptions (Lai et al., 2022a)’, is a relation extraction task targeted at scientific doc- uments divided into two sub-tasks: sub-task A is a named entity recognition (NER) task that aims to predict the span of symbols and descriptions, and sub-task B is a relation extraction (RE) task that aims to predict relations between symbols and descriptions. 1Our code is publicly available at https://github. com/ZIZUN/symlink. Extracting these entities and relations is done to discover relational facts from unstructured texts. This problem can be decomposed into NER (Tjong Kim Sang and De Meulder, 2003; Ratinov and Roth, 2009) and RE (Zelenko et al., 2002; Bunescu and Mooney, 2005). Early works employed a two- stage relation extraction system, training one model to extract entities (Florian et al., 2004) and another model to classify relations between these entities (Zhou et al., 2005; Chan and Roth, 2011). To reduce the error propagation of NER or better capture the interactions between NER and RE, joint models have been proposed as a promising approach that are based on an end-to-end method or on the setting of multi-task learning using shared representations (Wadden et al., 2019; Lin et al., 2020; Wang and Lu, 2020). Recently, it has been observed that RE based on the shared encoder is suboptimal, but the use of separated encoders for NER and RE has shown improved performance compared to shared en- coders, reexamining the effectiveness of the simple pipelined two-stage approach (Zhong and Chen, 2021; Ye …

Content truncated to 500 words