DSC IIT-ISM at SemEval-2020 Task 6: Boosting BERT with Dependencies for Definition Extraction

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 710–716 Barcelona, Spain (Online), December 12, 2020. 710 DSC IIT-ISM at SemEval-2020 Task 6: Boosting BERT with Dependencies for Definition Extraction Aadarsh Singh*, Priyanshu Kumar* and Aman Sinha Indian Institute of Technology (Indian School of Mines) Dhanbad, India {aadarshsingh191198, kpriyanshu256, amansinha091}@gmail.com Abstract We explore the performance of Bidirectional Encoder Representations from Transformers (BERT) at definition extraction. We further propose a joint model of BERT and Text Level Graph Convolutional Network so as to incorporate dependencies into the model. Our proposed model produces better results than BERT and achieves comparable results to BERT with fine tuned language model in DeftEval (Task 6 of SemEval 2020), a shared task of classifying whether a sentence contains a definition or not (Subtask 1). 1 Introduction Definition Extraction from free text (DEFT) (Spala et al., 2020) involves finding term definition pairs from free and semi-structured texts, especially those whose term-definition span crosses a sentence boundary and those which do not have a definition phrase. An example of a cross boundary sentence from the DEFT corpus is given below: In doing so , <DEF> monomers release water molecules as byproducts <DEF> (1). This type of reaction is known as <TERM> dehydration synthesis <TERM> , which means “ <QUALIFIER> to put together while losing water <QUALIFIER> . ”(0) It can be observed that the definition (enclosed by DEF tags) and the corresponding terms (enclosed by TERM tags) are present in different sentences, thus increasing the difficulty of definition extraction. The task is thus relatively different and complex from the conventional definition extraction (DE) task in which a definition could be broken into the following sub-parts: 1. The DEFINIENDUM field (DF) i.e., the word being defined. 2. The DEFINITOR field (VF) i.e., the verb phrase used to introduce the definition. 3. The DEFINIENS field (GF) i.e., genus phrase or the hypernym. 4. The REST field (RF) i.e., additional clauses that help to distinguish the definiendum from its genus. and thus easily be captured by common verb phrases (DEFINITOR) like “means”, “refers to”, “is”, etc. These kind of conventional definitions could easily be tagged as follows: <DF>Photosynthesis</DF> <VF>is</VF> <GF>the process </GF> <RF>by which green plants manufacture food. </RF> In the example presented above, the definition (REST field) and the corresponding term (DEFINIEN- DUM field) are present in the same sentence. Moreover, the presence of DEFINITOR(s) in the text also eases the task of extracting such definitions. * Equal contribution. This work is licensed under a Creative Commons Attribution 4.0 International Licence. Licence details: http:// creativecommons.org/licenses/by/4.0/ 711 The order of the tags in the WCL Corpus (Navigli et al., 2010), one of the conventional corpus for Definition Extraction task, is predefined viz. DEFINIENDUM, DEFINITOR, DEFINIENS, REST. On the other hand, there is no such predefined order for the DEFT corpus regarding the occurrence of BIO Tags. This absence of order makes it difficult for the models to identify the definitional sentences in the DEFT corpus and then associate the words …

Content truncated to 500 words