Improving Scientific Relation Classification with Task Specific Supersense

Improving Scientific Relation Classification with Task Specific Supersense Qin Dai1, Naoya Inoue1,2, Paul Reisert2, Kentaro Inui1,2 1Tohoku University, Japan 2RIKEN Center for Advanced Intelligence Project, Japan {daiqin, naoya-i, preisert, inui}@ecei.tohoku.ac.jp Abstract Classifying the relationship between entities is an important natural language processing (NLP) task. Scientific Relation Classification aims at automatically categorizing scientific semantic relationships among entities in sci- entific documents. Conventionally, only task unspecific supersense, such as supersense (or hyernym) from WordNet (e.g., ANIMAL is the supersense of “dog”), is used as a feature for re- lation classification. In this work, we hypothe- size that task specific supersense could also be utilized as an informative feature for relation classification. Specifically, we define a new entity type based on the property of a given task, and facilitate scientific relation classifi- cation with the task specific supersense. Our experiments on three different datasets prove the effectiveness of the task specific supersense on relation classification in scientific articles. 1 Introduction In recent years, along with the number of scientific papers increasing, it is prohibitively time-consuming for researchers to review and fully-comprehend all papers. To effectively and quickly access a large amount of scientific papers and acquire useful knowl- edge, a wide variety of computational studies for structuralizing scientific papers has been conducted, such as Argumentative Zoning (Teufel and oth- ers, 2000), BioNLP Shared Task (2017), ScienceIE Shared Task (Augenstein et al., 2017) and Semantic Relation Extraction and Classification in Scientific Papers (Gábor et al., 2018). One fundamental study is Relation Classification (RC). In this paper, we tackle the task of RC. RC is the task of capturing predefined semantic relations be- tween entities from text. Thus, our task consists of the following: given a sentence that has been an- notated with entity1 mentions, we aim towards cat- egorizing relations between entities. Suppose the following sentence: (1) An efficient entity bit-vector-based CKY-style parserX for entity context-free parsingY is presented. In Example 1, one of the scientific relations we aim to identify is the relation USAGE(X, Y), which means that bit-vector-based CKY-style parser is used for the action of context-free parsing. For notational convenience, we refer to a sentence where a relation is identified as a target sentence, and we refer to the related entity pair as a target entity pair. Many previous works on RC exist in the general domain (Kumar, 2017; Zhou et al., 2014). The ear- lier approaches depend on complex feature engineer- ing such as manually prepared lexical-syntactic pat- terns (Boschee et al., 2005; Suchanek et al., 2006; Chan and Roth, 2010, etc.). Recently, Neural Net- work (NN)-based approaches achieve close or even better performance to earlier approaches without complicated manually prepared features (Zeng et al., 2014; Zhang and Wang, 2015; Santos et al., 2015). In the context of scientific RC, Ammar et al. (Am- mar et al., 2017) enhanced Miwa and Bansal (Miwa 1In this work, entity refers not merely to concepts denoted by noun or noun phrase, it could be actions denoted by verb or verb phrase, and evaluation …

Content truncated to 500 words