Apollo at SemEval-2018 Task 9: Detecting Hypernymy Relations Using Syntactic Dependencies

Proceedings of the 12th International Workshop on Semantic Evaluation (SemEval-2018), pages 898–902 New Orleans, Louisiana, June 5–6, 2018. ©2018 Association for Computational Linguistics 1 Apollo at SemEval-2018 Task 9: Detecting Hypernymy Relations Using Syntactic Dependencies Mihaela Onofrei, Ionuț Hulub, Diana Trandabăț, Daniela Gîfu University Alexandru Ioan Cuza of Iași, Romania Institute of Computer Science of the Romanian Academy, Iași Branch Cognos Business Consulting S.R.L., 32 Bd. Regina Maria, Bucharest, Romania {mihaela.onofrei, ionut.hulub, daniela.gifu, dtrandabat} @info.uaic.ro Abstract This paper presents the participation of Apollo’s team in the SemEval-2018 Task 9 “Hypernym Discovery”, Subtask 1: “Gen- eral-Purpose Hypernym Discovery”, which tries to produce a ranked list of hy- pernyms for a specific term. We propose a novel approach for automatic extraction of hypernymy relations from a corpus by us- ing dependency patterns. The results show that the application of these patterns leads to a higher score than using the traditional lexical patterns. Keywords: hypernymy relations, semantic relations, corpus, taxonomy, syntactic de- pendencies. 1 Introduction This paper presents the Apollo team’s system for hypernym discovery which participated in task 9 of Semeval 2018 (Camacho-Collados et al., 2018) based on unsupervised machine learning. It is a rule-based system that exploits syntactic depend- ency paths that generalize Hearst-style lexical patterns. The paper is structured in 4 sections: this sec- tion presents existing approaches for automatic extraction of hypernymy relations, Section 2 contains the current system architecture. The next section presents the web interface of the project, and, finally, Section 4 briefly analyses the results and drafts some conclusions. Since language is a “vital organ”, constantly evolving and changing over time, there are many words which lose one of their meanings or attach a new meaning. For instance, when searching the word “apple” in WordNet (Miller, 1995), it ap- pears defined as “fruit with red or yellow or green skin and sweet to tart crisp whitish flesh” and “native Eurasian tree widely cultivated in many varieties for its firm rounded edible fruits” but searching in British National Corpus1, we will remark that the term is used more frequently as a named entity (referring to a “company”). From this point of view, we consider that de- veloping a system for hypernym discovery that uses linguistic features from a corpus could be more useful for this task than using a manually- crafted taxonomy. It is well known that in natural language pro- cessing (NLP), one of the biggest challenges is to understand the meaning of words. Also, detecting hypernymy relations is an important task in NLP, which has been pursued for over two decades, and it is addressed in the literature using two comple- mentary approaches: rule-based and distributional 1 https://corpus.byu.edu/bnc/ 898 2 methods. Rule-based methods (Hearst, 1992; Snow et al., 2004) base the decision on the lexico- syntactic paths connecting the joint occurrences of two or more terms in a corpus. In the case of su- pervised distributional methods (Baroni et al., 2012; Roller et al., 2014, Weeds et al., 2014; Levy et al., 2015, Kruszewski et …

Content truncated to 500 words