Using Distributional Thesaurus Embedding for Co-hyponymy Detection

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5766–5771 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5766 Using Distributional Thesaurus Embedding for Co-hyponymy Detection Abhik Jana, Nikhil Reddy Varimalla and Pawan Goyal Universit¨at Hamburg, Indian Institute of Technology Kharagpur, Indian Institute of Technology Kharagpur jana@informatik.uni-hamburg.de, nikhil.varimala@gmail.com, pawang@cse.iitkgp.ac.in Abstract Discriminating lexical relations among distributionally similar words has always been a challenge for natural language processing (NLP) community. In this paper, we investigate whether the network embedding of distributional thesaurus can be effectively utilized to detect co-hyponymy relations. By extensive experiments over three benchmark datasets, we show that the vector representation obtained by applying node2vec on distributional thesaurus outperforms the state-of-the-art models for binary classification of co-hyponymy vs. hypernymy, as well as co-hyponymy vs. meronymy, by huge margins. Keywords: Co-hyponymy detection, Distributional Thesaurus, Network Embedding 1. Introduction Distributional semantic models are used in a wide variety of tasks like sentiment analysis, word sense disambigua- tion, predicting semantic compositionality, etc. Automatic detection of lexical relations is one such fundamental task which can be leveraged in applications like paraphrasing, ontology building, metaphor detection etc. Both super- vised and unsupervised methods have been proposed by the researchers to identify lexical relations like hypernymy, co-hyponymy, meronymy etc. over the years. Recent at- tempts to solve this task deal with proposing similarity measures based on distributional semantic models (Roller et al., 2014; Weeds et al., 2014; Santus et al., 2016; Shwartz et al., 2017; Roller and Erk, 2016). For hypernymy de- tection, several works use distributional inclusion hypothe- sis (Geffet and Dagan, 2005), entropy-based distributional measure (Santus et al., 2014) as well as several embed- ding schemes (Fu et al., 2014; Yu et al., 2015; Nguyen et al., 2017). Image generality for lexical entailment de- tection (Kiela et al., 2015) has also been tried out for the same purpose. As far as meronymy detection is concerned, most of the attempts are pattern based (Berland and Char- niak, 1999; Girju et al., 2006; Pantel and Pennacchiotti, 2006) along with some recent works exploring the pos- sibility of using distributional semantic models (Morlane- Hond`ere, 2015). Similarly, for co-hyponymy detection, researchers have in- vestigated the usefulness of several distributional semantic models. One such attempt is made by Weeds et al. (2014), where they proposed a supervised framework and used sev- eral vector operations as features for the classification of hypernymy and co-hyponymy. Santus et al. (2016) pro- posed a supervised method based on a Random Forest al- gorithm to learn taxonomical semantic relations and they have shown that the model performs well for co-hyponymy detection. In another attempt, Jana and Goyal (2018b) pro- posed various complex network measures which can be used as features to build a supervised classifier model for co-hyponymy detection, and showed improvements over other baseline approaches. Recently, with the emergence of various network representation learning methods (Perozzi et al., 2014; Tang et al., 2015; Grover and Leskovec, 2016; Ribeiro et al., 2017), attempts have been …

Content truncated to 500 words