Network Features Based Co-hyponymy Detection

Network Features Based Co-hyponymy Detection Abhik Jana, Pawan Goyal Indian Institute of Technology Kharagpur, Indian Institute of Technology Kharagpur abhik.jana@iitkgp.ac.in, pawang@cse.iitkgp.ac.in Abstract Distinguishing lexical relations has been a long term pursuit in natural language processing (NLP) domain. Recently, in order to detect lexical relations like hypernymy, meronymy, co-hyponymy etc., distributional semantic models are being used extensively in some form or the other. Even though a lot of efforts have been made for detecting hypernymy relation, the problem of co-hyponymy detection has been rarely investigated. In this paper, we are proposing a novel supervised model where various network measures have been utilized to identify co-hyponymy relation with high accuracy performing better or at par with the state-of-the-art models. Keywords: Co-hyponymy detection, Distributional thesaurus network, Complex network measures. 1. Introduction Automatic detection of lexical relations is a fundamental task for natural language processing (NLP). Numerous ap- plications including paraphrasing, query expansion, recog- nizing textual entailment, ontology building, metaphor de- tection etc. are benefited by precise relation classification and relation discovery tasks. For example, it may be diffi- cult to interpret a sentence containing a metaphor, like “He drowned in a sea of grief” if we go by the literal mean- ing. But if we replace ‘drowned’ by its co-hyponym ‘over- whelmed’ and ‘sea’ by its co-hyponym ‘lot’, it immediately provides an inference. Note that, ‘drown’ and ‘overwhelm’ are (co-)hyponyms for the concept ‘cover’ whereas ‘sea’ and ‘lot’ are (co-)hyponyms for the concept ‘large indefi- nite amount’ as per WordNet (Miller, 1995). Lexical relations are of variety of types like hyponyms, hypernyms, co-hyponyms, meronyms etc. Among these, some relations are symmetric (co-hyponymy) and some are asymmetric (hypernymy, meronymy). With the advance- ment of distributional semantics representation of words, researchers have attempted to identify lexical relations in both supervised and unsupervised ways. One of the oldest attempt for detection of hypernymy ex- traction dealt with finding out ‘lexico-syntactic patterns’ proposed by Hearst (1992). A lot of attempts have been made for hypernymy extraction using knowledge bases like Wordnet, Wikipedia and hand crafted patterns or patterns learnt from the corpus (Cederberg and Widdows, 2003; Ya- mada et al., 2009). With the emergence of the trend of applying distributional hypothesis (Firth, 1957) to solve this relation classification task, researchers have started us- ing Distributional Semantic Models (DSM) and have come up with several directional measures (Roller et al., 2014; Weeds et al., 2014; Santus et al., 2016; Shwartz et al., 2017; Roller and Erk, 2016). Specifically for hypernymy detection, researchers also used a variant of distributional hypothesis, i.e., distributional inclusion hypothesis (Gef- fet and Dagan, 2005) according to which the contexts of a narrow term are also shared by the broad term. Recently, entropy-based distributional measure (Santus et al., 2014) has also been tried out for the same purpose. In some of the recent attempts (Fu et al., 2014; Yu et al., 2015; Nguyen et al., 2017), people have tried several embed- ding schemes for hypernymy detection. One interesting attempt was made by Kiela et al. …

Content truncated to 500 words