Taking Antonymy Mask off in Vector Space

PACLIC 28 !135 Taking Antonymy Mask off in Vector Space Enrico Santus Qin Lu The Hong Kong Polytechnic University The Hong Kong Polytechnic University Hong Kong Hong Kong e.santus@connect.polyu.hk csluqin@comp.polyu.edu.hk Alessandro Lenci Chu-Ren Huang Università di Pisa The Hong Kong Polytechnic University Pisa, Italy Hong Kong alessandro.lenci@ling.unipi.it churen.huang@polyu.edu.hk Abstract Automatic detection of antonymy is an important task in Natural Language Processing (NLP) for Information Retrieval (IR), Ontology Learning (OL) and many other semantic applications. However, current unsupervised approaches to antonymy detection are still not fully effective because they cannot discriminate antonyms from synonyms. In this paper, we introduce APAnt, a new Average- Precision-based measure for the unsupervised discrimination of antonymy from synonymy using Distributional Semantic Models (DSMs). APAnt makes use of Average Precision to estimate the extent and salience of the intersection among the most descriptive contexts of two target words. Evaluation shows that the proposed method is able to distinguish antonyms and synonyms with high accuracy across different parts of speech, including nouns, adjectives and verbs. APAnt outperforms the vector cosine and a baseline model implementing the co- occurrence hypothesis. 1 Introduction Antonymy is one of the fundamental relations shaping the organization of the semantic lexicon and its identification is very challenging for computational models (Mohammad et al., 2008; Deese, 1965; Deese, 1964). Yet, antonymy is essential for many Natural Language Processing (NLP) applications, such as Information Retrieval (IR), Ontology Learning (OL), Machine Translation (MT), Sentiment Analysis (SA) and Dialogue Systems (Roth and Schulte im Walde, 2014; Mohammad et al., 2013). In particular, the automatic identification of semantic opposition is a crucial component for the detection and generation of paraphrases (Marton et al., 2011), the understanding of contradictions (de Marneffe et al., 2008) and the detection of humor (Mihalcea and Strapparava, 2005). Several existing computational lexicons and thesauri explicitly encode antonymy, together with other semantic relations. Although such resources are often used to support the above mentioned NLP tasks, hand-coded lexicons and thesauri have low coverage and many scholars have shown their limits: Mohammad et al. (2013), for example, have noticed that “more than 90% of the contrasting pairs in GRE closest-to-opposite questions are not listed as opposites in WordNet”. Copyright 2014 by Enrico Santus, Qin Lu, Alessandro Lenci and Chu-Ren Huang 28th Pacific Asia Conference on Language, Information and Computation pages 135–144 PACLIC 28 !136 The automatic identification of semantic relations is a core task in computational semantics. Distributional Semantic Models (DSMs) have often been exploited for their well known ability to identify semantically similar lexemes using corpus-derived co-occurrences encoded as distributional vectors (Santus et al., 2014a; Baroni and Lenci, 2010; Turney and Pantel, 2010; Padó and Lapata, 2007; Sahlgren, 2006). These models are based on the Distributional Hypothesis (Harris, 1954) and represent lexical semantic similarity in function of distributional similarity, which can be measured by vector cosine (Turney and Pantel, 2010). However, these models are characterized by a major shortcoming. That is, they are not able to discriminate among different kinds of semantic relations linking distributionally similar …

Content truncated to 500 words