Representing Verbs with Rich Contexts: an Evaluation on Verb Similarity

Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing, pages 1967–1972, Austin, Texas, November 1-5, 2016. c⃝2016 Association for Computational Linguistics Representing Verbs with Rich Contexts: an Evaluation on Verb Similarity Emmanuele Chersoni Aix-Marseille University emmanuelechersoni@gmail.com Enrico Santus The Hong Kong Polytechnic University esantus@gmail.com Alessandro Lenci University of Pisa alessandro.lenci@unipi.it Philippe Blache Aix-Marseille University philippe.blache@univ-amu.fr Chu-Ren Huang The Hong Kong Polytechnic University churen.huang@polyu.edu.hk Abstract Several studies on sentence processing sug- gest that the mental lexicon keeps track of the mutual expectations between words. Current DSMs, however, represent context words as separate features, thereby loosing important information for word expectations, such as word interrelations. In this paper, we present a DSM that addresses this issue by defining verb contexts as joint syntactic dependencies. We test our representation in a verb similarity task on two datasets, showing that joint con- texts achieve performances comparable to sin- gle dependencies or even better. Moreover, they are able to overcome the data sparsity problem of joint feature spaces, in spite of the limited size of our training corpus. 1 Introduction Distributional Semantic Models (DSMs) rely on the Distributional Hypothesis (Harris, 1954; Sahlgren, 2008), stating that words occurring in similar con- texts have similar meanings. On such theoretical grounds, word co-occurrences extracted from cor- pora are used to build semantic representations in the form of vectors, which have become very popular in the NLP community. Proximity between word vec- tors is taken as an index of meaning similarity, and vector cosine is generally adopted to measure such proximity, even though other measures have been proposed (Weeds et al., 2004; Santus et al., 2016). Most of DSMs adopt a bag-of-words approach, that is they turn a text span (i.e., a word window or a parsed sentence) into a set of words and they regis- ter separately the co-occurrence of each word with a given target. The problem with this approach is that valuable information concerning word interrelations in a context gets lost, because words co-occurring with a target are treated as independent features. This is why works like Ruiz-Casado et al. (2005), Agirre et al. (2009) and Melamud et al. (2014) pro- posed to introduce richer contexts in distributional spaces, by using entire word windows as features. These richer contexts proved to be helpful to seman- tically represent verbs, which are characterized by highly context-sensitive meanings, and complex ar- gument structures. In fact, two verbs may share in- dependent words as features despite being very dis- similar from the semantic point of view. For instance kill and heal share the same object nouns in The doc- tor healed the patient and the The poison killed the patient, but are highly different if we consider their joint dependencies as a single context. Nonetheless, richer contexts like these suffer from data sparsity, therefore requiring either larger corpora or complex smoothing processes. In this paper, we propose a syntactically savvy no- tion of joint contexts. To test our representation, we implement several DSMs and we evaluate them …

Content truncated to 500 words