MELODI: Semantic Similarity of Words and Compositional Phrases using Latent Vector Weighting

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 98–102, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics MELODI: Semantic Similarity of Words and Compositional Phrases using Latent Vector Weighting Tim Van de Cruys IRIT, CNRS tim.vandecruys@irit.fr Stergos Afantenos IRIT, Toulouse University stergos.afantenos@irit.fr Philippe Muller IRIT, Toulouse University philippe.muller@irit.fr Abstract In this paper we present our system for the SemEval 2013 Task 5a on semantic similar- ity of words and compositional phrases. Our system uses a dependency-based vector space model, in combination with a technique called latent vector weighting. The system computes the similarity between a particular noun in- stance and the head noun of a particular noun phrase, which was weighted according to the semantics of the modifier. The system is en- tirely unsupervised; one single parameter, the similarity threshold, was tuned using the train- ing data. 1 Introduction In the course of the last two decades, vector space models have gained considerable momentum for se- mantic processing. Initially, these models only dealt with individual words, ignoring the context in which these words appear. More recently, two different but related approaches emerged that take into account the interaction between different words within a par- ticular context. The first approach aims at building a joint, compositional representation for larger units beyond the individual word level (e.g., the com- posed, semantic representation of the noun phrase crispy chips). The second approach, different but re- lated to the first one, computes the specific meaning of a word within a particular context (e.g. the mean- ing of the noun bank in the context of the adjective bankrupt). In this paper, we describe our system for the Sem- Eval 2013 Task 5a: semantic similarity of words and compositional phrases – which follows the latter ap- proach. Our system uses a dependency-based vector space model, in combination with a technique called latent vector weighting (Van de Cruys et al., 2011). The system computes the similarity between a par- ticular noun instance and the head noun of a par- ticular noun phrase, which was weighted according to the semantics of the modifier. The system is en- tirely unsupervised; one single parameter, the simi- larity threshold, was tuned using the training data. 2 Related work In recent years, a number of methods have been de- veloped that try to capture the compositional mean- ing of units beyond the individual word level within a distributional framework. One of the first ap- proaches to tackle compositional phenomena in a systematic way is Mitchell and Lapata’s (2008) ap- proach. They explore a number of different mod- els for vector composition, of which vector addition (the sum of each feature) and vector multiplication (the elementwise multiplication of each feature) are the most important. Baroni and Zamparelli (2010) present a method for the composition of adjectives and nouns. In their model, an adjective is a linear function of one vector (the noun vector) to another vector (the vector for …

Content truncated to 500 words