UO UA: Using Latent Semantic Analysis to Build a Domain-Dependent Sentiment Resource

Abstract

In this paper we present our contribution to SemEval-2014 Task 4: Aspect Based Sentiment Analysis (Pontiki et al., 2014), Subtask 2: Aspect Term Polarity for Laptop domain. The most outstanding feature in this contribution is the automatic building of a domain-depended sentiment resource using Latent Semantic Analysis. We induce, for each term, two real scores that indicate its use in positive and negative contexts in the domain of interest. The aspect term polarity classification is carried out in two phases: opinion words extraction and polarity classification. The opinion words related with an aspect are obtained using dependency relations. These relations are provided by the Stanford Parser1. Finally, the polarity of the feature, in a given review, is determined from the positive and negative scores of each word related to it. The results obtained by our approach are encouraging if we consider that the construction of the polarity lexicon is performed fully automatically.