UoM: Using Explicit Semantic Analysis for Classifying Sentiments

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 535–538, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UoM: Using Explicit Semantic Analysis for Classifying Sentiments Sapna Negi University of Malta Msida, MSD2080, MALTA sapna.negi13@gmail.com Mike Rosner University of Malta Msida, MSD2080, MALTA mike.rosner@um.edu.mt Abstract In this paper, we describe our system sub- mitted for the Sentiment Analysis task at Se- mEval 2013 (Task 2). We implemented a com- bination of Explicit Semantic Analysis (ESA) with Naive Bayes classifier. ESA represents text as a high dimensional vector of explicitly defined topics, following the distributional se- mantic model. This approach is novel in the sense that ESA has not been used for Senti- ment Analysis in the literature, to the best of our knowledge. 1 Introduction Semantic relatedness measure gives the comparison of different terms or texts on the basis of their meaning or the content. For instance, it can be said that the word ”computer” is semantically more related to ”laptop” than ”flute”. Sentiment analysis refers to the task of determining the overall contextual polarity of the written text. In this paper, we propose the use of semantic relatedness models, specifically Explicit Semantic Analysis (ESA), to identify textual polarity. There are different approaches to model semantic relatedness like WordNet based models (Banerjee and Banerjee, 2002), distributional semantic models (DSMs) etc. DSMs follow the distributional hypothesis, which says that words occurring in the same contexts tend to have similar meanings (Harris, 1954). There- fore, considering sentiment classification problem, distributional hypothesis suggests that the words or phrases referring to positive polarity would tend to co-occur, and similar assumptions can be made for the negative terms. DSMs generally utilize large textual corpora to extract the distributional information relying on the co-occurrence information and distribution of the terms. These models represent the text in the form of high-dimensional vectors highlighting the co-occurrence information. Semantic relatedness between two given texts is calculated by using these vectors, thus, following that the the semantic meaning of a text can be inferred from its usage in different contexts. There are several different computational models following distributional semantics hypothesis. Latent Semantic Analysis (LSA), Latent Dirichlet Allocation (LDA) (Blei et. al., 2003), Explicit Semantic Analysis (ESA) are some examples of such models. However, in this work, we investigated the use of ESA for the given task of sentiment analysis (SA). There are two sub-tasks defined in Task 2 at SemEval 2013 (SemEval, 2013). We participated in Message Polarity Classification sub-task, where we are required to automatically classify the sentiment of a given message into positive, negative, or neutral. The task deals with the short texts coming from Twitter and SMS (Short Message Service). We are provided with 8,000 - 12,000 twitter messages annotated with their sentiment label for the purpose of training the models. In this work, we present our approach for sentiment classification which uses a combination of ESA and Naive Bayes classifier. The rest of …

Content truncated to 500 words