UNITOR: Combining Syntactic and Semantic Kernels for Twitter Sentiment Analysis

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 369–374, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics UNITOR: Combining Syntactic and Semantic Kernels for Twitter Sentiment Analysis Giuseppe Castellucci(†), Simone Filice(‡), Danilo Croce(⋆), Roberto Basili(⋆) (†) Dept. of Electronic Engineering (‡) Dept. of Civil Engineering and Computer Science Engineering (⋆) Dept. of Enterprise Engineering University of Rome, Tor Vergata Rome, Italy {castellucci,filice,croce,basili}@info.uniroma2.it Abstract In this paper, the UNITOR system participat- ing in the SemEval-2013 Sentiment Analysis in Twitter task is presented. The polarity de- tection of a tweet is modeled as a classifica- tion task, tackled through a Multiple Kernel approach. It allows to combine the contribu- tion of complex kernel functions, such as the Latent Semantic Kernel and Smoothed Par- tial Tree Kernel, to implicitly integrate syn- tactic and lexical information of annotated ex- amples. In the challenge, UNITOR system achieves good results, even considering that no manual feature engineering is performed and no manually coded resources are em- ployed. These kernels in-fact embed distri- butional models of lexical semantics to deter- mine expressive generalization of tweets. 1 Introduction Web 2.0 and Social Networks technologies allow users to generate contents on blogs, forums and new forms of communication (such as micro-blogging) writing their opinion about facts, things, events. The analysis of this information is crucial for companies, politicians or other users in order to learn what peo- ple think, and consequently to adjust their strategies. In such a scenario, the interest in the analysis of the sentiment expressed by people is rapidly growing. Twitter1 represents an intriguing source of informa- tion as it is used to share opinions and sentiments about brands, products, or situations (Jansen et al., 2009). 1http://www.twitter.com On the other hand, tweet analysis represents a challenging task for natural language processing systems. Let us consider the following tweets, evok- ing a positive (1), and negative (2) polarity, respec- tively. Porto amazing as the sun sets... http://bit.ly/c28w (1) @knickfan82 Nooooo ;( they delayed the knicks game until Monday! (2) Tweets are short, informal and characterized by their own particular language with “Twitter syntax”, e.g. retweets (“RT”), user references (“@”), hash- tags (“#”) or other typical web abbreviations, such as emoticons or acronyms. Classical approaches to sentiment analysis (Pang et al., 2002; Pang and Lee, 2008) are not directly ap- plicable to tweets: most of them focus on relatively large texts, e.g. movie or product reviews, and per- formance drops are experimented in tweets scenario. Some recent works tried to model the sentiment in tweets (Go et al., 2009; Pak and Paroubek, 2010; Kouloumpis et al., 2011; Davidov et al., 2010; Bifet and Frank, 2010; Croce and Basili, 2012; Barbosa and Feng, 2010; Agarwal et al., 2011). Specific ap- proaches and feature modeling are used to achieve good accuracy levels in tweet polarity recognition. For example, the use of n-grams, POS tags, polar- ity lexicon and tweet specific features (e.g. hash- tag, retweet) are …

Content truncated to 500 words