TJP: Using Twitter to Analyze the Polarity of Contexts

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 375–379, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics TJP: Using Twitter to Analyze the Polarity of Contexts Tawunrat Chalothorn Jeremy Ellman University of Northumbria at Newcastle University of Northumbria at Newcastle Pandon Building, Camden Street Pandon Building, Camden Street Newcastle Upon Tyne, NE2 1XE, UK Newcastle Upon Tyne, NE2 1XE, UK Tawunrat.chalothorn@unn.ac.uk Jeremy.ellman@unn.ac.uk Abstract This paper presents our system, TJP, which participated in SemEval 2013 Task 2 part A: Contextual Polarity Disambiguation. The goal of this task is to predict whether marked con- texts are positive, neutral or negative. Howev- er, only the scores of positive and negative class will be used to calculate the evaluation result using F-score. We chose to work as ‘constrained’, which used only the provided training and development data without addi- tional sentiment annotated resources. Our ap- proach considered unigram, bigram and trigram using Naïve Bayes training model with the objective of establishing a simple- approach baseline. Our system achieved F- score 81.23% and F-score 78.16% in the re- sults for SMS messages and Tweets respec- tively. 1 Introduction Natural language processing (NLP) is a research area comprising various tasks; one of which is sen- timent analysis. The main goal of sentiment analy- sis is to identify the polarity of natural language text (Shaikh et al., 2007). Sentiment analysis can be referred to as opinion mining, as study peoples’ opinions, appraisals and emotions towards entities and events and their attributes (Pang and Lee, 2008). Sentiment analysis has become a popular research area in NLP with the purpose of identify- ing opinions or attitudes in terms of polarity. This paper presents TJP, a system submitted to SemEval 2013 for Task 2 part A: Contextual Polar- ity Disambiguation (Wilson et al., 2013). TJP was focused on the ‘constrained’ task, which used only training and development data provided. This avoided both resource implications and potential advantages implied by the use of additional data containing sentiment annotations. The objective was to explore the relative success of a simple ap- proach that could be implemented easily with open-source software. The TJP system was implemented using the Py- thon Natural Language Toolkit (NLTK, Bird et al., 2009). We considered several basic approaches. These used a preprocessing phase to expand con- tractions, eliminate stopwords, and identify emoti- cons. The next phase used supervised machine learning and n-gram features. Although we had two approaches that both used n-gram features, we were limited to submitting just one result. Conse- quently, we chose to submit a unigram based ap- proach followed by naive Bayes since this performed better on the data. The remainder of this paper is structured as fol- lows: section 2 provides some discussion on the related work. The methodology of corpus collec- tion and data classification are provided in section 3. Section 4 outlines details of the experiment and results, followed by the conclusion and ideas for future work …

Content truncated to 500 words