NileTMRG at SemEval-2016 Task 7: Deriving Prior Polarities for Arabic Sentiment Terms

Proceedings of SemEval-2016, pages 486–490, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics NileTMRG at SemEval-2016 Task 7: Deriving Prior Polarities for Arabic Sentiment Terms Samhaa R. El-Beltagy Center for Informatics Sciences Nile University, Juhayna Square, Sheikh Zayed City Giza, Egypt samhaa@computer.org Abstract This paper presents a model that was devel- oped to address SemEval Task 7: “Determin- ing Sentiment Intensity of English and Arabic Phrases”, with focus on ‘Arabic Phrases’. The goal of this task is to determine the degree to which some given term is associated with pos- itive sentiment. The underlying premise be- hind the model that we have adopted is that determining the context (positive or negative) in which a term usually occurs can determine its strength. Since the focus is on Twitter terms, Twitter was used to collect tweets for each term for which a strength value was to be derived. An Arabic sentiment analyzer, was then used to assign a polarity to each of these tweets, thus defining their context. We then experimented with normalized point wise mu- tual information with and without linear re- gression to assign intensity scores to input terms. The output of the model that we’ve adopted ranked at two out of the three pre- sented systems for this task with a Kendall score of 0.475. 1 Introduction During the past few years, interest in sentiment analysis has surged. Sentiment lexicons are often an essential component for building sentiment analysis systems. Entries in sentiment lexicons can vary significantly in terms of how strongly they re- flect positive or negative sentiment. For example, while both the terms “good” and “amazing” reflect a positive sentiment, the term “amazing” is strong- er and more positive than “good”. In this paper, we address early work that has been conducted to de- termine the intensity of Arabic twitter sentiment terms with respect to positive sentiment. Since the desired task is to determine the “positive strength” of a word, the results can be interpreted as follows: the closer the term score is to 1, the stronger it is as a positive indicator; the closer it is to 0, the strong- er it is as a negative indicator. Terms that can be used in both positive and negative contexts will usually fall somewhere in the middle. While this task was introduced in SemEval--2015 as Task 10- sub-task E (Rosenthal et al. 2015) for English terms, this is the first time that it has been introduced for Arabic terms. Addressing this task for Arabic has to take place in the absence of many resources that are available for English. In this work we make of use an Arabic sentiment analyzer (El-Beltagy et al. 2016) that was developed by our team as well as of a sentiment lexicon that was also developed within by same team and which is pub- licly available for research purposes (El-Beltagy 2016). The main idea behind this work is to try to collect a representative number of tweets for each term …

Content truncated to 500 words