Adjective Intensity and Sentiment Analysis

Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing, pages 2520–2526, Lisbon, Portugal, 17-21 September 2015. c⃝2015 Association for Computational Linguistics. Adjective Intensity and Sentiment Analysis Raksha Sharma, Mohit Gupta, Astha Agarwal, Pushpak Bhattacharyya Dept. of Computer Science and Engineering IIT Bombay, Mumbai, India {raksha,mohitgupta,astha11,pb}@cse.iitb.ac.in Abstract For fine-grained sentiment analysis, we need to go beyond zero-one polarity and find a way to compare adjectives that share a common semantic property. In this paper, we present a semi-supervised ap- proach to assign intensity levels to adjec- tives, viz. high, medium and low, where adjectives are compared when they belong to the same semantic category. For exam- ple, in the semantic category of EXPER- TISE, expert, experienced and familiar are respectively of level high, medium and low. We obtain an overall accuracy of 77% for intensity assignment. We show the sig- nificance of considering intensity informa- tion of adjectives in predicting star-rating of reviews. Our intensity based prediction system results in an accuracy of 59% for a 5-star rated movie review corpus. 1 Introduction Sentence intensity becomes crucial when we need to compare sentences having the same polarity ori- entation. In such scenarios, we can use inten- sity of words to judge the intensity of a sentence. Words that bear the same semantic property can be used interchangeably to upgrade or downgrade the intensity of the expression. For example, good and outstanding both are positive words from the QUALITY category, but the latter can be used to intensify positive expression in a sentence. There are several manually or automatically created lexical resources (Liu, 2010; Wilson et al., 2005b; Wilson et al., 2005a; Taboada and Grieve, 2004) that assign a fixed positive (+1) or nega- tive (−1) polarity to words, making no distinction among them in terms of their intensity. This pa- per presents a semi-supervised approach to assign intensity levels to adjectives, viz. high, medium and low, which share the same semantic property. We have used the semantic frames of FrameNet- 1.5 (Baker et al., 1998) to obtain these semantic categories. Our approach is based on the idea that the most intense word has higher contextual simi- larity with high intensity words than with medium or low intensity words. We use the intensity an- notated movie review corpus to obtain the most intense word for a semantic category. Then, co- sine similarity between word vectors of the most intense word and other words of the category is used to assign intensity levels to those words. Our approach with the used resources is shown in fig- ure 1. Figure 1: Intensity Analysis System Our Contribution: Corpus based approaches suf- fer from the data sparsity problem. Our approach tackles this problem by using word vectors for in- tensity assignment (Section 2.3). It also provides a better overall accuracy (77%) than current state of the art when compared with gold-standard in- tensity levels (Section 6.2). In addition to this, we show that accuracy of the star rating prediction task improves when we …

Content truncated to 500 words