ReNew: A Semi-Supervised Framework for Generating Domain-Specific Lexicons and Sentiment Analysis

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 542–551, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics ReNew: A Semi-Supervised Framework for Generating Domain-Specific Lexicons and Sentiment Analysis Zhe Zhang Department of Computer Science North Carolina State University Raleigh, NC 27695-8206 zzhang13@ncsu.edu Munindar P. Singh Department of Computer Science North Carolina State University Raleigh, NC 27695-8206 singh@ncsu.edu Abstract The sentiment captured in opinionated text provides interesting and valuable informa- tion for social media services. However, due to the complexity and diversity of linguistic representations, it is challeng- ing to build a framework that accurately extracts such sentiment. We propose a semi-supervised framework for generat- ing a domain-specific sentiment lexicon and inferring sentiments at the segment level. Our framework can greatly reduce the human effort for building a domain- specific sentiment lexicon with high qual- ity. Specifically, in our evaluation, work- ing with just 20 manually labeled reviews, it generates a domain-specific sentiment lexicon that yields weighted average F- Measure gains of 3%. Our sentiment clas- sification model achieves approximately 1% greater accuracy than a state-of-the-art approach based on elementary discourse units. 1 Introduction Automatically extracting sentiments from user- generated opinionated text is important in build- ing social media services. However, the complex- ity and diversity of the linguistic representations of sentiments make this problem challenging. High-quality sentiment lexicons can improve the performance of sentiment analysis models over general-purpose lexicons (Choi and Cardie, 2009). More advanced methods such as (Kanayama and Nasukawa, 2006) adopt domain knowledge by ex- tracting sentiment words from the domain-specific corpus. However, depending on the context, the same word can have different polarities even in the same domain (Liu, 2012). In respect to sentiment classification, Pang et al. (2002) infer the sentiments using basic features, such as bag-of-words. To capture more complex linguistic phenomena, leading approaches (Naka- gawa et al., 2010; Jo and Oh, 2011; Kim et al., 2013) apply more advanced models but assume one document or sentence holds one sentiment. However, this is often not the case. Sentiments can change within one document, one sentence, or even one clause. Also, existing approaches in- fer sentiments without considering the changes of sentiments within or between clauses. However, these changes can be successfully exploited for in- ferring fine-grained sentiments. To address the above shortcomings of lexicon and granularity, we propose a semi-supervised framework named ReNew. (1) Instead of us- ing sentences, ReNew uses segments as the basic units for sentiment classification. Segments can be shorter than sentences and therefore help cap- ture fine-grained sentiments. (2) ReNew leverages the relationships between consecutive segments to infer their sentiments and automatically generates a domain-specific sentiment lexicon in a semi-su- pervised fashion. (3) To capture the contextual sentiment of words, ReNew uses dependency re- lation pairs as the basic elements in the generated sentiment lexicon. Sentiment Segment 1 2 3 4 5 NEG NEU POS transition cue transition cue Figure 1: Segments in a Tripadvisor review. Consider a part of a review from …

Content truncated to 500 words