Combining Social Cognitive Theories with Linguistic Features for Multi-genre Sentiment Analysis

Copyright 2012 by Hao Li, Yu Chen, Heng Ji, Smaranda Muresan, and Dequan Zheng 26th Pacific Asia Conference on Language,Information and Computation pages 127–136 Combining Social Cognitive Theories with Linguistic Features for Multi-genre Sentiment Analysis Hao Li1, Yu Chen2, Heng Ji1, Smaranda Muresan3, Dequan Zheng2 1Computer Science Department and Linguistics Department, Queens College and Graduate Center, City University of New York, U.S.A 2School of Computer Science and Technology, Harbin Institute of Technology, China 3School of Communication and Information, Rutgers University, U.S.A {haoli.qc,hengjicuny}@gmail.com {chenyu,dqzheng}@mtlab.hit.edu.cn smuresan@rci.rutgers.edu Abstract With the rapid development of social media and social networks, spontaneously user gen- erated content like tweets and forum posts have become important materials for tracking people’s opinions and sentiments online. In this paper we investigate the limitations of tra- ditional linguistic-based approaches to senti- ment analysis when applied to these informal genres. Inspired by various social cognitive theories, we combine local linguistic features and global social evidence in a propagation scheme to improve sentiment analysis result- s. Without using any additional labeled da- ta, this new approach obtains significant im- provement (up to 12% higher accuracy) for various genres in the domain of presidential election. 1 Introduction Sentiment analysis is an important step for both Nat- ural Language Processing (NLP) tasks such as opin- ion question answering (Yu and Hatzivassiloglou, 2003) and practical applications such as commercial product reputation mining (Morinaga et al., 2002), movie review mining (Pang et al., 2002) and politi- cal election prediction (Tumasjan et al., 2010). With the prevalence of social media, spontaneous- ly user generated content such as tweets or forum posts have become an invaluable source of people’s sentiments and opinions. However, as with other NLP tasks, sentiment analysis on such informal gen- res presents several challenges: (1) informal text ex- pressions; (2) lexical diversity (e.g., for example, in our training data only 10% of words in the discus- sion forums and tweets appear more than ten times, while in movie reviews over 20% of words appear more than ten times); (3) unpredictable shift in top- ics/issues. The prevalence of debate in both forum posts and tweets leads to the use of more compli- cated discourse structures involving multiple targets and sentiments, as well as the second-person voice. These difficulties are magnified in tweets due to nec- essarily compressed contexts (tweets are limited to 140 characters). In this paper, we tackle these challenges from t- wo perspectives. First, we approach the sentiment analysis task by identifying not only a specific “tar- get” (e.g., presidential candidate) but also its associ- ated “issues” (e.g., foreign policy) before detecting sentiment. This approach is similar to the idea of modeling “aspect” in product reviews (Titov and M- cDonald, 2008; Wang et al., 2011). Second, a detailed error analysis has shown that currently available sentiment lexicons and various shallow linguistic features are not sufficient to ad- vance simple bag-of-words baseline approaches due to the diverse ways in which sentiment can be ex- pressed as well as the prevalence of debate in social media. …

Content truncated to 500 words