Detecting and Blocking False Sentiment Propagation

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 354–362, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Detecting and Blocking False Sentiment Propagation Hye-Jin Min CS Department, KAIST Daejeon, Republic of Korea hjmin@nlp.kaist.ac.kr Jong C. Park CS Department, KAIST Daejeon, Republic of Korea park@cs.kaist.ac.kr Abstract Sentiment detection of a given expression in- volves interaction with its component constit- uents through rules such as polarity propaga- tion, reversal or neutralization. Such composi- tionality-based sentiment detection usually performs better than a vote-based bag-of- words approach. However, in some contexts, the polarity of the adjectival modifier may not always be correctly determined by such rules, especially when the adjectival modifier char- acterizes the noun so that its denotation be- comes a particular concept or an object in cus- tomer reviews. In this paper, we examine ad- jectival modifiers in customer review sentenc- es whose polarity should either be propagated (SHIFT) or not (UNSHIFT). We refine polari- ty propagation rules in the literature by con- sidering both syntactic and semantic clues of the modified nouns and the verbs that take such nouns as arguments. The resulting rules are shown to work particularly well in detect- ing cases of ‘UNSHIFT’ above, improving the performance of overall sentiment detection at the clause level, especially in ‘neutral’ sen- tences. We also show that even such polarity that is not propagated is still necessary for identifying implicit sentiment of the adjacent clauses. 1 Introduction Detecting the sentiment of a given grammatical unit such as phrase, clause or sentence has re- ceived much attention in opinion mining and sentiment analysis. While earlier work simply detected the overall polarity by computing the majority of the polarity of words within the ex- pression, researchers are now looking more into the composition of polarity of words within the expression (Wilson et al., 2005; Moilanen and Pulman, 2007; Choi and Cardie, 2008). They have utilized either word features (e.g., ‘Context Valence Shifters’) or grammatical structures (e.g., ‘the Principle of Compositionality’). It is shown that a machine learning approach with these fea- tures performs better than a vote-based bag-of- words approach. While the importance of salient features such as ‘negation’ or ‘intensifier’ is fully recognized, it is not yet clearly understood when the polarity of a particular word is propagated or is sacrificed. The polarity of an adjectival modifier is often propagated to the polarity of the modified noun or noun phrases with no inherent polarity. How- ever, sometimes the polarity is not propagated to that of the enclosing clause or sentence at all. For example, the polarity of the word ‘real’ is not propagated to that of the sentence “Real 501’s are made of 14 oz canvas-like material.” in a cus- tomer review of the pants, even though there is no salient sentiment word except for the word ‘real’. Our observation shows that stopped prop- agation of this kind in customer reviews often appears because of the following reasons: 1) the word in question is mainly used …

Content truncated to 500 words