Fine-Grained Sentiment Analysis with Structural Features

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 336–344, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Fine-Grained Sentiment Analysis with Structural Features C¨acilia Zirn† Mathias Niepert† †KR & KM Research Group University of Mannheim Mannheim, Germany caecilia@informatik.uni-mannheim.de Heiner Stuckenschmidt† Michael Strube‡ ‡Heidelberg Institute for Theoretical Studies Heidelberg, Germany michael.strube@h-its.org Abstract Sentiment analysis is the problem of de- termining the polarity of a text with re- spect to a particular topic. For most ap- plications, however, it is not only nec- essary to derive the polarity of a text as a whole but also to extract negative and positive utterances on a more fine- grained level. Sentiment analysis systems working on the (sub-)sentence level, how- ever, are difficult to develop since shorter textual segments rarely carry enough in- formation to determine their polarity out of context. In this paper, therefore, we present a fully automatic framework for fine-grained sentiment analysis on the sub- sentence level combining multiple senti- ment lexicons and neighborhood as well as discourse relations to overcome this prob- lem. We use Markov logic to integrate po- larity scores from different sentiment lex- icons with information about relations be- tween neighboring segments, and evaluate the approach on product reviews. The ex- periments show that the use of structural features improves the accuracy of polar- ity predictions achieving accuracy scores of up to 69%. 1 Introduction Sentiment analysis systems have continuously im- proved the quality of polarity classifications of en- tire product reviews. For numerous real-world ap- plications, however, classification on such a coarse level is not suitable. Even in their most enthu- siastic reviews, users still tend to mention nega- tive aspects of a particular product. Conversely, in very negative reviews there might still be men- tions of several positive aspects of the product. Moreover, different opinions can even be uttered in the same sentence. Consider, for instance, the sentence “Despite the pretty design I would never recommend it, because the sound quality is unac- ceptable” which expresses both positive and nega- tive opinions about a product. Thus, to determine both negative and positive utterances in product reviews, classification on the subsentence level is needed. Sentiment Analysis on Subsentence Level. As basic classification unit for our fine-grained sen- timent analysis system we choose discourse seg- ments. There are various theories describing dis- course, discourse segmentation and discourse re- lations. The most well-known theory aiming to describe some aspects of text coherence is the Rhetorical Structure Theory (RST) introduced by Mann and Thompson (1988). According to this theory, every text consists of elementary segments that are connected by relations. Segments joined by a relation form a unit, which is itself connected to other segments. This leads to a hierarchical tree structure that spans over the whole text. The ex- ample sentence given above could be divided into the three segments s1 = Despite the pretty design, s2 = I would never recommend it and s3 = be- cause the sound quality is unacceptable, …

Content truncated to 500 words