Sentiment Relevance

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 954–963, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Sentiment Relevance Christian Scheible Institute for Natural Language Processing University of Stuttgart, Germany scheibcn@ims.uni-stuttgart.de Hinrich Sch¨utze Center for Information and Language Processing University of Munich, Germany Abstract A number of different notions, including subjectivity, have been proposed for dis- tinguishing parts of documents that con- vey sentiment from those that do not. We propose a new concept, sentiment rele- vance, to make this distinction and argue that it better reflects the requirements of sentiment analysis systems. We demon- strate experimentally that sentiment rele- vance and subjectivity are related, but dif- ferent. Since no large amount of labeled training data for our new notion of sen- timent relevance is available, we investi- gate two semi-supervised methods for cre- ating sentiment relevance classifiers: a dis- tant supervision approach that leverages structured information about the domain of the reviews; and transfer learning on feature representations based on lexical taxonomies that enables knowledge trans- fer. We show that both methods learn sen- timent relevance classifiers that perform well. 1 Introduction It is generally recognized in sentiment analy- sis that only a subset of the content of a doc- ument contributes to the sentiment it conveys. For this reason, some authors distinguish the categories subjective and objective (Wilson and Wiebe, 2003). Subjective statements refer to the internal state of mind of a person, which cannot be observed. In contrast, objective statements can be verified by observing and checking reality. Some sentiment analysis systems filter out objective lan- guage and predict sentiment based on subjective language only because objective statements do not directly reveal sentiment. Even though the categories subjective/objective are well-established in philosophy, we argue that they are not optimal for sentiment analysis. We in- stead introduce the notion of sentiment relevance (S-relevance or SR for short). A sentence or lin- guistic expression is S-relevant if it contains infor- mation about the sentiment the document conveys; it is S-nonrelevant (SNR) otherwise. Ideally, we would like to have at our disposal a large annotated training set for our new con- cept of sentiment relevance. However, such a resource does not yet exist. For this reason, we investigate two semi-supervised approaches to S-relevance classification that do not require S- relevance-labeled data. The first approach is dis- tant supervision (DS). We create an initial label- ing based on domain-specific metadata that we ex- tract from a public database and show that this improves performance by 5.8% F1 compared to a baseline. The second approach is transfer learning (TL) (Thrun, 1996). We show that TL improves F1 by 12.6% for sentiment relevance classification when we use a feature representation based on lex- ical taxonomies that supports knowledge transfer. In our approach, we classify sentences as S- (non)relevant because this is the most fine-grained level at which S-relevance manifests itself; at the word or phrase level, S-relevance classification is not possible because of scope and context …

Content truncated to 500 words