Sentiment Retrieval using Generative Models

Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing (EMNLP 2006), pages 345–354, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Sentiment Retrieval using Generative Models Koji Eguchi National Institute of Informatics Tokyo 101-8430, Japan eguchi@nii.ac.jp Victor Lavrenko Department of Computer Science University of Massachusetts Amherst, MA 01003, USA lavrenko@cs.umass.edu Abstract Ranking documents or sentences accord- ing to both topic and sentiment relevance should serve a critical function in helping users when topics and sentiment polari- ties of the targeted text are not explicitly given, as is often the case on the web. In this paper, we propose several sentiment information retrieval models in the frame- work of probabilistic language models, as- suming that a user both inputs query terms expressing a certain topic and also speci- fies a sentiment polarity of interest in some manner. We combine sentiment relevance models and topic relevance models with model parameters estimated from training data, considering the topic dependence of the sentiment. Our experiments prove that our models are effective. 1 Introduction The recent rapid expansion of access to informa- tion has significantly increased the demands on re- trieval or classification of sentiment information from a large amount of textual data. The field of sentiment classification has recently received con- siderable attention, where the polarities of senti- ment, such as positive or negative, were identified from unstructured text (Shanahan et al., 2005). A number of studies have investigated sentiment classification at document level, e.g., (Pang et al., 2002; Dave et al., 2003), and at sentence level, e.g., (Hu and Liu, 2004; Kim and Hovy, 2004; Nigam and Hurst, 2005); however, the accuracy is still less than desirable. Therefore, ranking ac- cording to the likelihood of containing sentiment information is expected to serve a crucial func- tion in helping users. We believe that our work is the first attempt at sentiment retrieval that aims at finding sentences containing information with a specific sentiment polarity on a certain topic. Intuitively, the expression of sentiment in text is dependent on the topic. For example, a nega- tive view for some voting event may be expressed using ‘flaw’, while a negative view for some politi- cian may be expressed using ‘reckless’. Moreover, sentiment polarities are also dependent on topics or domains. For example, the adjective ‘unpre- dictable’ may have a negative orientation in an au- tomotive review, in a phrase such as ‘unpredictable steering’, but it could have a positive orientation in a movie review, in a phrase such as ‘unpredictable plot’, as mentioned in (Turney, 2002) in the con- text of his sentiment word detection. We propose sentiment retrieval models in the framework of generative language modeling, not only assuming query terms expressing a certain topic, but also assuming that the polarity of sen- timent interest is specified by the user in some manner, where the topic dependence of the sen- timent is considered. To the best of our knowl- edge, there have been no other studies on a re- trieval model unifying …

Content truncated to 500 words