Locally contextualized smoothing of language models for sentiment sentence retrieval

Recently, a number of documents are published on the web. One of the crucial techniques to access to such information is sentiment sentence retrieval. It is very useful to retrieve positive or negative opinions to a specific topic at sentence level. Considering the property that sentiment polarities are often locally consistent in a document, we focus on using local context information for retrieving sentiment-bearing sentences. For this objective, we propose a new smoothing method, extending Dirichlet prior smoothing, to improve effectiveness of the retrieval. We demonstrate through experiments that our proposed smoothing method achieves statistically significant improvements in sentiment sentence retrieval, compared with a conventional smoothing method without local context.