Using Predicate-Argument Structures for Context-Dependent Opinion Retrieval

Current opinion retrieval techniques do not provide context-dependent relevant results. They use frequency of opinion words in documents or at proximity to query words, such that opinionated documents containing the words are retrieved regardless of their contextual or semantic relevance to the query topic. Thus, opinion retrieved for the qualitative analysis of products, performance measurement for companies, and public reactions to political decisions can be largely biased. We propose a sentence-level linear relevance model that is based on subjective and semantic similarities between predicate-argument structures. This ensures opinionated documents are not only subjective but semantically relevant to the query topic. The linear relevance model performs a linear combination of a popular relevance model, our proposed transformed terms similarity model, and a popular subjectivity mechanism. Evaluation and experimental results show that the use of predicate-argument structures improves performance of opinion retrieval task by more than 15% over popular TREC baselines.