Opinion retrieval in social networks is a very useful field for industry because it can provide a facility for monitoring opinions about a product, person or issue in real time. An opinion retrieval system generally retrieves topically relevant and subjective documents based on topical relevance and a degree of subjectivity. Previous studies on opinion retrieval only considered the intrinsic features of original tweet documents and thus suffer from the data sparseness problem. In this paper, we propose a method of utilizing the extrinsic information of the original tweet and solving the data sparseness problem. We have found useful extrinsic features of related tweets, which can properly measure the degree of subjectivity of the original tweet. When we performed an opinion retrieval experiment including proposed extrinsic features within a learning-to-rank framework, the proposed model significantly outperformed both the baseline system and the state-of-the-art opinion retrieval system in terms of Mean Average Precision (MAP) and Precision@K(P@K) metrics.
