Mining Opinion Attributes from Texts Using Multiple Kernel Learning

In this paper we propose a novel framework for recognizing complex opinion attributes from product reviews. Instead of focusing on linguistic properties of text fragments and their direct representations, we focus on these fragments´ similarities which we obtain from multiple sources of lexical and semantic information. The problem is formulated as that of multiclass classification and is based on multiple similarity matrices. We apply multiple kernel learning algorithm which seeks optimal combinations of matrices using linear programming and support vector machines for classification. Experiments demonstrate benefits from multiple sources of information. Overall, the approach is promising especially in the case of reviews of product types with complex and wordy attribute expressions.