Exploiting Syntactic and Semantic Kernels for Target-Polarity Word Collocation Extraction

Target-polarity word (T-P) collocation extraction is a basic sentiment analysis task, which aims to extract the targets and their modifying polarity words by analyzing the relationships between them. Recent studies rely primarily on syntactic rule matching. However, this kind of method has two disadvantages: (1) the syntactic rules are limited and hard matching is always used during the matching procedure that can result in the low recall value, and (2) this method omits the effect of the latent lexical semantic information on the T-P collocation extraction task. To solve the structured syntactic information loss and the lexical semantic information loss problems, in this paper we study the impact of syntactic and semantic information and incorporate them into kernels to automatically extract T-P collocations. We define a syntactic kernel based on the tree structures between targets and polarity words, and a semantic kernel based on the word semantic representations of them to exploit the good way to represent the two kinds of information with support vector machines. The experiments on four types of product reviews show that based on the basic features, a suitable combination of the syntactic and semantic kernels is very promising for T-P collocation extraction.