Fine-Grained Opinion Mining as a Relation Classification Problem

Abstract

The main focus of this paper is to investigate methods for opinion extraction at a more detailed level of granularity, retrieving not only the opinionated portion of text, but also the target of that expressed opinion. We describe a novel approach to fine-grained opinion mining that, after an initial lexicon based processing step, treats the problem of finding the opinion expressed towards an entity as a relation classification task. We detail a classification workflow that combines the initial lexicon based module with a broader classification part that involves two different models, one for relation classification and the other for sentiment polarity shift identification. We provided detailed descriptions of a series of classification experiments in which we use an original proximity based bag-of-words model. We also introduce a new use of syntactic features used together with a tree kernel for both the relation and sentiment polarity shift classification tasks.