Opinion Retrieval Based On Genetic Programming

As a novel information retrieval task, opinion retrieval has attracted considerable amount of attentions in recent years. Current researches mainly adopted the classical two-stage framework, i.e., first retrieving topic relevant documents and then re-ranking them according to the appropriate combination of their topic and opinion relevant scores using some combination function. Two critical issues in such a framework are how to calculate topic and opinion relevant scores and how to optimize the combination function to integrate the two scores effectively to make the top returned results both topic and opinion relevant. One major problem in existing works is that the score combination functions are defined in advance regardless of domains. However, there is no evidence that these two scores should be combined in a unique way.In this paper, we propose to tackle the opinion relevant score calculation problem based on the similarity measure of the usage of opinion words. Also, we propose to learn the combination functions automatically for retrieval tasks of different domains. We employ the popular Genetic Programming framework for the learning task. To perform the whole opinion retrieval task, we also design a novel opinion retrieval system to compute the topic and opinion relevant scores and then learn the optimal combination function to integrate these two scores.During the experiments, we compare our system with other state-of-the-art work on a public dataset and the experimental results show that our system performs comparatively well with others.