Structure-Aware Review Mining and Summarization

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 653–661, Beijing, August 2010 Structure-Aware Review Mining and Summarization Fangtao Li1, Chao Han1, Minlie Huang1, Xiaoyan Zhu1, Ying-Ju Xia2, Shu Zhang2 and Hao Yu2 1State Key Laboratory of Intelligent Technology and Systems� 1Tsinghua National Laboratory for Information Science and Technology� 1Department of Computer Science and Technology, Tsinghua University 2Fujitsu Research and Development Center fangtao06@gmail.com; zxy_dcs@tsinghua.edu.cn Abstract In this paper, we focus on object feature 1 1 Introduction based review summarization. Different from most of previous work with linguistic rules or statistical methods, we formulate the review mining task as a joint structure tagging prob- lem. We propose a new machine learning framework based on Conditional Random Fields (CRFs). It can employ rich features to jointly extract positive opinions, negative opi- nions and object features for review sentences. The linguistic structure can be naturally inte- grated into model representation. Besides li- near-chain structure, we also investigate con- junction structure and syntactic tree structure in this framework. Through extensive experi- ments on movie review and product review data sets, we show that structure-aware mod- els outperform many state-of-the-art ap- proaches to review mining. With the rapid expansion of e-commerce, people are more likely to express their opinions and hands-on experiences on products or services they have purchased. These reviews are impor- tant for both business organizations and personal costumers. Companies can decide on their strat- egies for marketing and products improvement. Customers can make a better decision when pur- 1 Note that there are two meanings for word “feature”. We use “object feature” to represent the target entity, which the opinion expressed on, and use “feature” as the input for machine learning methods. chasing products or services. Unfortunately, reading through all customer reviews is difficult, especially for popular items, the number of re- views can be up to hundreds or even thousands. Therefore, it is necessary to provide coherent and concise summaries for these reviews. Figure 1. Feature based Review Summarization Inspired by previous work (Hu and Liu, 2004; Jin and Ho, 2009), we aim to provide object fea- ture based review summarization. Figure 1 shows a summary example for movie “Gone with the wind”. The object (movie) features, such as “movie”, “actor”, with their correspond- ing positive opinions and negative opinions, are listed in a structured way. The opinions are ranked by their frequencies. This provides a con- cise view for reviews. To accomplish this goal, we need to do three tasks: 1), extract all the ob- ject features and opinions; 2), determine the sen- timent polarities for opinions; 3), for each object feature, determine the relevant opinions, i.e. ob- ject feature-opinion pairs. For the first two tasks, most previous studies employ linguistic rules or statistical methods (Hu and Liu, 2004; Popescu and Etzioni 2005). They mainly use unsupervised learning methods, which lack an effective way to address infre- quent object features and opinions. They are also hard to incorporate rich overlapping features. Gone With The Wind: Movie: Positive: …

Content truncated to 500 words