Grouping Product Features Using Semi-Supervised Learning with Soft-Constraints

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 1272–1280, Beijing, August 2010 Grouping Product Features Using Semi-Supervised Learning with Soft-Constraints* Zhongwu Zhai†, Bing Liu‡, Hua Xu† and Peifa Jia† †State Key Lab of Intelligent Tech. & Sys. Tsinghua National Lab for Info. Sci. and Tech. Dept. of Comp. Sci. & Tech., Tsinghua Univ. zhaizhongwu@gmail.com ‡Dept. of Comp. Sci. University of Illinois at Chicago liub@cs.uic.edu Abstract In opinion mining of product reviews, one of- ten wants to produce a summary of opinions based on product features/attributes. Howev- er, for the same feature, people can express it with different words and phrases. To produce a meaningful summary, these words and phrases, which are domain synonyms, need to be grouped under the same feature group. This paper proposes a constrained semi- supervised learning method to solve the prob- lem. Experimental results using reviews from five different domains show that the proposed method is competent for the task. It outper- forms the original EM and the state-of-the-art existing methods by a large margin. 1 Introduction* One form of opinion mining in product reviews is to produce a feature-based summary (Hu and Liu, 2004a; Liu, 2010). In this model, product features are first identified, and positive and neg- ative opinions on them are aggregated to produce a summary on the features. Features of a product are attributes, components and other aspects of the product, e.g., “picture quality”, “battery life” and “zoom” of a digital camera. In reviews (or any writings), people often use different words and phrases to describe the same product feature. For example, “picture” and “photo” refer to the same feature for cameras. Grouping such synonyms is critical for effective opinion summary. Although WorldNet and other *Supported by National Natural Science Foundation of Chi- na (Grant No: 60875073). This work was done when the first author was visiting Bing Liu’s group at the University of Illinois at Chicago. thesaurus dictionaries can help to some extent, they are far from sufficient due to a few reasons. First, many words and phrases that are not syn- onyms in a dictionary may refer to the same fea- ture in an application domain. For example, “ap- pearance” and “design” are not synonymous, but they can indicate the same feature, design. Second, many synonyms are domain dependent. For example, “movie” and “picture” are syn- onyms in movie reviews, but they are not syn- onyms in camera reviews as “picture” is more likely to be synonymous to “photo” while “mov- ie” to “video”. Third, determining which expres- sions indicate the same feature can be dependent on the user’s application need. For example, in car reviews, internal design and external design can be regarded as two separate features, but can also be regarded as one feature, called “design”, based to the level of details that the user needs to study. In camera reviews, one may want to study battery as a whole (one feature), or as more than one feature, e.g., battery weight, and battery life. …

Content truncated to 500 words