Generalizing Syntactic Structures for Product Attribute Candidate Extraction

Human Language Technologies: The 2010 Annual Conference of the North American Chapter of the ACL, pages 377–380, Los Angeles, California, June 2010. c⃝2010 Association for Computational Linguistics Generalizing Syntactic Structures for Product Attribute Candidate Extraction Yanyan Zhao, Bing Qin, Shen Hu, Ting Liu Harbin Institute of Technology, Harbin, China {yyzhao,bqin,shu,tliu}@ir.hit.edu.cn Abstract Noun phrases (NP) in a product review are always considered as the product attribute candidates in previous work. However, this method limits the recall of the product at- tribute extraction. We therefore propose a novel approach by generalizing syntactic structures of the product attributes with two strategies: intuitive heuristics and syntactic structure similarity. Experiments show that the proposed approach is effective. 1 Introduction Product attribute extraction is a fundamental task of sentiment analysis. It aims to extract the product at- tributes from a product review, such as “picture qual- ity” in the sentence “The picture quality of Canon is perfect.” This task is usually performed in two steps: product attribute candidate extraction and candidate classification. Almost all the previous work pays more attention to the second step, fewer researchers make in-depth research on the first step. They simply choose the NPs in a product review as the product attribute can- didates (Hu and Liu, 2004; Popescu and Etzioni, 2005; Yi et al., 2003). However, this method lim- its the recall of the product attribute extraction for two reasons. First, there exist other structures of the product attributes except NPs. Second, the syntactic parsing is not perfect, especially for the Non-English languages, such as Chinese. Experiments on three Chinese datasets1 show that nearly 15% product at- tributes are lost, when only using NPs as the can- didates. Obviously, if using the candidate classifi- cation techniques on these NP candidates, it would 1It refers to the training data in Section 3.1. lead to poor performance (especially for recall) for the final product attribute extraction. Based on the above discussion, it can be observed that product attribute candidate extraction is well worth studying. In this paper, we propose an ap- proach by generalizing the syntactic structures of the product attributes to solve this problem. Figure 1 lists some syntactic structure samples from an an- notated corpus, including the special forms of NPs in Figure 1(a) and other syntactic structures, such as VP or IP in Figure 1(b). We can find that the syntac- tic structures can not only cover more phrase types besides NP, but also describe the detailed forms of the product attributes. NP NN 屏幕 (screen) NP NN 屏幕 NP NN 分辨率 (screen resolution) NP QP CD 单 NP NN 声道 (single track) NP ADJP JJ 前 NP NN 座椅 (front seats) NP VB 拍照 NP NN 功能 (photographing function) VP NP NN 屏幕 VP VB 显示 (screen display) IP (a) syntactic structure samples of NP (b) syntactic structure samples of other phrases Figure 1: Syntactic structure samples of the product at- tributes (acquired by an automatic phrase parser). In order to exploit more and useful syntactic struc- tures, two …

Content truncated to 500 words