A Study On The Method Of Generating Chinese Opinion Abstracts For Product Reviews

As one of the research focus in natural language processing and opinion mining, the goal of opinion summarization is to aggregate a variety of opinion information scattered in difference opinionated texts or reviews, and further to generate a streamlined and readable text summary. Opinion summarization not only play a very important role in opinion mining systems like opinion question answering and opinion information retrieval, but also have prospects for applications in business intelligence and recommendation systems.This thesis aims to explore some key issues in opinion summarization for Chinese product reviews, mainly including explanatory opinion extraction, opinion information clustering and opinion summarization generation. To this end, this thesis concerns the following three main aspects.(1) Multi-granularity Chinese explanatory opinion extraction. Explanatory opinion mining is an emerging field in opinion mining which focuses on mining the underlying reasons for opinions. To acquire explanatory opinion information for subsequent opinion summarization, this thesis proposes to investigate explanatory opinion extraction at sentence and phrase level. In particular, we incorporate word embeddings into the supported vector machines to perform explanatory sentence classification, and further exploit a weakly-supervised semantic pattern matching method to recognize explanatory segments within explanatory opinionated sentences. Experimental results over product reviews in mobilephone and car domains show that the proposed approach significantly outperforms existing state-of-the-art methods for explanatory opinion extraction.(2) Opinion aggregation and its application in polarity classification. Opinion aggregation is to group together the reviews for a certain product attributes. To approach this, we first consider features at three linguistic levels of similarities, namely literal level, semantic level and context level, and thus present a two-stage hierarchical clustering algorithm for Chinese opinion aggregation. Furthermore, to avoid data sparseness in sentiment classification of short opinionated texts, we further introduce opinion clusters into sentiment polarity classification and propose a cluster-based approach to Chinese sentiment classification under the framework of support vector machines. Experimental results show the incorporation of multiple levels of similarities and the introduction of product attribute clusters is of great value to opinion aggregation and polarity classification, respectively.(3) Chinese opinion text summarization generation. On the basis of explanatory opinion extraction and opinion aggregation, we finally investigate the problem of text summarization generation for Chinese product reviews. In particular, we first utilize a graph ranking algorithm to sort opinion sentences in attribute clusters, and then remove redundant opinions using the maximal marginal relevance algorithm. Finally, we define a number of summary templates and further exploit them to generate text summaries. Experimental results show that our method can generate a readable and general summary.