基于语义分析的评价对象-情感词对抽取  

Extracting Target-Opinion Pairs Based on Semantic Analysis

作　　者：江腾蛟[1,2] 万常选[1,2] 刘德喜[1,2] 刘喜平[1,2] 廖国琼[1,2] 

机构地区：[1]江西财经大学信息管理学院,南昌330013 [2]江西财经大学数据与知识工程江西省高校重点实验室,南昌330013

出　　处：《计算机学报》2017年第3期617-633,共17页Chinese Journal of Computers

基　　金：国家自然科学基金项目(61562032;61662027;61662032;61173146;61363039;61363010;61462037);江西省自然科学基金重大项目(20152ACB20003);江西省高等学校科技落地计划项目(KJLD12022;KJLD14035)资助~~

摘　　要：评价对象-情感词对是情感词及其所修饰评价对象的组合,评价对象-情感词对的识别是细粒度情感分析的一个原子任务和关键任务.现有的研究大多集中在商品评论上,随着金融大数据的涌现,金融评论的情感挖掘意义凸显.与商品评论不同,中文金融评论中评价对象数目繁多且构成形式复杂,虚指评价对象和隐式评价对象也更常见;情感词的词性更丰富,其在句中的句法成分也更灵活、语义更丰富.针对金融评论的这些特点,该文提出了基于浅层语义与语法分析相结合的评价对象-情感词对抽取方法.考虑到金融评论多动词情感词,设计了语义角色标注与依存句法分析相结合的评价对象-情感词对抽取规则,解决了评价对象构成的复杂性问题;基于语义和领域知识对虚指评价对象进行了判别和替换,以明确其实际的指向和含义;基于特殊情感词搭配表、上下文搭配表及频繁搭配表提出了隐式评价对象识别的新思路,能有效地识别出缺省和隐含评价对象.在大规模的中文金融评论上进行了详细的实验测试,实验结果表明了该方法的有效性.A Target-opinion pair is a combination of opinion word and the target it modified. The extraction of target-opinion pairs is an atomic and key task of fine-grained sentiment analysis. Most existing work on extracting target-opinion pairs considers product reviews. As more and more financial data accumulate on the Web, sentiment mining of financial reviews becomes an important task. In this work, we put focus on Chinese financial reviews. Compared with product reviews, Chinese financial reviews have some unique characteristics. First, the number of targets is very large in Chinese financial reviews and the structure of targets is usually more complex. Second, it is very common for a Chinese review to have ambiguous and implicit targets. What is more, the opinion words in Chinese financial reviews are more flexible in POS （Part-Of-Speech） and syntactic roles, and richer in semantics. In this paper, based on shallow semantic and syntactic parsing, we propose a new method for extracting target-opinion pairs from Chinese financial reviews. Considering that many opinion words in financial reviews are verbs, we design extracting rules of target-opinion pairs based on semantic role labeling and dependency parsing, which reserves the structural complexity of targets. Our method distinguishes and replaces ambiguous targets based on semantics and domain knowledge, making the actual references and meanings of ambiguous targets explicit. To identify the implicit targets, we propose a novel approach based on the one- to-one correspondences of some special opinion words and targets, contextual co-occurrences of targets and frequent co-occurrences of targets, which is able to effectively identify the default and implied targets. We conduct comprehensive experiments on Chinese financial reviews, and experimental results show that the proposed methods are effective.

关 键 词：情感分析 中文金融评论 评价对象-情感词对 语义角色标注 依存句法分析 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]