Abstract: Mining opinion targets from massive amount of product reviews is a key problem for opinion mining. Nevertheless nowaday the results of opinion target extraction only offer a little useful information. This paper presents a bidirectional bootstrapping approach which is based on context-sensitive to extract product names and features. The approach uses initial seeds and POS templates to get candidate opinion targets, then uses context-sensitive approach to extract opinion targets from the sentences which contain candidate opinion targets, and perform boundary detection. At the same time, it extracts and evaluates contexts templates of opinion targets, and puts the good templates into a template set. It iterates the approach until no new opinion targets emerge. The experimental results show that context-sensitive approach improves the performance by 10% compared with context-free approaches in extracting opinion targets.

Key words: opinion target extraction, product review, bootstrapping, context-sensitive

摘要： 从大量的产品评论中进行观点评价对象的自动抽取是观点挖掘研究的重要课题，然而目前观点评价对象抽取结果只提供少量信息，因此提出一种基于上下文相关的双向自举方法同时获取产品名称和产品属性。该方法利用初始种子集、词性模板集获得候选观点评价对象，采用上下文相关的方法对文中所有包含候选观点评价对象的语句抽取出观点评价对象并进行边界识别，同时抽取观点评价对象的词性模板并计算分数，将分值高的模板加入模板集，这样重复迭代直到没有出现新的观点评价对象为止。实验结果表明采用上下文相关方法进行观点评价对象抽取相对于上下文无关的方法性能提高10%以上。

关键词: 观点评价对象抽取, 产品评论, 自举, 上下文相关