Research On Approaches To Opinion Target Extraction In Opinion Mining

Opinion mining, also known as sentiment analysis, is to analyze people’s sentiments, attitudes or opin-ions toward things such as products, services, individuals, issues and topics. It is important for both research and application. There are coarse-grained and fine-grained opinion mining. Although coarse-grained opinion mining has been well-developed, a lot of problems still need to be addressed in fine-grained opinion mining.Opinion target extraction, also called aspect extraction, is a key task in fine-grained opinion mining, which aims to extract fine-grained opinion targets from opinion texts, such as products, and also their com-ponents, attributes and features. There are two main approaches to aspect extraction, i.e., supervised and unsupervised. The former are mainly based on Hidden Markov Model and Conditional Random Field, and the latter are mainly based on Topic Modeling and syntactical rules. Recent work has shown that the syntacti-cal rule based approach performs very well, however, this approach is still facing some challenges. The first challenge is how to implement the aspect extraction rules efficiently. The second challenge is how to select a subset of high quality aspect extraction rules from a set of rules which all vary in quality. The third challenge is how to employ a huge number of unlabeled reviews on the Internet to help extract aspects. In order to address these challenges, we provide the following solutions in this thesis. To the best of our knowledge, we are the first to propose these solutions.(1) A new aspect extraction framework is proposed based on logic programming, specially, Answer Set Programming (ASP) in order to implement aspect extraction rules efficiently. Firstly, represent the Part-of-Speech tags and dependency relations of words in the reviews as ASP facts. Secondly, translate existing aspect extraction rules into ASP rules. Finally, implement the extraction rules automatically using existing ASP solvers. Experimental results show that the proposed approach is efficient and elegant.(2) Two automatic rule selection approaches are proposed in order to automatically select a subset of high quality rules from a set of aspect extraction rules which all vary in quality. The first approach employs a greedy algorithm, and the second one employs a local search algorithm, specifically, simulated annealing. Experimental results show that the proposed approaches can effectively select a subset of high quality rules from a given rule set to achieve significantly better results than the original full rule set.(3) An aspect recommendation approach is proposed based on semantic similarities and associations in order to improve aspect extraction employing a huge number of unlabeled reviews on the Internet. Two types of knowledge are first learnt from a huge number of unlabeled reviews, including semantic similarities and associations between words. Then they are employed together with a set of seed aspects to recommend potential aspects to new domains. Experimental results show that the proposed approach can employ the knowledge learnt from reviews in other domains to recommend high quality aspects to new domains.