With the rapid development of Internet commerce, more and morepeople prefer to buy the products they want through e-commerce sites. Mostconsumers refer to other consumer reviews to learn about the products orservices. At the same time, manufacturers and sellers can get user feedbackfrom the comments to improve and their products. Therefore the productreviews contains valuable information for both consumers and sellers.Reviews mining becomes a research hotspot.Our research is on aspect-based reviews mining. Aspect-based reviewsmining is an extension of reviews mining. The major subtasks ofaspect-based reviews are aspects extraction and grouping, aspects sentimentanalysis. In this thesis, we discuss and explore the solutions to the keyproblems related to this two subtasks.First, we propose a semi-automatic clustering method for productaspects with help external knowledge base. People tend to use differentwords to express the same aspects of a product in reviews, so it is necessaryto cluster or group these words into target aspect. Manual methods aretime-consuming and hard to be adapted on other training data, automaticmethods produce much noise, we propose a semi-automatic method toovercome these disadvantages.After that, we propose a sentiment strength analysis method based onfuzzy sets. Dictionary based approaches are applicable to opinion word ingeneral domain, but has problems to cope with the opinion words in specificdomains. Our approach is domain-independent and able to quantify thesentiment strength expressed. Finally, we apply the sentiment analysis methods on aspect rating prediction and compare the pros and cons ofdifferent methods. |