Research On Aspect Extraction And Disambiguation In Sentiment Analysis

Sentiment analysis is a process of getting the preference and sentiment ofobjects through analyzing the comments.There are three kinds of sentiment analysis based on the grain, which areword-level, sentence-level and document-level. The study of sentiment analysisindicates that the sentiment information is incomplete through sentiment analysisby word and sentence because the information the methods extract is limited. Andthe analysis by document concentrates on the algorithm of analyzing the polarity ofthe whole review. It is too macroscopic to getting the polarities of aspects in thereview. The key to solve two problems is aspect sentiment analysis.We propose the algorithm of aspect extraction and target completion to solvethe two problems. We can get the find-grained sentiment information while we willnot lose the information of the whole review.We discuss the algorithm of aspect extraction in the first part of the thesis.We talk about the related work on aspect extraction. And we study the supervised-learning method to extract the aspects. We compare the Maximum Entropy Modelsand Conditional Random Fields to choose the better one.The aspect lexicon is another topic of my thesis. We build an aspect lexiconby semi-supervised learning method. The aspect lexicon is built by some simplerules and a small-scale sentiment lexicon. We discuss the pruning of the aspectlexicon. We study the point in detail and deeply, which is a new point few peoplestudy. Then we extract aspects combining supervised-learning and aspect lexicon.We explore the target completion problem. This problem is novel andcomplex. We propose some methods to solve it refer to some ideas fromcoreference resolution and some characteristics which are special in the sentimentanalysis field. We study the center theory and the sentiment consistency.