Text Sentiment Polarity Analysis Based On Chinese Reviews In Hotel Domain

With the development of Internet,especially in the time of Web2.0,there are a lot of information generated on the Internet.Most of these information are created by Internet users in the form of texts.Users would comment on products when shopping online.Users would comment on news when reading them.Users would also express their emotions on the social networking platforms.In order to mine useful information from these texts,many different technologies and research directions in Natural Language Processing are created.Text sentiment analysis is one of the hottest topic.There are many application scenarios for text sentiment analysis,for example,Product Recommendation System,Online Public Opinion Analysis System and Decision Making System.China,as the country has the largest internet population,has more and more needs for Chinese text sentiment analysis.This thesis studies the application of Chinese text sentiment analysis in hotel domain and main tasks are included:(1)Create a feature lexicon and sentiment lexicon in the hotel domain.We use some public Chinese sentiment lexicons as the basic lexicon and extend the basic lexicon by the method of measuring semantic similarity of candidate sentiment words.(2)Extract the units of sentiment elements from a hotel review in a fine granularity and compute these units' emotional value using the dependency parsing,feature lexicon and sentiment lexicon.According to emotional value of the units composed by feature word,sentiment word and modifiers,we can gain the emotional evaluation on each feature of the hotel.(3)Classify the sentiment polarity of a whole hotel review based on the methods of machine learning and sentiment lexicon.Firstly,compute each review's value of emotional tendency and pick up those reviews who has high emotional value in each class as train-set.Secondly,combine multiply SVM-KNN classifiers by the method of multi-features fusion to classify the candidate reviews.The experiment shows that optimizing the selection of train-set can improve the classifier's performance and the method of combining multi-classifiers by multi-features fusion has an advantage compared with single classifier.(4)Finally,design and implement a hotel retrieval system based on text sentiment analysis technologies in this thesis.So far,most of hotel reservation websites only provide objective retrieval conditions,for example the position or the price of hotel and so on,to retrieve hotels information.The system on this thesis not only provide objective retrieval conditions,but also provide subjective retrieval conditions according to the text sentiment analysis.For example,whether the hotel has high cost performance,whether the environment is comfortable and so on.The system can retrieve hotels' information and show the result to user according to both of objective retrieval conditions and subjective retrieval conditions that user selects.