Latent semantic model has been applied to cross-language information retrieval, image annotation, image retrieval, sentiment analysis and other fields. In this paper, we propose the Multitype-LDA model for cross-media information retrieval and the CTS-LDA model for sentiment analysis. The main contributions of this thesis include following issues: We propose the Multitype-LDA to deal with corpus which have multi-type “words”.The words can be words from different languages, the features from different medium or the mixtures of the words and features. The Multitype-LDA model can find the semantic relation between the different types of "words", and can be applied to cross-language or cross-media information retrieval. We apply the Multitype-LDA model to the image notation. We also find that combining the original document model with the Multitype-LDA model is effective in image image retrieval. We also construct a Multitype-LDA based cross language information retrieval model.The model can bridge the semantic gap of different languages. The experiments indicate that the Multitype-LDA based cross language information retrieval model is very effective. We propose a topic sentiment mixture model which we call TS-LDA model for the sentiment analysis.The TS-LDA model can find the topic and the sentiment of the online reviews simultaneously. To enable the user benefit from the expert opinions and the ordinary opinions, we propose the CTS-LDA model, which stand for concept and topic sentiment LDA model. In CTS-LDA model, the expert opinions are concepts and the ordinary opinions are topics. The CTS-LDA model can obtain the sentiment polarity of the concepts or topics in each review. Experiments show that CTS-LDA model is very suitable for sentiment analysis. We also develop a demo system for online sentiment analysis of the product reviews. We call the system as CTS. In CTS, we apply the CTS-LDA model for sentiment analysis, and use the semantic web to build knowledge base to store and query the results, and illustrate the results by phical interface. In a word, in this thesis, we have made a lot of fruitful attempts and significant progresses on latent semantic model.
修改评论