The Research On Latent Semantic Markov Network Retrieval Model

With the development of network technology, we can obtain information conveniently. However, there still exists a challenge for us to get information efficiently. Information retrieval can help us get specific information we need.Traditional retrieval models make a simple match between a query and words in the documents, their effects are not so good influenced by uncertain factors like the relevance between the words. Graphical models can effectively interpret and infer probabilistic information. They lay in the useful information by learning and inferring the possibility of knowledge to improve the retrieval performance. Bayesian network model is one of the effective graphical models. But its anisotropy makes learning and reference become a complicated process, and lacks effective semantic interpretation.We propose the latent semantic Markov network retrieval model in this paper. Markov network is a powerful tool for representation and inference of possibility knowledge. Its isotropy is useful for network structuring and interprets the correlation between different knowledge better.Our model learns from document set, and extracts latent semantics between index items and between documents through singular value decomposition to construct a Markov knowledge network. Our experiments show that the latent semantic MNM is highly applicable, and get optimal results in the retrieval of English documents. Comparing with the BM25, our model makes some improvements in retrieval performances.New points in this thesis:1) We propose the latent semantic MNM. This model has a highly descriptive characteristic, existing classic models can all be treated as specialties of it.2) Through experiments and analysis, we confirm the retrieval performances of the latent semantic MNM, and compare it with other retrieval models. The experiments show that latent semantic in this model can portray the authentic semantic relevance in the document sets well.