Multi-Document Automatic Summarization Based On The Term-Sentences—Document Tri-layer Graph Model

In recent years, with the increasing of network information, obtaining information from the World Wide Web has become the most direct and essential channel, but to find the information which you want quickly and effectively from huge amounts of information is difficult to achieve. While, multi-document summarization is an effective solution to this problem, therefore, the research of multi-document summarization has important theoretical significance and practical value.The main task of multi-document automatic summarization is to find the the sentences which can present the theme best in the documents, so how to pick out the most representative sentences has become the most important problem. People always pick the sentences by grading for each sentence in current studies, the sentence scoring method based on graph model is a hotspot in the research now. This method builds a sentence figure in which each sentence as a vertex, the similarity between sentence as the edge weight, the final score which the summaried sentences are selected based on is calculated for each sentence by multiple iterations, then sorting the sentences with limited length in the documents according to the importance degree of all sentences,the more important sentences as the higher score will be feedback to user. The solution, however, is only on the level of sentences to sort sentence, without considering for more information.To solve the above problems existing in graph model, we study the word information in sentence and the document information each sentence belongs to, and try to build a tri-layer graph model which containing word, sentences and documents, in this way, we can improve the quality of multi-document summarization. The experimental result on two multi-document summarization datsset DUC’2003 and DUC’2004 shows that the proposed tri-layer graph model is significantly better than traditional methods.