Using SOM based graph clustering for extracting main ideas from documents

In this paper, we would like to present a graph clustering system for grouping the similar documents and extracting the main ideas in documents. To cluster the documents, we need a model for representing the documents. The traditional approaches used a word set based model or a vector based model for representing the documents. These models discard the important structural information of documents such as word position, the semantic relations of words in document... Recently, some research works using the graph for representing the documents have been appeared. We use the graph to be created by analyzing the co-occurrence and position of two words in a section of document. After representing the documents by using graph, we used self organizing map (SOM) with two-dimensional output layer for grouping the graphs. One of the advantages of SOM is to cluster the data without specifying the number of clusters. Besides, two-dimensional SOM output layer can be put on the computer display and it can help to access the similar documents on the computer display. We use the graph distance based on the maximum common sub-graph (mcs) which is discovered by maximal frequent sub-graph algorithm and the updated operation of neurons on SOM ouput layer based on the weighted means graphs and the genetic algorithm.