Constructing keywords network for query-by-example mode text searching

Text searching, categorization, and summarization are important problems in information retrieval research. One of the most common approaches to text analysis is to exploit the term frequency-inverse document frequency (tf-idf) vector model, which is very effective and efficient in representing a large document through a small vector. The tf-idf approach has the crucial drawback that it only considers the text in terms of the structure of composition. However, each natural language has its own syntactic structure. Thus, it is not sufficient to replace the text with a set of important keywords without taking into account their relative relation to the thesis and meaning of the text. In this paper, we propose a text search model based on a keyword graph model, which is based on the cognitive process (writing) model. We show how to construct a keyword graph from a text by assigning edges between two vertices (keywords) if their regions of influence overlap. Our approach allows the use of the text as a query attribute. In our model, if a user wants to find text similar to a given query text in a large repository, the query document can be searched without selecting keywords. This query-by-example in text searching is an important contribution of our work. Experiments show that our keyword graph model is superior to the tf-idf model in clearly and effectively revealing the similarity between documents. Our experiments use more than 2,000 speeches obtained from the United States White House, and show that our approach is superior to prevalent text search methods in terms of accuracy of syntactic similarity and the semantic structure of object texts.