When processing raw documents in Information Retrieval (IR) System, a term-weighting scheme is used to calculate the importance of each term which occurs in a document. However, most term-weighting schemes assume that a term is independent of the other terms. Term dependency is an indispensable consequence of language use [1]. Therefore, this assumption can make the information of a document being lost. In this paper, we propose new approach to refine term weights of documents using term dependencies discovered from a set of documents. Then, we evaluate our method with two experiments based on the vector space model [2] and the language model [3].

In this paper, we propose a new approach for topic distillation on World Wide Web. Topic distillation is to find quality documents related to the user query topic. Our approach is based on Bharat's topic distillation algorithm [1]. We present the analysis of hyperlink graph structure using hierarchy concept tree to solve the mixed hubs problem that is also remained in the Bharat's algorithm. For assigning better weights to hyperlinks which point to relevant documents among hyperlinks in a document, we try to find the relationship in documents connected by hyperlinks using content analysis and we assign weights to hyperlinks based on the relationship. We evaluated this algorithm using 50 topics on WT10g corpus and obtained improved results.