In this study, relation measures between words and between sentences were integrated with the Latent Semantic Analysis to construct the relational matrix between words and sentences. Such a novel approach improved the semantic analysis of the similarity between sentences, aiding the extraction of the representative sentences of the document for text mining. In the experiment, thirteen categories of documents were applied and two performance indices including sentence similarity and document classification rate were used to evaluate the proposed idea. The experimental results show that the proposed method is capable of extracting less similar and more representative sentences of the document.