Tang poetry semantic correlation computing is critical in many applications, such as searching, clustering, automatic generation of poetry and so on. Aiming to increase computing efficiency and accuracy of semantic relatedness, we improved the process of latent semantic analysis (LSA). In this paper, we adopted "representation of words semantic" instead of "words-by-poems" to The ability to quantify semantic relatedness of words in poems should be an integral part of semantic analysis, and underlies many fundamental tasks in NLP, including information retrieval, word sense disambiguation, and text clustering, etc. In contrast to semantic similarity, which is the special case of relatedness, the notion of relatedness is more general than that of similarity like Budanitsky et al [7] argued, as the latter subsumes many different kind of specific relations, including metonymy, antonym, functional association, and others. In this paper we deal with semantic relatedness.Semantic relatedness computing of natural language texts requires encoding vast amount of world knowledge. Until recently, prior work of linguistic resources using pursued two main directions. One is lexical databases such as WordNet [8], Wikipedia [9], encodes relations between words such as synonymy, hypernymy, and the other is large-scale text corpora, provide statistical corpus for computer learning like Latent Semantic Analysis (LSA) [10].But in general computing of modern language semantic relatedness, the least resources used are knowledge-free approaches that rely exclusively on the corpus data themselves. Under the corpus-based approach, word relationships are often derived from their co-occurrence distribution in a corpus [11]. With the introduction of machine readable dictionaries, lexicons, thesauri, and taxonomies, these manually built pseudo-knowledge bases provide a natural framework for organizing words or concepts into a semantic space. Kozima and Furugori [12] measured word distance by adaptive scaling of a vector space generated from LDOCE (Longman Dictionary of Contemporary English). Morris and Hirst [13] used Roget's thesaurus to detect word semantic relationships. With the recently developed lexical taxonomy WordNet [14], many researches have taken the advantage of this broad-coverage taxonomy to study word/concept relationships [15].