Paper similarity detection depends on grammatical and semantic analysis, word segmentation, similarity detection, document summarization and other technologies, involving multiple disciplines. However, there are some problems in the existing main detection models, such as incomplete segmentation preprocessing specification, impact of the semantic orders on detection, near-synonym evaluation, difficulties in paper backtrack and etc. Therefore, this paper presents a two-step segmentation model of special identifier and Sharpley value specific to above problems, which can improve segmentation accuracy. In the aspect of similarity comparison, a distance matrix model with row-column order penalty factor is proposed, which recognizes new words through search engine exponent. This model integrates the characteristics of vector detection, hamming distance and the longest common substring and carries out detection specific to near-synonyms, word deletion and changes in word order by redefining distance matrix and adding ordinal measures, making sentence similarity detection in terms of semantics and backbone word segmentation more effective. Compared with the traditional paper similarity retrieval, the present method has advantages in accuracy of word segmentation, low computation, reliability and high efficiency, which is of great academic significance in word segmentation, similarity detection and document summarization.