【目的】利用文本語義組塊特征提升Cosine文本相似度計算性能。【方法】獲取NSF資助的關于碳納米管研究領域的項目數據,進行詞干還原、詞性標注等預處理;利用條件隨機場模型實現文本內容的語義組塊標注;在此基礎上實現基于語義組塊特征的改進Cosine文本相似度計算,并與未標注的數據進行相似度計算比較,分析實驗結果。【結果】實驗證明基于語義組塊特征的改進Cosine相似度計算結果比原始文本Cosine相似度計算結果相似度均有不同程度的提升,在實驗數據中最高的相似度提升了26%。【局限】依賴于語義組塊標注性能。【結論】本文方法能有效提升文本間語義相似度,降低向量空間模型維度,提高計算效率,并且具有良好的泛化能力和魯棒性。

[Objective] This paper aims to improve the performance of Cosine text similarity computing method with the help of text semantic chunk feature. [Methods] First, we retrieved the project data of carbon nanotubes studies, which were pre-processed with stemming and POS techniques. Then, we identified the semantic chunk of text contents with the conditional random field model. Third, we calculated the similarity of texts based on semantic chunk feature. Finally, we compared our results with those generated by the unlabeled data. [Results] The proposed method improved the performance of Cosine similarity calculation by up to 26%. [Limitations] Our study relies on semantic chunks to annotate the computing performance. [Conclusions] The proposed method could effectively identify similar texts, and reduce the dimensions of vector space model, which improves the computing efficiency. The new method is robust and could be transferred to other fields.