科技情报分析中LDA主题模型最优主题数确定方法研究

Identifying Optimal Topic Numbers from Sci-Tech Information with LDA Model

【目的】有效确定科技情报分析中LDA主题模型的最优主题数目。【方法】利用主题相似度度量潜在主题之间的差异,同时结合困惑度提出一种确定LDA最优主题数目的方法,该方法既考虑主题抽取效果同时也考虑模型对新文档的泛化能力。【结果】获取国内新能源领域的科技文献作为数据集,实证结果表明本文提出的最优LDA主题数确定方法与单纯使用困惑度相比,具有更高的主题抽取查准率(91.67%)、F值(86.27%)及科技文献推荐精度(71.25%)。【局限】未针对其他类型的数据集进行新方法的验证,如微博短文本、XML文档等。【结论】本文方法能够有效地从科技文献数据集中抽取辨识度较高的主题,并能够提高科技文献推荐效果。

[Objective] This paper tries to identify the optimal number of topics for the Latent Dirichlet Allocation （LDA） model to analyze scientific and technical information. [Methods] First, we used the topic similarity to measure the differences among the latent topics. Second, we proposed a method determining the optimal topic numbers and tried to utilize this model to documents from Chinese literature in the field of new energy. [Results] The proposed method achieved higher precision ratio and higher F-score in topic extration, which improved the performance of literature recommendation systems. [Limitations] We did not examine the new mothod with other datasets, such as microblog posts and XML documents. [Conclusions] The proposed method could identify more recognizable topics and improve the performance of scientific and technical literature recommendation systems.