基于半监督图聚类的项目主题模型构建方法

项目文档主题表征的好坏直接影响后续评审专家的推荐效果.为有效利用项目文档片段之间的关联关系进行项目主题分析,提出一种基于半监督图聚类的项目主题模型构建方法.该方法首先分析项目文档的结构特点,提取项目名称、项目关键字等能表征主题的结构信息,结合专家证据文档、专家主题关系网等能表征专家主题的外部资源,定义及提取项目文档片段之间的关联关系特征;然后,利用不同类型的关联关系计算项目文档片段之间的相关性,构建项目文档片段间的无向图模型;最后,利用已标记关联关系特征作为聚类的监督信息,采用半监督图聚类算法对项目文档片段进行聚类,从而实现项目主题的提取.项目主题提取对比实验结果验证了所提方法的有效性,项目文档结构化特征、专家证据文档以及专家主题关系网对项目主题模型的构建具有一定的指导作用.

Project Topic Model Construction Based on Semi-supervised Graph Clustering

The quality of project topic model has a direct impact on recommended effect of the follow-up evaluation experts.In order to effectively exploit the association relationships among project document fragments to analyze project topics,we proposed a project topic model construction method based on semi-supervised graph clustering.We first analyzed structural characteristics of project documents to extract project name,project keywords and other structural information that responds project topics.Combined with expert evidence documents,expert topic relationship networks and other external resources which can indicate expert topics,we defined and extracted the association relationship features among project document fragments.Then,we used different association relationships to calculate correlation among project document fragments and built undirected graph model for project document fragments.Finally,using the marked association relationship features as supervised information for clustering,we applied semi-supervised graph clustering algorithm to cluster for project document fragments to realize the construction of the project topic model.The comparative experimental results of project topic extraction verify the effectiveness of the proposed method.Structural features of the project documents,expert evidence documents and expert topic relationship networks have certain guidance function for the construction of the project topic model.