基于贝叶斯信念网络的话题识别模型

Topic detection model based on Bayesian belief network

针对贝叶斯信念网络应用于话题识别进行了研究,提出了新的话题识别模型。模型的拓扑结构包括新报道、报道术语、事件术语、话题四层节点,用弧标明索引关系。在贝叶斯概率和条件独立性假设的基础上,模型运用条件概率计算新报道和已有话题簇的相似度,从而实现话题识别。考虑到核心报道、核心事件的重要性,对不同层次的权重计算进行了调整。实验采用DET曲线评测法对模型性能进行测试,实验结果显示,调整后的权重计算可在一定程度上提高新模型的性能,与向量空间模型相比,在相同阈值下新模型的漏报率与误报率有所降低。

According to the research of Bayesian belief network was applied to topic detection, this paper proposed a new topic detection model. The topology of the new model included four level nodes ： new story, story term, event term and topic, arcs indicated the indexing relationships. To achieve the task of topic detection, the new model applied conditional probability based on Bayesian probability and conditional independence assumption to compute the similarity between new story and topic clus- ters. Considering the importance of seminal stories and seminal events,it adjusted weight computations in different levels,and evaluated the new model and the vector space model by DET curves. Experimental results show that adjusted weight computa- tions will improve ＇the performance of the new model, and at the same threshold, the new model has lower miss probability and false alarm probability compared to the vector space model.