融合扩展信息瓶颈理论的话题关联检测方法研究

话题关联检测的关键任务在于判断给定报道对是否属于同一话题. 现有判断方法往往忽略种子事件与其直接相关事件之间的层次关系．为此，通过分析报道内部语义分布规律及篇章结构，并依据语义分布规则，利用语义分布规律改进信息瓶颈（Information bottleneck，IB）算法，用于子话题逻辑语义单元的划分，并利用这些逻辑语义单元表示报道，进行话题关联检测. 实验证明该方法有较快的收敛速度，并在一定程度上提高了系统性能．

A Topic Link Detection Method Based on Improved Information Bottleneck Theory

Topic link detection aims to detect whether two given stories talk about the same topic, whose key task is how to represent the story utilizing a proper model. In the previous works, the hierarchical relationship between seed events and its directly related events is ignored. Thus, this paper analyzes the regular pattern of semantic distribution and the structure of a story, and proposes a method to divide a story into several sections of sub-topic features based on the regular pattern of semantic distribution and improved information bottleneck (IB) theory. Then, the story represented by the attributes is utilized to do topic link detection. Experimental result shows that this method has a fast convergent rate, and can improve the performance of the system.