一种基于词语能量值变化的微博热点话题发现方法研究

摘要

With the popularity of microblog, hot topic detection on microblog has been a hot area of research. Regarding the instantaneity of microblog as a point of penetration, the paper proposes a method of hot topic detection based on change of term energy by studying the change of term energy at different time domain. Based on traditional topic aging theory, the method divides all microblog data into different microblog windows, and introduces the concept of acceleration in physics, which uses the acceleration of terms to describe the change of the speed of the terms in the adjacent window. The paper combines the term acceleration and term weight into a compound weight to quantize term energy better. The paper uses double-conditional probability context similarity computing method based on single-conditional probability, and adds document distribution similarity to decrease the probability of topic confusion. The experiments show that the method is effective and stable in robustness. Compared with single-conditional probability context similarity model, the modiifed context similarity model has better clustering effect in different keyword detection methods.%随着微博的迅速发展，微博上的热点话题发现成为目前的研究热点之一。文章以微博的实时性强作为研究的切入点，通过研究不同时域上词语的能量值变化，提出一种基于词语能量值变化的微博热点话题检测方法。该方法基于传统的话题生命周期理论，按微博的时间先后顺序对微博进行划分；引入了物理学科中加速度的概念，用词语的加速度来刻画词语在相邻窗口之间速度的变化；综合考虑词语的加速度和权重值来构造词语的复合权值，更适合量化词语的能量值；在单条件概率的基础上，使用了双条件概率的上下文相似度计算方法，并增加文档分布相似度来减少话题混淆的概率。实验表明了文章方法的有效性和稳定鲁棒性。与单条件概率的上下文相似度模型相比，改进之后的上下文相似度模型在不同的关键词检测方法中均具有更好的聚类效果。