基于词语对狄利克雷过程的时序摘要

时序摘要是按照时间顺序生成摘要, 对话题的演化发展进行概括. 已有的相关研究忽视或者不能准确发现句子中隐含的子话题信息. 针对该问题, 本文建立了一种新的主题模型, 即词语对狄利克雷过程, 并提出了一种基于该模型的时序摘要生成方法. 首先通过模型推理得到句子的子话题分布; 然后利用该分布计算句子的相关度和新颖度; 最后按时间顺序抽取与话题相关且新颖度高的句子组成时序摘要. 实验结果表明, 本文方法较目前的代表性研究方法生成了更高质量的时序摘要.

Temporal Summarization Based on Biterm Dirichlet Process

Temporal summarization aims at extracting sentences chronologically to give an overview about the evolution of a topic. Existing researches either neglect the information of latent subtopics, or fail to accurately discover them. In this paper, we develop a novel topic model called biterm Dirichlet process and generate the temporal summary based on it. Firstly, we get the subtopic distribution in each sentence through posterior inference. Secondly, we calculate each sentence's relevance and novelty degree according to its subtopic distribution. Finally, we chronologically extract the sentences which are relevant and novel to generate the temporal summary. Experiments demonstrate the better performance of our approach compared with currently representative methods.