Update Summarization using a Multi-level Hierarchical Dirichlet Process Model

Proceedings of COLING 2012: Technical Papers, pages 1603–1618, COLING 2012, Mumbai, December 2012. Update Summarization Using a Multi-level Hierarchical Dirichlet Process Model Jiwei Li1 Sujian Li1 Xun Wang1 Ye Tian1 Baobao Chang1 (1) Key Laboratory of Computational Linguistics, Peking University, Ministry of Education, CHINA {bdlijiwei, lisujian, xunwang, ytian, chbb}@pku.edu.cn ABSTRACT Update summarization is a new challenge which combines salience ranking with novelty detection. Previous researches usually convert novelty detection to the problem of redundancy removal or salience re-ranking, and seldom explore the birth, splitting, merging and death of aspects for a given topic. In this paper, we borrow the idea of evolutionary clustering and propose a three-level HDP model named h-uHDP, which reveals the diversity and commonality between aspects discovered from two different epochs (i.e. epoch history and epoch update). Specifically, we strengthen modeling the sentence level in the h-uHDP model to adapt to the sentence extraction based framework. Automatic and manual evaluations on TAC data demonstrate the effectiveness of our update summarization algorithm, especially from the novelty criterion. KEYWORDS : Update summarization, Hierarchical Dirichlet process, Novelty detection. 1603 1 Introduction Update summarization aims to generate a short and concise summary for the latest updating topic-related documents (hereafter update documents for short), under the assumption that the user has already read the earlier historical documents (history documents for short) about the same topic. Recently, there have been many attempts to explore different approaches to generate update summaries. The predominant approaches are mainly built upon the sentence extraction framework. Update summarization for an evolving topic differs from previous generic summarization for a static topic in that the latter aims to acquire the salient information in one topic, while the former cares for both the salience and the novelty of information. By developing traditional summarization techniques, massive efforts on update summarization have been made to dig out new information (Boudin et al., 2008; Fisher and Boark, 2008; Wan, 2007; Li et al., 2008; Du et al., 2010; Li et al., 2012). The typical examples include the scaled Maximal Marginal Relevance (MMR) algorithm which excludes those sentences similar to the history documents, and some extensions of TextRank such as TimedTextRank (Wan, 2007), PNR2 (Li et al., 2008), MRSP (Du et al., 2010) which re-rank the salience scores of sentences by employing various kinds of reinforcement between sentences. One problem with these approaches is that they tend to regard update summarization more as a redundancy removal problem than a novelty detection problem. Another problem is that these approaches are mainly based on the computation of lexical similarities between sentences and fail to consider higher level information to avoid semantic redundancy in update summarization. To solve these two problems, we borrow the techniques of evolutionary clustering which focuses on detecting the dynamics of a given topic. Normally, one topic is described from various specific aspects 1 , accompanied with the background information running the whole topic (Chemudugunta et al., 2007; Li et al., 2010). For example, the topic “Quebec independence” may involve the specific aspects …

Content truncated to 500 words