Topic-based Multi-Document Summarization with Probabilistic Latent Semantic Analysis

International Conference RANLP 2009 - Borovets, Bulgaria, pages 144–149 Topic-based Multi-Document Summarization with Probabilistic Latent Semantic Analysis Leonhard Hennig DAI Labor, TU Berlin Berlin, Germany leonhard.hennig@dai-labor.de Abstract We consider the problem of query-focused multi- document summarization, where a summary containing the information most relevant to a user’s information need is produced from a set of topic-related documents. We propose a new method based on probabilistic latent semantic analysis, which allows us to represent sentences and queries as probability distributions over la- tent topics. Our approach combines query- focused and thematic features computed in the latent topic space to estimate the summary- relevance of sentences. In addition, we evaluate several different similarity measures for comput- ing sentence-level feature scores. Experimental results show that our approach outperforms the best reported results on DUC 2006 data, and also compares well on DUC 2007 data. Keywords text summarization, probabilistic latent semantic analysis, plsa 1 Introduction Automatically producing summaries from large tex- tual sources is an extensively studied problem in IR and NLP [17, 12]. In this paper, we investigate the problem of multi-document summarization, where a summary is created from a set of related documents and optionally fulfills a specific information need of a user. In particular, we focus on generating an extrac- tive summary by selecting sentences from a document cluster [8]. Multi-document summarization is an in- creasingly important task: With the rapid growth of online information, and many documents covering the same topic, the condensation of information from dif- ferent sources into an informative summary helps to reduce information overload. Automatically created summaries can either consist of the most important information overall (generic summarization) or of the information most relevant with respect to a user’s in- formation need (query-focused summarization). A major aspect of identifying relevant information is to find out what a text is about. A document will generally contain a variety of information cen- tered around a main theme, and covering different aspects of the main topic. Similarly, human sum- maries tend to cover different topics of the original source text to increase the informative content of the summary. Various approaches have exploited features based on the identification of topics (or thematic foci) to construct generic or query-focused summaries. Of- ten, thematic features rely on identifying and weight- ing important keywords [21], or creating topic signa- tures [14, 10]. Sentences are scored by combinations of keyword scores, or by computing similarities between sentences and queries. Yet it is well known that term matching has severe drawbacks due to the ambivalence of words and to differences in word usage and personal style across authors. This is especially important for automatic summarization, as summaries produced by humans may differ significantly, potentially not shar- ing very many terms [16]. Latent Semantic Indexing (LSI) is an approach to overcome these problems by mapping documents to a latent semantic space, and has been shown to work well for text summarization [9, 23]. However, LSI has a number of drawbacks, namely its unsatisfactory sta- tistical foundations. …

Content truncated to 500 words