A Risk Minimization Framework for Extractive Speech Summarization

Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, pages 79–87, Uppsala, Sweden, 11-16 July 2010. c⃝2010 Association for Computational Linguistics A Risk Minimization Framework for Extractive Speech Summarization Shih-Hsiang Lin and Berlin Chen National Taiwan Normal University Taipei, Taiwan {shlin, berlin}@csie.ntnu.edu.tw Abstract In this paper, we formulate extractive summarization as a risk minimization problem and propose a unified probabilis- tic framework that naturally combines su- pervised and unsupervised summarization models to inherit their individual merits as well as to overcome their inherent limita- tions. In addition, the introduction of vari- ous loss functions also provides the sum- marization framework with a flexible but systematic way to render the redundancy and coherence relationships among sen- tences and between sentences and the whole document, respectively. Experi- ments on speech summarization show that the methods deduced from our framework are very competitive with existing summa- rization approaches. 1 Introduction Automated summarization systems which enable user to quickly digest the important information conveyed by either a single or a cluster of docu- ments are indispensible for managing the rapidly growing amount of textual information and mul- timedia content (Mani and Maybury, 1999). On the other hand, due to the maturity of text sum- marization, the research paradigm has been ex- tended to speech summarization over the years (Furui et al., 2004; McKeown et al., 2005). Speech summarization is expected to distill im- portant information and remove redundant and incorrect information caused by recognition er- rors from spoken documents, enabling user to efficiently review spoken documents and under- stand the associated topics quickly. It would also be useful for improving the efficiency of a num- ber of potential applications like retrieval and mining of large volumes of spoken documents. A summary can be either abstractive or extrac- tive. In abstractive summarization, a fluent and concise abstract that reflects the key concepts of a document is generated, whereas in extractive summarization, the summary is usually formed by selecting salient sentences from the original document (Mani and Maybury, 1999). The for- mer requires highly sophisticated natural lan- guage processing techniques, including semantic representation and inference, as well as natural language generation, while this would make ab- stractive approaches difficult to replicate or ex- tend from constrained domains to more general domains. In addition to being extractive or ab- stractive, a summary may also be generated by considering several other aspects like being ge- neric or query-oriented summarization, single- document or multi-document summarization, and so forth. The readers may refer to (Mani and Maybury, 1999) for a comprehensive overview of automatic text summarization. In this paper, we focus exclusively on generic, single- document extractive summarization which forms the building block for many other summarization tasks. Aside from traditional ad-hoc extractive sum- marization methods (Mani and Maybury, 1999), machine-learning approaches with either super- vised or unsupervised learning strategies have gained much attention and been applied with empirical success to many summarization tasks (Kupiec et al., 1999; Lin et al., 2009). For super- vised learning strategies, …

Content truncated to 500 words