Positional Language Modeling for Extractive Broadcast News .ppt

Positional Language Modeling for Extractive Broadcast News Speech Summarization Shih-Hung Liu, Kuan-Yu Chen, Berlin Chen, Hsin-Min Wang, Hsu-Chun Yen, Wen-Lian Hsu Academia Sinica National Taiwan University National Taiwan Normal University Introduction Language Modeling Framework for Summarization Traditional LM Relevance Modeling Positional Language Modeling Passage-based LM Position-specific LM Positional Relevance Modeling Experiments Conclusions Outline 2 Introduction Language Modeling Framework for Summarization Traditional LM Relevance Modeling Positional Language Modeling Passage-based LM Position-specific LM Positional Relevance Modeling Experiments Conclusions Outline 3 Content Creation vs. Content Management Too much information kills information! Automatic summarization figures prominently in dealing with the information overload problem Facilitate people to browse multimedia documents and distill their themes both efficiently and effectively Introduction: Information Overload 4 Speech Multimedia Text Some of the above figures are adapted from the presentation slides of Prof. Nenkova et al. at ACL 2011 Spectrum of Summarization Research 5 S.-H. Lin, “Speech Summarization - Features, Models and Applications,” Ph.D. Dissertation, 2011 The wide array of extractive speech summarization methods may roughly fall into three main categories: Approaches simply based on sentence structure or position information LEAD Approaches based on unsupervised machine-learning Vector-based methods VSM, LSA, CBOW, NMF, MMR Graph-based methods TextRank, LexRank, MRW, MDS Combinatorial optimization-based methods ILP, Submodular Language modeling framework Approaches based on supervised machine-learning GMM, BC, SVM, CRF Introduction 6 For more detailed review on the developments automatic summarization, please also refer to Nenkova and McKeown , “Automatic Summarization,” Foundations and Trends in Information Retrieval, 2011 Introduction Language Modeling Framework for Su