Information extraction from broadcast news.pdf

Information Extractionfrom Broadcast NewsBy Yoshihiko Gotoh and Steve RenalsUniversity of Sheeld, Department of Computer ScienceRegent Court, 211 Portobello Street, Sheeld S1 4DP, UKfy.gotoh,s.renalsg@dcs.shef.ac.ukThis paper discusses the development of trainable statistical models for extract-ing content from television and radio news broadcasts. In particular we concentrateon statistical nite state models for identifying proper names and other named en-tities in broadcast speech. Two models are presented: the rst represents name classinformation as a word attribute; the second represents both word-word and class-class transitions explicitly. A common n-gram based formulation is used for bothmodels. The task of named entity identication is characterized by relatively sparsetraining data and issues related to smoothing are discussed. Experiments are re-ported using the DARPA/NIST Hub{4E evaluation for North American BroadcastNews. Keywords: named entity; information extraction; language modelling1. IntroductionSimple statistical models underlie many successful applications of speech and lan-guage processing. The most accurate document retrieval systems are based on uni-gram statistics. The acoustic model of virtually all speech recognition systems isbased on stochastic nite state machines referred to as hidden Markov models(HMMs). The language (word sequence) model of state-of-the-art large vocabularyspeech recognition systems uses an n-gram model ([n 1]th order Markov model),where n is typically 4 or less. Two important features of these simple models aretheir trainability and scalability: in the case of language modelling, model para-meters are frequently estimated from corpora containing up to 109 words. Theseapproaches have been extensively investigated and optimized for speech recogni-tion, in particular, resulting in systems that can perform certain tasks (e.g., largevocabulary dictation from a cooperative speaker) with a high degree of accuracy.More r