Robust Named Entity Extraction from Large Spoken Archives

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 491–498, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Robust Named Entity extraction from large spoken archives Benoˆıt Favre Thales, MMP Laboratory Colombes, France benoit.favre@ fr.thalesgroup.com Fr´ed´eric B´echet LIA, University of Avignon Avignon, France frederic.bechet@ univ-avignon.fr Pascal Noc´era LIA, University of Avignon Avignon, France pascal.nocera@ univ-avignon.fr Abstract Traditional approaches to Information Ex- traction (IE) from speech input simply consist in applying text based methods to the output of an Automatic Speech Recog- nition (ASR) system. If it gives satis- faction with low Word Error Rate (WER) transcripts, we believe that a tighter inte- gration of the IE and ASR modules can increase the IE performance in more dif- ficult conditions. More specifically this paper focuses on the robust extraction of Named Entities from speech input where a temporal mismatch between training and test corpora occurs. We describe a Named Entity Recognition (NER) system, de- veloped within the French Rich Broad- cast News Transcription program ESTER, which is specifically optimized to pro- cess ASR transcripts and can be integrated into the search process of the ASR mod- ules. Finally we show how some meta- data information can be collected in or- der to adapt NER and ASR models to new conditions and how they can be used in a task of Named Entity indexation of spoken archives. 1 Introduction Named Entity Recognition (NER) is a crucial step in many Information Extraction (IE) tasks. It has been a specific task in several evaluation pro- grams such as the Message Understanding Confer- ences (MUC), the Conferences on Natural Language Learning (CoNLL), the DARPA HUB-5 program or more recently the French ESTER Rich Transcription program on Broadcast News data. Most of these conferences have studied the impact of using tran- scripts generated by an Automatic Speech Recogni- tion (ASR) system rather than written texts. It ap- pears from these studies that unlike other IE tasks, NER performance is greatly affected by the Word Error Rate (WER) of the transcripts processed. To tackle this problem, different ideas have been pro- posed: modeling explicitly the ASR errors (Palmer and Ostendorf, 2001) or using the ASR system alternate hypotheses found in word lattices (Sar- aclar and Sproat, 2004). However performance in NER decreases dramatically when processing high WER transcripts like the ones that are obtained with unmatched conditions between the ASR train- ing model and the data to process. This paper in- vestigates this phenomenon in the framework of the NER task of the French Rich Transcription program of Broadcast News ESTER (Gravier et al., 2004). Several issues are addressed: • how to jointly optimize the ASR and the NER models ? • what is the impact in term of ASR and NER performance of a temporal mismatch between the corpora used to train and test the models and how can it be recovered by means of meta- data information ? • Can metadata information be used for indexing large …

Content truncated to 500 words