Use this URL to cite or link to this record in EThOS:
Title: Named entity extraction for speech
Author: Horlock, James
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 2005
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Named entity extraction, the task of finding specific entities within documents, has proven of great benefit for numerous information extraction and information retrieval tasks. As well as multiple language evaluations named entity extraction has been investigated on a variety of media forms with varying success. In general, these media forms have all been based upon standard text and assumed that any variation from standard text constitutes noise. We investigate how it is possible to find named entities in speech data. Where others have focussed on applying named entity extraction techniques to transcriptions of speech, we investigate a method for finding the named entities direct from the word lattices associated with the speech signal. The results show that it is possible to improve named entity recognition at the expense of word error rate (WER) in contrast to the general trend that F-score is directly proportional to WER. We use a hidden Markov model (HMM) style approach to the task of named entity extraction and show how it is possible to utilise a HMM to find named entities within speech lattices. We further investigate how it is possible to improve results by considering an alternative derivation of the joint probability of words and entities than is traditionally used. This new derivation is particularly appropriate to speech lattices as no presumptions are made about the sequence of words. The HMM style approach that we use requires using a number of language models in parallel. We have developed a system for discriminately retraining these language based upon the results of the output, and we show how it is possible to improve named entity recognition by iterations over both training data and development data. We also consider how part of speech (POS) can be used within word lattices. We devise a method of labelling a word lattice with POS tags and adapt the model to make use of these POS tags when producing the best path through the lattice. The resulting path provides the most likely sequence of words, entities and POS tags and we show how this new path is better than the previous path which ignored the POS tags.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID:  DOI: Not available