Language Modelling and Error Handling in Spoken Dialogue Systems

Sammanfattning: Language modellingfor speech recognition is an area of research currently dividedbetween two main approaches: stochastic and grammar-basedapproaches are each being differently preferred for theirrespective strengths and weaknesses. At the same time, dialoguesystems researchers are becoming aware of the potential value ofhandling recognition failures better to improve the userexperience. This work aims to bring together these two areas ofinterest, in investigating how language modelling approaches can beused to improve the way in which speech recognition errors arehandled.Three practical ways of combining approaches to language modellingin spoken dialogue systems are presented. Firstly, it isdemonstrated that a stochastic language model-based recogniser canbe used to detect out-of-vocabulary material in a grammar-basedsystem with high precision. Ways in which the technique could beused are discussed. Then, two approaches to providing users withrecognition failure assistance are described. In the first, poorrecognition results are re-recognised with a stochastic languagemodel, and a decision tree classifier is then used to select acontext-specific help message. The approach thereby improves ontraditional pproaches, where only general help is provided onrecognition failure. A user study shows that the approach iswell-received. The second differs from the first in its use oflayered recognisers and a modified dialogue, and uses LatentSemantic Analysis for the classification part of the task.Decision-tree classification outperforms Latent Semantic Analysisin the work presented here, though it is suggested that there isthe potential to improve LSA performance such that it mayultimately prove superior.