Language Modelling and Error Handling in Spoken Dialogue Systems
Genevieve Gorrell · 2004
Language modelling for speech recognition is an area of research currently divided between two main approaches: stochastic and grammar-based approaches are each being differently preferred for their respective strengths and weaknesses. At the same time, dialogue systems researchers are becoming aware of the potential value of handling recognition failures better to improve the user experience. This work aims to bring together these two areas of interest, in investigating how language modelling approaches can be used to improve the way in which speech recognition errors are handled.Three practical ways of combining approaches to language modelling in spoken dialogue systems are presented. Firstly, it is demonstrated that a stochastic language model-based recogniser can be used to detect out-of-vocabulary material in a grammar-based system with high precision. Ways in which the technique could be used are discussed. Then, two approaches to providing users with recognition failure assistance are described. In the first, poor recognition results are re-recognised with a stochastic language model, and a decision tree classitier is then used to select a context-specific help message. The approach thereby improves on traditional approaches, where only general help is provided on recognition failure. A user study shows that the approach is well-received. The second differs from the first in its use of layered recognisers and a modified dialogue, and uses Latent Semantic Analysis for the classification part of the task. Decision-tree classification outperforms Latent Semantic Analysis in the work presented here, though it is suggested that there is the potential to improve LSA performance such that it may ultimately prove superior.