A Spoken Language System for Automated Call Routing 3. Language Modeling

Abstract

1. ABSTRACT We are interested in the problem of understanding u-ently spoken language. In particular, we consider people's responses to the open-ended prompt of 'How May I help you?'. We then further restrict the problem to classifying and automatically routing such a call, based on the meaning of the user's response. Thus, we aim at extracting a relatively small number of semantic actions from the utterances of a very large set of users who are not trained to the system's capabilities and limitations. In this paper, we describe the main components of our speech understanding system: the large vocabulary recognizer and the language understanding module performing the call-type classiica-tion. In particular, we propose automatic algorithms for selecting phrases from a training corpus in order to enhance the prediction power of the standard word n-gram The phrase language models are integrated into stochas-tic nite state machines which outperform standard word n-gram language models. From the speech recognizer output we recognize and exploit automatically-acquired salient phrase fragments to make a call-type classiication. This system is evaluated on a database of 10K uently spoken utterances collected from interactions between users and human agents. 2. INTRODUCTION The typical approaches to the problem of topic classi-cation are word and concept spotting. Although these techniques work quite well for small applications, they do not scale up to large tasks and are limited in scope. On the other hand, we view this problem as understanding speech by taking into account the information conveyed by the whole utterance, with the ultimate goal of building automatically trained language models integrating both recognition and understanding. For this reason, we use a large vocabulary speech recognition front-end followed by an understanding module which performs a stochastic mapping between salient fragments and call-types. The problem of automated call routing has been addressed in 2] and the issues concerning the understanding and dialog mechanisms are part of ongoing research ((1], 10]). We have created a database of 10K spoken transactions of people responding to a human agent's greeting of 'How May I help you?' 2]. The rst utterance of each transaction has been transcribed and marked with a call-type by labelers. There are 14 call-types plus an other class as a complement. In particular, we focused our study on the classiication of the user's rst utterance in these dialogs. The spoken sentences vary widely in duration, with a distribution distinctively skewed around a …

Cite this paper

@inproceedings{Riccardi1997ASL, title={A Spoken Language System for Automated Call Routing 3. Language Modeling}, author={Giuseppe Riccardi and Allen L. Gorin and Andrej Ljolje and Martyn Riley}, year={1997} }