How may I help you?

Abstract

We are interested in providing automated services via natural spoken dialog systems. By natural, we mean that the machine understands and acts upon what people actually say, in contrast to what one would like them to say. There are many issues that arise when such systems are targeted for large populations of non-expert users. In this paper, we focus on the task of automatically routing telephone calls based on a user's fluently spoken response to the open-ended prompt of “How may I help you?”. We first describe a database generated from 10,000 spoken transactions between customers and human agents. We then describe methods for automatically acquiring language models for both recognition and understanding from such data. Experimental results evaluating call-classification from speech are reported for that database. These methods have been embedded within a spoken dialog system, with subsequent processing for information retrieval and formfilling.

Résumé

Nous sommes intéressés par la production de services automatisés par des systèmes de dialogue utilisant la parole naturelle. Nous entendons par naturel que la machine comprend et agit selon ce que les personnes effectivement disent, en opposition à ce que l'on aimerait qu'ils disent. Plusieurs problèmes apparaissent quand de tels systèmes sont visés pour une population large d'utilisateurs qui ne sont pas des experts. Dans ce papier, nous focalisons sur la tâche de routage automatique des appels téléphoniques se basant sur la réponse spontanée des utilisateurs à la question ouverte “How may I help you?”. Nous décrivons d'abord la base de données générées par 1000 transactions orales entre des utilisateurs et des agents humains. Nous décrivons ensuite les méthodes pour l'acquisition automatique, à partir des données, des modèles de langage pour la reconnaissance et la compréhension. Les résultats expérimentaux pour l'évaluation de la classification des appels sont rapportés pour cette base de données. Ces méthodes ont été incorporées dans un système de dialogue oral avec des traitements subséquents pour le tri des informations et le remplissage des formes.

Introduction

There are a wide variety of interactive voice systems in the world, some residing in laboratories, many actually deployed. Most of these systems, however, either explicitly prompt the user at each stage of the dialog, or assume that the person has already learned the permissible vocabulary and grammar at each point. While such an assumption is conceivable for frequent expert users, it is dubious at best for a general population on even moderate complexity tasks. In this work, we describe progress towards an experimental system which shifts the burden from human to machine, making it the device's responsibility to respond appropriately to what people actually say.

The problem of automatically understanding fluent speech is difficult, at best. There is, however, the promise of solution within constrained task domains. In particular, we focus on a system whose initial goal is to understand its input sufficiently to route the caller to an appropriate destination in a telecommunications environment. Such a call router need not solve the user's problem, but only transfer the call to someone or something which can. For example, if the input is “Can I reverse the charges on this call?”, then the caller should be connected to an existing automated subsystem which completes collect calls. Another example might be “How do I dial direct to Tokyo?”, whence the call should be connected to a human agent who can provide dialing instructions. Such a call router should be contrasted with traditional telephone switching, wherein a user must know the phone number of their desired destination, or in recent years navigate a menu system to self-select the desired service. In the method described here, the call is instead routed based on the meaning of the user's speech.

This paper proceeds as follows. In Section 2, an experimental spoken dialog system is described for call-routing plus subsequent automatic processing of information retrieval and form-filling functions. The dialog is based upon a feedback control model, where at each stage the user can provide both information plus feedback as to the appropriateness of the machine's response (Gorin, 1995a). In Section 3, a database is described of 10 K fluently spoken transactions between customers and human agents for this task. In particular, we describe the language variability in the first customer utterance, responding to the prompt of “How may I help you?” in a telecommunications environment.

In Section 4, we describe the spoken language understanding (SLU) algorithms which we exploit for call classification. A central notion in this work is that it is not necessary to recognize and understand every nuance of the speech, but only those fragments which are salient for the task (Gorin, 1995a). This leads to a methodology where understanding is based upon recognition of such salient fragments and combinations thereof.

There are three main components in our SLU methodology. First is to automatically acquire salient grammar fragments from the data, modeling those parts of the language which are meaningful for the task plus their statistical associations to the machine actions. Second is to recognize these fragments in fluent speech, searching the output of a large vocabulary speech recognizer. The statistical language model which constrains this recognizer embeds automatically-acquired fragments in a stochastic finite state machine, providing an efficient approximation to an n-gram model with variable length units (Riccardi et al., 1996). Third, we exploit these multiple recognized fragments to classify the call-type of an utterance. Since the SLU is embedded within a dialog system, the classifier provides both the best (rank 1) and secondary (rank 2, etc.) decisions. Finally, in Section 5, we report on experimental results for call-classification from the above-mentioned speech database, training on 8 K utterances and testing on 1 K.