Learning new user formulations in automatic Directory Assistance

Abstract

Telecom ltalia has deployed since the beginning of year 2001 a nationwide automatic Directory Assistance (DA) system that routinely serves customers asking for residential and business listings. DA for business listings is a challenging task: one of its main problems is that customers formulate their requests for the same listing with a great variability. Since it is difficult to reliably predict a priori the user formula tions, in this paper we ' propose a procedure for detecting, from field data, user formulations that were not foreseen by the designers. These formulations can be added, as variants, to the denominations already included in the system to reduce its failures.We show that using a large database ass ociating phonetic tran scriptions of user utterances with the phone number provided by the automatic service, a completely unsupervised approach detects most of the old formulations. Furthermore, our proce dure is able to filter a huge amount of calls routed to the op erators, and to detect B limited number of phonetic strings that are good candidate to be included as new formulation variants in the system vocabulary. l. IntroductionDA is the most used service that Telecom operators otTer to their customers. This service is expensive because it relies on a multitude of human operators. A strategy for reducing these costs is to reduce the work time of the operators providing information collected by an Automatic Speech Recognizer, another strategy is to rely as much as possible to complete automation.Several ASR technology providers have worked on problems related to DA, but most of the work has been concentrated 01\ locality and penon name recogniti on [3), [5], for example. This is not surprising because it is difficult to reliably predict a priori the user formulations for business listings. However, since the vast majority of the calls to the DA services are related to business listings, to reduce the service costs, most of the efforts should be devoted to this category of calls.The second strategy has been selected by Telecom Italia, that has deployed since the beginning of year 2001 a nationwide automatic DA system, jointly developed with Loquendo (formerly CSELT), that routinely serves customers asking for residential and business listings. Whenever the automatic system is unable to terminate the transaction with the cus tomer, the call is routed to a human operator. A description of the system, related to the management of the residential calls, has been presented in [1 J. 0-7803-7402-9/02/$17.00 «':)2002 IEEE 1-17The analysis of the traffi c has shown that about 80% of the DA customer accesses are related to business listings. it is important, thus, to improve the percentage of success of the automatic system for this class of calls. The Loquendo approac h to DA for business listing is based on large vocabulary isolated word recognition technology, where the sequence of words of a business listing is concate nated and transcribed as a single word, with possible silences in between . Since the co...

Semantic Frame‐Based Spoken Language Understanding

“…In (Popovici et al 2002), an unsupervised learning algorithm was proposed to obtain the linguistic variants of listings that were not modeled in the Telecom Italia's DA system. A phone-looped model was exploited to obtain the phonetic transcriptions for the utterances that failed the automated service and got routed to the operators.…”

“…To discover these uncovered intents, a language model based acoustic clustering algorithm was proposed. Unlike the algorithm in (Popovici et al 2002) that clusters the one-best phonetic transcriptions, it treats the word transcription and the cluster they belong to as hidden variables, and optimizes the parameters associated with them with respect to an objective function. Specifically, given a fixed number of clusters, it builds a cluster-specific language model P (w | c) and a cluster prior model P (c) to maximize P (x) = c,w P (x, w, c) = c,w P (x | w) P (w | c) P (c), the likelihood of the observed acoustic signal x.…”