Automatic induction of language model data for a spoken dialogue system

Abstract

When building a new spoken dialogue application, large amounts of domain specific data are required. This paper addresses the issue of generating in-domain training data when little or no real user data are available. The twostage approach taken begins with a data induction phase whereby linguistic constructs from out-of-domain sentences are harvested and integrated with artificially constructed in-domain phrases. After some syntactic and semantic filtering, a large corpus of synthetically assembled user utterances is induced. The second stage involves sampling the synthetic corpus towards the goal of obtaining data that would be representative of the statistics of applicationspecific real user interactions. The sampling methods proposed employ an example-based generation framework, a simulated user model and information extracted from development data. Evaluation is conducted on recognition performance in a restaurant information domain. We show that word error rate can be reduced when limited amounts of real user training data are augmented with synthetic data derived by our methods.

Semantic Class Induction for Language Model Adaptation in a Chinese Voice Search System

“…For statistical language model adaptation, there are a lot of works (see [11] for a nice review). Recently there have been a few studies that are oriented toward dialogue systems, e.g., [12,13,14]. One of the recent trends is to exploit web data [15,13], esp.…”

“…But the data from the web is less likely to be truly in-domain but near in-domain. Wang et al [14] transformed out-of-domain data and then went through filtering and re-sampling to obtain near in-domain data. We use no out-of-domain data and get data only through induction and then generation.…”

“…There are some language model adaptation (LMA) work oriented to the dialogue systems e.g. Wang et al(2006), Hakkani-Tür et al(2006), Bellegarda(2004. So far major effort has been spent on adaptation for large vocabulary speech recognition or transcription tasks.…”

Language model adaptation using auto-induced semantic structures in a voice search system

“…There are some LMA work oriented to the dialogue systems [10], [11], [12]. In our experiments, a trigram model was trained on the large text collection (T).…”

“…E.g., Wang et al took a two-step approach to generate in-domain data through out-of-domain data transformation [12]. [15] extract useful information from previous domains and World Wide Web without any in-domain data.…”