Alex: Bootstrapping a Spoken Dialogue System for a New Domain by Real Users

Abstract

When deploying a spoken dialogue system in a new domain, one faces a situation where little to no data is available to train domain-specific statistical models. We describe our experience with bootstrapping a dialogue system for public transit and weather information in real-word deployment under public use. We proceeded incrementally, starting from a minimal system put on a toll-free telephone number to collect speech data. We were able to incorporate statistical modules trained on collected data -in-domain speech recognition language models and spoken language understanding -while simultaneously extending the domain, making use of automatically generated semantic annotation. Our approach shows that a successful system can be built with minimal effort and no in-domain data at hand.

Incremental LSTM-based dialog state tracker

“…6 Throughout this paper we use α = 0.1. 7 The measurements are taken at the end of each dialog turn, provided the component has already been mentioned in some of the SLU n-best lists in the dialog. Note we do not use the SLU n-best list in our model at all, but we adapt this metric to be able to compare to the other trackers in DSTC2.…”

“…For each dialog state component in each dialog, the measurements are taken at the end of each dialog turn 7 .…”

“…However, they still possess two undesirable properties. First, they can only track the dialog state turnby-turn (as opposed to a more complicated word-by-word approach), which limits their interaction with users: For example, in a typical dialog system [7,8], the dialog system can neither provide affirmative natural feedback while the user is speaking, nor can the system interpret additional information said by the user while the system is speaking, both of which is very natural in human-human communication. And second, some of the trackers use an intermediate semantic representation and a spoken language understanding (SLU) component [9].…”

“…We use the live ones and refer to them as live-ASR 6. Throughout this paper we use α = 0.1 7. The measurements are taken at the end of each dialog turn, provided the component has already been mentioned in some of the SLU n-best lists in the dialog.…”

Neural Generation for Czech: Data and Baselines

“…Our goal was to create a dataset comparable in size and domain to existing English data-to-text NLG datasets used in experiments with neural systems. Since there are few to none Czech speakers on crowdsourcing platforms (Pavlick et al, 2014;Dušek et al, 2014), we were not able to use them for data collection. Recruiting freelance translators seemed easier than training annotators; therefore, we turned to localizing and translating an existing dataset instead of creating a new one from scratch.…”