Computational Linguistics Colloquium

Friday, 23 October, 9:30
Conference Room, Building C7 4
Please note the unusual date and time!

Natural Language Generation as Planning Under Uncertainty for Spoken Dialogue Systems

I present a new model for Natural Language Generation (NLG) in Spoken Dialogue Systems, based on statistical planning, given noisy feedback from the current generation context (e.g. a user, a surface realiser, and a TTS engine). The model is adaptive and incremental at the turn level, and optimises NLG actions with respect to a data-driven objective function using Reinforcement Learning. We study its use in a standard NLG problem: how to present information (in this case a set of search results) to users, given the complex trade-offs between utterance length, amount of information conveyed, and cognitive load.

I first present results from a proof-of-concept study of this model using data from the MATCH project. We show that the optimised model outperforms several baselines derived from previous work in this area. Furthermore, we study incremental Information Presentation in a Wizard-Of-Oz experiment. Our results indicate that users prefer highly contextual-adaptive NLG behaviour for Information Presentation, and that the wizards' strategy and attribute selection can be modelled as a hierarchical decision process. We currently train our model using this data.

The presented NLG model is part of an end-to-end statistical architecture for Spoken Dialogue Systems, being developed in the CLASSiC project. One major advantage of this architecture is the unified treatment of uncertainty across the entire system. For example, NLG prompts can be re-ranked according to their expected TTS quality (see Boidin et al., Interspeech 2009 [PDF]).
In future work we will test our model in this statistical architecture.

If you would like to meet with the speaker, please contact Magdalena Wolska.