Combining Confidence Scores with Contextual Features for Robust Multi-Device Dialogue

Proceedings of the Australasian Language Technology Workshop 2005, pages 233–240, Sydney, Australia, December 2005. Combining Confidence Scores with Contextual Features for Robust Multi-Device Dialogue∗ Lawrence Cavedon National ICT Australia, Victoria Research Lab and CS&IT, RMIT University Melbourne VIC, Australia lawrence.cavedon@nicta.com.au Matthew Purver, Florin Ratiu CSLI, Stanford University Cordura Hall, 210 Panama St. Stanford CA 94305, USA {mpurver,fratiu}@stanford.edu Abstract We present an approach to multi-device dia- logue that evaluates and selects amongst candi- date dialogue moves based on features at mul- tiple levels. Multiple sources of information can be combined, multiple speech recognition and parsing hypotheses tested, and multiple de- vices and moves considered to choose the high- est scoring hypothesis overall. The approach has the added benefit of potentially re-ordering n-best lists of inputs, effectively correcting er- rors in speech recognition or parsing. A current application includes conversational interaction with a collection of in-car devices. 1 Introduction In this paper, we describe recent enhancements to the CSLI Dialogue Manager (CDM) infras- tructure to increase robustness, in particular in (but not exclusive to) multi-device settings. Dialogue contributions may be processed using multiple information sources (e.g. deep syn- tactic parsing and shallow topic classification), scored at multiple levels (e.g. acoustic, seman- tic and context-based), and bid for by multiple agents, with the overall highest-confidence bid chosen. The CDM provides a multi-device infrastruc- ture, with customization to new applications and addition of plug-and-play devices eased by a declarative dialogue-move scripting language (Mirkovic and Cavedon, 2005). However, decid- ing which device an utterance is directed at is not always straightforward. One of our current application areas is a conversational interface to in-car devices, including entertainment, restau- rant recommendation, navigation and telematic systems (Weng et al., 2004); in such an envi- ronment, a request such as “Play X” might be ∗ This work was performed while all the authors were employed at CSLI, Stanford University, and was par- tially supported by the US government’s NIST Advanced Technology Program. directed at an MP3 player or a DVD player. Eye-gaze (useful in multi-human dialogue) is not available, and we cannot rely on explicit device naming. One option is to use the reso- lution of NP arguments as disambiguating in- formation (in our “Play X” example, whether X is a song or a movie). However, the NP- resolution process itself is often device-specific (see below), preventing NPs from being prop- erly resolved until device has been determined. Our proposed solution, inspired by ap- proaches to multi-agent task allocation such as Contract Net (Smith, 1980), is to allow all devices to perform shallow processing of the incoming utterance, each producing multi- ple possible candidate dialogue moves. Poten- tial device-move combinations are then scored against a number of features, including speech- recognition and parse confidence, discourse con- text, current device-under-discussion, and NP argument analysis. The device associated with the highest-scoring dialogue move is given first option to process the utterance. A disambigua- tion question may be generated if no device is a clear winner, or a confirmation question if the …

Content truncated to 500 words