Meaningful Conversation with a Mobile Robot

Meaningful Conversation with a Mobile Robot Johan Bos, Ewan Klein, Tetsushi Oka ICCS, School of Informatics, University of Edinburgh 2 Buccleuch Place, Edinburgh EH8 9LW Scotland, United Kingdom Obos,ewan,okatl@inf.ed.ac.uk Abstract We describe an implementation integrat- ing a spoken dialogue system with a mo- bile robot, which the user can direct to specific locations, ask for information about its status, and supply information about its environment. The robot uses an internal map for navigation, and com- municates its current orientation and ac- cessible locations to the dialogue system using a topological map as interface. 1 Introduction Most research on spoken dialogue has focused on humans talking to virtual agents, often only reached at the end of a telephone line. Interesting challenges and opportunities arise when the inter- locutor is a physically embodied mobile agent— for example, a robot. When we enter into dia- logue with a robot, we can talk about the phys- ical environment that we share with the robot, and we get a palpable indicator of dialogue suc- cess when an utterance such as go to the corri- dor produces the desired effect. In short, spoken dialogue research with mobile robots opens up a new vista for human-computer interaction design which goes beyond the current preoccupation with visual interfaces. In this paper, we give a short overview of the kind of dialogue that can be held with Godot, our robot, together with the architec- ture and technologies that we have implemented. 2 The Spoken Dialogue System The spoken dialogue system allows the user to move Godot by giving it commands, to ask about its current location, or to provide it with new information. This section explains how speech processing, language modelling, language under- standing and dialogue management have been im- plemented. 2.1 Speech Recognition Nuance's speaker-independent speech recognition system (www. . nuance . corn) allows language mod- els to be specified using Speech Grammar rules. Rather than writing these directly, we compile them from a unification grammar for English. Moreover, instead of adopting the Speech Gram- mar slot-filling paradigm for semantic interpreta- tion, our grammar builds a sophisticated, com- positional semantics involving A terms which are passed as the value of a single slot for the recog- nised sentence. As a result the output of the speech recognition stage is a semantic representation, and no further parsing is required before handing it over to the dialogue manager. In order to reduce perplexity, different grammars are loaded at differ- ent states in the dialogue. 2.2 Natural Language Understanding Discourse Representation Theory (DRT) (Kamp and Reyle, 1993) is used for meaning represen- tation in the system. The current implementation covers a wide variety of linguistic phenomena, including context-sensitive phenomena involving anaphora, presupposition, quantification, and plu- ral descriptions. One crucial benefit of DRT is that it supports inference, using a translation from Dis- course Representation Structures (DRSs) to for- mulas of first-order logic. Inference helps to de- tect inconsistencies in the dialogue and assists in disambiguation. Inference invokes standard theorem proving techniques …

Content truncated to 500 words