Incremental Generation of Self-corrections Using Underspecification

Abstract

The focus of the present paper is on the interface between the first two components of a language production model, i.e. conceptualiser and formulator. We propose a method of how the incremental input for the formulator can be represented within an underspecification formalism. The proposed method can be used in dynamically changing environments, e.g. an airport supervision and information system, to produce descriptions of ongoing events. Because such a highly dynamic setting requires a real time production, an incremental conceptualiser that produces increments piece by piece is used in order to speed up the production process. Moreover, changes in the environment have to be recognised and quickly reported to the user. For this reason an underspecified representation for the increments is crucial. Underspecification allows information to be added monotonically as soon as a change in the environment has been noticed, e.g. 'CK-314 is on time . . . uh . . . is delayed.' IntroductionInteractive dialogue systems have to cope with numerous challenges. They can only be considered to adequately simulate a human-human-like dialogue if they meet a high standard regarding robustness and naturalness of natural language understanding and production. We are concerned here with the production side of this enterprise and propose a method of how already generated information can be dynamically corrected. Current dialogue systems lack the following vital requirements for building a robust and natural system: (1) the system needs to react quickly and appropriately to changes in the environment, (2) the output should be generated similarly to what a user would expect from a human communication partner.As part of fulfilling these requirements, we employ incrementality as a processing principle for our model of language production and propose underspecification as the underlying representation formalism. More specifically, we build on earlier work on the incremental conceptualiser INC (Guhe and Habel 2001) and employ the underspecification formalism CLLS (Constraint Language for Lambda Structures) (Egg, Koller and Niehren 2001) in order to produce self-corrections.Incrementality is advantageous for systems in a real time processing environment, because smaller pieces (increments) can already be processed by subsequent components before the entire planning process is finished (Kempen and Hoenkamp 1987, Levelt 1989, De Smedt 1990). An underspecification formalism is well suited for incremental language generation, because an underspecified representation can easily be extended without overwriting earlier made planning decisions: subsequently added increments can be inserted where the underspecified representation left room.

Incremental Coordination: Attention-Centric Speech Production in a Physically Situated Conversational Agent

“…For instance, Skantze and Hjalmarsson (2010) showed that an incremental generator using filled pauses and self-corrections achieved (in a wizard of Oz experiment) shorter response times and was perceived as more efficient than a non-incremental generator. Guhe and Schilder (2002) have also used incremental generation for self-corrections.…”

Incremental generation of interconnected preverbal messages

“…24. Other discourse refOs are used in Guhe and Schilder (2002a), where they represent correction terms. These in turn are similar to discourse refOs establishing coordination relations in the generation of VP ellipses, simply called list in Guhe and Schilder (2002b).…”