Prof. Günther Görz
Friedrich-Alexander University (FAU) at Erlangen-Nuremberg

CONALD: A Configurable Plan-Based Dialogue System

Abstract

CONALD is a spoken language dialogue system to conduct rational dialogues addressing an application system, e.g. to collect information, to negotiate about resources, or to control a technical system. To enable flexible, robust and generic dialogue management, we use a plan-based approach to conduct task-oriented dialogues which aim at accomplishing concrete tasks. For linguistic analysis and generation, we claim that a ``pragmatics-first'' view on rational interaction provides an appropriate framework for flexible and scalable dialogue modeling. CONALD combines deep syntactic and semantic analysis, discourse processing, and language generation and features a complex semantics-pragmatics interface. Semantics is defined in terms of an extended version of DRT. Word lattices as produced by a speech recognizer are analyzed by a chunk parser which in turn transforms what can be recognized from the user's utterance into a semantic representation, from which the dialogue system derives a goal for the application. For the reasoning part, i.e. knowledge representation and inference for the interpretation of dialogue as well as for planning to satisfy user goals in the application domain, we argue for a computational logic framework. Discourse and application knowledge are considered independent; user utterances affect operations in an application domain when their speech acts are executed. Dialogue coherence is determined by analyzing the content of contributions to a dialogue with the help of a partial logic: for each new utterance, the dialogue manager evaluates how it contributes to the completion of one of the currently active plans. Due to the clear functional separation between the language model, the dialogue model, and the domain model, the feedback cycle between the dialogue manager and the application is of vital importance, as is the close interaction on the linguistic side between the speech parser, the dialogue manager and the text generation module. CONALD has been implemented in a multi-agent system framework which facilitates quick configurability for various applications.