Hybrid natural language generation in a spoken language dialog system

Abstract

Natural Language Generation (NLG) systems have almost reached the state of "market-readiness" now, mostly because hybrid systems of different types have emerged as a de-facto standard. But still relatively few dialog systems make use of NLG techniques. In this paper, we discuss the output part of our spoken language dialog system by presenting an example scenario including dialog management, NLG, and speech synthesis. Our approach to hybrid NLG couples shallow and deep processing with respect to the linguistic and pragmatic system resources and also on the architectural level and thus increases processing efficiency (compared to pure deep generation) as well as generative power (compared to pure shallow generation). Our system has been applied to three different domains, namely home A/V management, model train controlling, and B2B e-procurement. © 2004 Springer-Verlag.