Natural language generation typically employs templates or grammars. Grammars can be overly complex if they need to cover only limited domains in which a template-based approach may be sufficient. In addition, derivation from grammars may be infeasible in incremental use-cases in which utterances can grow very long. We implemented a template-based incremental natural language generation component that is able to take available timing into account in order to deliver as much information (with as low delay) as possible given the timing constraints of a highly dynamic domain. We present a simulation framework that uses types for template messages, which can be open-ended, and be concatenated with continuations later on. Our system produces more elegant sentences than the nonincremental baseline within a small (but extensible) domain.
The CarChase 2 system was developed at the Natural Language Systems Division (NATS) of the Department of Informatics at the University of Hamburg. It uses the InproTK toolkit (Timo Baumann et al.) for the incremental speech synthesis.
- Download Paper about the CarChase 2 System: the paper includes a description of the CarChase 2 system and its implementation as well as a detailed evaluation including results of a conducted survey.
- Source Code on GitHub: the source repository includes the necessary source code, libraries and instructions to run the CarChase 2 system.