A System for Rapid Development of Large Scale Rule Bases for Template-Based NLG for Conversational Agents

Tim Krones

Mastersthesis Saarland University 11/2014.


Long-term social interaction between conversational agents and users requires adaptivity and variation of system output (Kruijff-Korbayová et al. 2011, Kruijff-Korbayová et al. 2012): Agents should produce natural language output that is appropriate and relevant, and should not repeat themselves in recurring situations. In the context of the ALIZ-E project (, designing natural language output involves writing rules which match abstract representations of situational knowledge to appropriate verbal responses. We present a new graphical system for collaborative creation, maintenance, and long-term evolution of large-scale rule bases that addresses many challenges involved in working with these rules: Specialized editing features facilitate fast creation of large amounts of variation and help minimize errors that are likely to occur when editing plain text representations of rules. By abstracting away from native rule syntax, the system aims to make the process of working with rules accessible to non-experts. Support for searching and filtering rule bases in various ways facilitates maintenance and long-term evolution. Results of a first set of evaluation experiments suggest that using our system, non-experts can gain a basic conceptual understanding of rules and productively design natural language output with relatively little training.


thesis-krones-final.pdf (pdf, 2 MB )

Deutsches Forschungszentrum für Künstliche Intelligenz
German Research Center for Artificial Intelligence