We describe robustness techniques used in the CommandTalk system at the recognition level, the parsing level, and the dialogue level, and how these were influenced by the lack of domain data. We used interviews with subject matter experts (SME's) to develop a single grammar for recognition, understanding, and generation, this eliminating the need for a robust parser. We broadened the coverage of the recognition grammar by allowing word insertions and deletions, and we implemented clarification and correction subdialogues to increase robustness at the dialogue level. We discuss the applicability of these techniques to other domains.
|Title of host publication||Proceedings of the ANLP-NAACL 2000 Workshop on Conversational Systems|
|Place of Publication||Stroudsburg, PA, USA|
|Publisher||Association for Computational Linguistics|
|Number of pages||5|
|Publication status||Published - 2000|