Extending the SUSI system with negative knowledge

The Philips SUSI system is a set of modules for the construction of automatic inquiry systems such as the train timetable information system TABA and the Philips Automatic Directory Information System (PADIS). These systems conduct a dialogue in continuous spontaneous speech with the user, giving access to information in a database. To improve the system behaviour we have extended the knowledge representation and update functions of the SUSI system so that they collect and use the negative knowledge during the dialogue. Negative knowledge in this context is the information that results from denials by the user in the dialogue, such as "No" and "Not Mr. Smiths". The negative knowledge handling has been implemented and tested for the SUSI based application PADIS. For PADIS, we found that certain dialogues improve significantly because of the use of negative knowledge. INTRODUCTION The Philips SUSI system is a set of modules from which the Philips Automatic Directory Information System (PADIS) [2] and TABA [1] automatic inquiry system have been constructed. In this paper we discuss the extension of the SUSI system with negative knowledge. The PADIS system used for call completion within Philips Research Labs Aachen, will be used to illustrate the effects of the changes in the SUSI modules. PADIS can handle spontaneous continuous speech and is a mixed-initiative dialogue system. In addition to call completion, PADIS can also give information about email addresses, room numbers or telephone numbers. An example PADIS dialogue is dialogue 1. In this paper we focus on the negative knowledge that can be gathered from a PADIS dialogue and used by the system. Negative knowledge in this context is the information that results from denials by the user in the dialogue. Turn U2 in dialogue 1 shows us such a denial. 1 Most PADIS dialogues consist of one or two turns. The average concept recognition rate is 98%. In this paper we give examples of the lengthier dialogues with more recognition errors than usual to illustrate the effects of using negative knowledge. S1: U1: