The figure shows the main processing components and interactions with other modules of VERBMOBIL. The basic information type is the dialogue act which represents the main intention of an utterance. Various structures are represented within the module: the dialogue memory stores the sequences of the dialogue's utterances and a thematic structure for the propositional content. The plan recognizer builds up an intentional structure and determines the dialogue phase. The statistic component predicts possible follow-up dialogue acts and turns.

Most knowledge sources of the module are automatically retrieved from a corpus annotated with dialogue information. Currently, approximately 500 dialogues are labeled with this information.

• The dialogue component stores, processes, and provides contextual information for natural language systems;
• It follows a hybrid approach, combining knowledge based and statistical methods;
• It merges results from different processing modes;
• It is robust against unforeseen input;
• It extracts knowlegde sources automatically from annotated corpora.

German Research Center for Artificial Intelligence GmbH Stuhlsatzenhausweg 3 66123 Saarbrücken http://www.dfki.uni-sb.de

Contact Dr. Norbert Reithinger Phone (0 681) 302-5346 Fax (0 681) 302-5341 E-Mail reithinger@dfki.uni-sb.de WWW http://www.dfki.uni-sb.de/verbmobil/