Project Summary: 
Natural Lanfuaoe Understandinf: Intcoratin¢ S~/ntaz, Semantics, and Discourse. 
Institution: UNISYS 
Principal Investigators: Dr. Lynette Hirschman, Dr. Martha Palmer 
Technical Summary: The focus of the UNISYS research is on integrating multiple knowledge sources, 
including syntax, semantics, pragmatics, a domain model, and both domain-specific and domain- 
independent knowledge sources, to produce a system capable of understanding messages in a res- 
tricted domain. 
Objeeti~ea: 
(1) Creation of an integrated, portable system for the detailed understanding of multi-paragraph 
text and messages. 
(2). Develop a principled computational treatment of semantic and discourse phenomena, includ- 
ing treatment of reference and temporal information. 
(3) Development of an integrated loglc-based analysis algorithm which uses syntactic, semantic and 
discourse information to maximize robustness and search focus. 
(4) Build a spoken language system through integration of a large natural language system with a 
speech recognition system. 
Recent Accomplishments: 
* System: Development of modular, interleaved system architecture, enhancing portability by 
enforcing segregation of core PUNDIT from domaln-speclfic modules. Demonstrated by port 
to domain of Navy Trouble Failure Reports. 
• Applications: Processing of CAS17~P messages, producing tabular summary; Navy RAINFORM 
(sighting messages), and Trouble and Failure reports, producing database updates. 
• Technology Transfer Activity: PUNDIT User's Guide released, system installed at National 
Library of Medicine, Unisys applications group; system requested by RADC, CECOM, Harvard, U. 
Penn., Cambridge U., Swedish Inst. of Computer Science. 
• Treatment of fragmentary input as found in message traffic, by means of minimal extensions to 
the syntactic, semantic and pragmatic components. 
• Semantic coverage of nominallzed verbs, adjectival participles and noun predicates, based on 
the treatment of the underlying predicate, using reference resolution and temporal processing to 
complete the interpretation. 
• Interactive selection component, interleaved with syntax, uses semantic (selectlonal) information 
to filter parses, producing dramatic, 6-fold, decrease in number of parses. 
• Modules for processing intra-sentential temporal information and referring expressions (including 
definite and indefinite noun phrases, phrases with omitted determiners, and reference to events). 
• Integration of contextual information from fixed message fields with that from free-text fields: sys- 
tem now handles complex multl-paragraph message formats. 
Plan8 for nezt year: 
Demonstrate extensibUity of PUNDIT to new messages, showing correct information capture by 
filling in 70~ of target DB fields with a false positive rate of less than 5~. 
Integrate PUNDIT with knowledge representation and reasoning systems to provide inferencing 
capabilities for correct filling of DB fields. 
Treatment of intersentential temporal relations and implementation of a domain-independent 
discourse management component, for handling a variety of message formats. 
Demonstrate PUNDIT on Resource Management domain, in preparation for Spoken Language 
Understanding. 
265 
