Linguistic Knowledge Sources for 
Spoken Language Understanding 
Deborah A. Dahl 
Unisys Center for Advanced Information Technology 
Paoli, Pennsylvania 19301 
PROJECT GOALS 
The objective of the Unisys Spoken Language Systems effort 
is to develop and demonstrate technology for the understand- 
ing of goal-directed spontaneous speech. The Unisys spoken 
language architecture couples speech recognition systems with 
the Unisys discourse understanding system, PUNDIT. PUNDIT 
is a broad-coverage language understanding system used in a 
variety of message understanding applications and has been 
extended to handle spoken input. Its power comes from the 
integration of syntax, semantics, and pragmatics processing, 
the ability to port rapidly to new task donudns, and from an 
open, modula,r a~chitecture. PUNDIT is unique in its ability to 
handle connected discourse; it includes a reference resolution 
module that tracks discom~se entities and distinguishes refer- 
ences to previously mentioned entities from the introduction 
of new entities. The PUNDIT front-end supports tunt-taking 
dialogue and permits the system to include both questions and 
answers in hnilding an integrated discourse context, required 
for the handling of interactive communication. PUNDIT has 
been interfaced to several speech recognition systems (MIT 
SUMMIT and ITT VRS-1280) to perform applications in 
dixection-finding assistance, all travel planning and air traffic 
control. 
RECENT RESULTS 
We have begun to explore a new approach to com- 
paring speech recognition systems and for investigating 
speech/language interactions in spoken language systems. 
This approach involves coupling a single natural language sys- 
tem to multiple speech recognition systems. At this point we 
have worked with output f~om MIT, BBN and Lincoln Labs, 
and are about to start working with output from Dragon. We 
believe this will be extremely useful for giving us a handle on 
the tradeoffs between recognizer improvements and improve- 
ments in spoken language understanding. 
We have also recently made improvements in semantic pro- 
cessing which allow semantics to handle, in a rule-governed 
way, a variety of syntactic constructions which exhibit non- 
transparent a~gument structure. These improvements led to 
an improvement of oux benclunark ATIS score to 48.3~ fxom 
36%. 
Our pragmatics processing has been enhanced to introduce 
information presented in displays into the dialog and to sup- 
port subsequent references by the user to items in the display. 
We have also developed a new approach to training the 
semantic component of the system using hand-generated rules 
developed for 900 sentences of the ATIS corpus to provide 
a basis for hypothesising the argument structure of unknown 
verbs. 
In collaboration with the University of Pennsylvania, we 
have recently implemented a probabilistic chart parser with 
an unknown word model and have partially integrated it into 
the system. This parser differs from previous attempts at 
stochastic parsers in that it uses a richer form of conditional 
probabilities based on context to predict likelihood. 
PLANS 
During the year to come we plan to: 
s Evaluate speech/language integration with multiple 
speech recognisers 
s Use Lincoln Labs stack decoder to provide tight 
speech/language coupling 
• Begin development of a new performance task - informa- 
tion retrieval interface 
• Extend trainable semantics 
• Improve dialogue processing for full dialogues 
427 
