Overview of the DARPA Speech and Natural Language Workshop 
Lynette Hirshman 
Unisys Defense Systems 
The aim of the first joint Speech and Natural Language Workshop was to bring together 
these two research communities, to interchange technical information, to reflect on past 
successes and to define future directions for Spoken Language research. The overall 
organization of the technical program was designed 1) to establish some common 
reference points by assessing the current state of the art in both Speech Recognition and 
Natural Language Processing; 2) to cross-educate researchers in the discipline of lesser 
familiarity; 3) to highlight areas of common interest, namely prosodics, spoken language 
systems, and development of shared resources; and 4) to present current research results 
in both fields. 
To support these aims, the program was divided into roughly four segments. 
Part 1 (Day 1) focused on the establishment of common ground between the natural 
language and speech groups and discussion of shared resources and performance 
evaluation. 
Part 2 focused specifically on progress in the area of Spoken Language Systems (morning 
of Day 2). 
Part 3 (afternoon of Day 2 and morning of Day 3) focused on research results in the 
separate areas of speech and natural language. 
Part 4 (rest of Day 3) focused on topics of joint interest: plans for selecting appropriate 
spoken language applications, stories of technology transfer, and finally, future plans. 
In addition, two evening sessions were held. The first was for demonstrations of running 
systems (including video tapes). The second was an ad hoc meeting called by George 
Doddington (Texas Instruments) to discuss the collection of a new speech/natural 
language database to drive research and evaluation. This ad hoc meeting sparked a heated 
but profitable exchange between the groups focused on spoken language systems and 
those focused on more conventional speech recognition tasks. The disagreement centered 
on the need to evaluate the contribution of natural language UNDERSTANDING, as 
opposed to the conventional speech recognition metric of CORRECT TRANSCRIPTION. 
The spoken language researchers pushed for definition of a task to be performed via 
spoken language, where metrics could be developed to measure task performance 
independent of the notion of verbatim transcription. By the following day, there seemed 
to be agreement that an application such as travel planning would provide an appropriate 
task for purposes of research and evaluation of spoken language systems. The domain of 
air traffic control was also discussed at some length. 
The workshop was successful in bringing these two groups of researchers together for 
technical discussions. The reaction to the tutorial sessions was generally very positive, 
with some frustration expressed at the time constraints, which precluded exploring any 
topic in depth. A number of researchers also expressed frustration at holding concurrent 
Speech and Natural Language sessions, because this prevented them from attending both 
sessions. This reaction was taken as a sign of serious interest in bridging the gap between 
these disciplines. Overall, the workshop represented a major step towards the realization 
of Spoken Language Systems on the part of the DARPA-funded speech and natural 
language research communities. 
2 
