Session 8: Spoken Language Systems II 
Charles T. Hemphill, Chair 
Texas Instruments 
P. O. Box 655474, MS 238 
Dallas, TX 75265 
This session consisted of four talks from six papers 
representing the gamut of technology in spoken language 
systems. Perhaps surprisingly, each of the papers present 
techniques that are compatible with one another and SLS 
system builders can benefit from all of these. 
Mike Philips of MIT presented recent progress on the 
VOYAGER system. In particular, MIT has incorporated a 
"top-N" approach to increase the level of integration be- 
tween speech and natural language. They use Viterbi 
search to compute the upper bound estimate for an A* 
search. Readers should contrast this with papers from 
BBN and AT&T in the first SLS session. MIT has also 
developed an interesting way of creating a high- 
coverage/low-perplexity word-pair grammar based on sen- 
tences generated from their natural language grammar. 
Debbie Dahl of Unisys presented some observations on 
training and evaluation of a spoken language system. 
Using "top-N" Voyager data from MIT, they explored the 
relationship between training and coverage. They con- 
verged on 70% coverage after approximately 1000 sen- 
tences. This convergence depends on the domain and per- 
haps the data collection paradigm, as discussed in Unisys' 
paper in session five. Unisys has also tabulated empirical 
data for correct responses versus false alarms to help deter- 
mine a beneficial value of N for the "top-N" algorithm. 
Alex Rudnicky of CMU presented developments in 
spoken language interaction. While the work discussed 
above primarily addressed spoken language speed and ac- 
curacy, CMU has concentrated on the user interface and 
system integration issues. They have designed a spoken 
language system architecture that permits the rapid design 
of spoken language applications. With this architecture, 
they have studied computer-human interface design for five 
different applications. The modularity of the architecture 
facilitates research and development for each of the in- 
dividual components. 
Finally, Dave Stallard from BBN presented their recent 
developments in applying unification grammars to spoken 
language systems. It should be noted that most of the 
concepts presented can be found in the unification grammar 
literature, and BBN indicated this by saying that the work 
contains ideas "so old that they are new." In the first paper, 
BBN presents various situations that illustrate the advan- 
tage of using e-productions in a unification grammar. In 
the second paper, BBN presents various approaches for 
reducing the time and space requirements of a unification 
grammar system. These include: the compilation of rules 
into rule-groups that share mutually subsumable con- 
sfituents, a limited form of feature disjunction in unifica- 
tion, and prediction constrained by features. 
205 
