A Real-Time Spoken-Language System 
Interactive Problem-Solving 
Patti Price and Robert C. Moore 
SRI International 
for 
Objective: 
SRI is developing a system to improve complex problem- 
solving through tile use of interactive spoken language in 
conjunction with other media. This requires real-time 
performance, large vocabulary, high semantic accuracy 
and habitability, as well as robustness to expected and 
unexpected variability. SRI's spoken-language system is 
being developed in the air travel planning domain along 
two overlapping research and development lines, one fo- 
cussed on an SLS kernel for database query, and tile 
other on the full interactive system. 
Recent Accomplishments 
• Improved natural language understanding in the air 
travel planning domain (about 90% of a test-set of 
319 sentences collected at T1 now parse, 64% re- 
ceive a semantic interpretation, over 25% retrieve 
the correct answer from the database. 
• Implemented corrective training to improve recogni- 
tion performance; on the standard training set this 
improves speaker-independent perplexity 60 perfor- 
mance from 6.7% error to 5.1% error, and for a 
larger training set (about 11,000 sentences), im- 
proves speaker-independent recognition from 5.3% 
error to 4.1% error. 
• Evaluated the effects on performance of larger train- 
ing sets: using about 11,000 training sentences com- 
pared to about 4000 training sentences improves 
from 6.7% error to 5.3% error. 
• Started initial experiments in consistency modeling: 
showed significant improvement for separate model- 
ing of male and female speech for both the standard 
and the large training sets (performance improves 
from 6.7e~ error to 5.7% error on the standard train- 
ing and from 5.3% error to 4.3% error on the large 
training set). Speaker- independent perplexity 60 
performance can be expected to improve still fur- 
ther when gender-specific modeling with corrective 
training on large data sets is implemented. 
• Started implementing procedures for porting to new 
vocabularies, including the ATIS vocabulary, in- 
cluding the automatic generation of baseforms, ap- 
plications of rules and the creation of word-models 
based on existing training data from other domains 
(currently about 9000 verified words are i, the dic- 
tionary and another 20,000 from tile Brown corpus 
most frequent words in process). 
Incorporated in the SRI Speech Understanding Sys- 
tem techniques (developed under all NSF contract) 
that use statistical information about tile speech sig- 
nal to improve recognition accuracy in noisy envi- 
ronments; started the real-time implementation of 
these algorithms. 
Started port of phonology software from LISP to 
C, and demonstrated performance improvement 
through the use of more detailed, statistically 
trained phonological models. 
With help from TI, implemented a functional equiv- 
alent of the TI data-collection mechanism; collected 
and transcribed data from 10 subjects using the 
TI protocols. Designed experiments for variations 
on this mechanism and started data collection and 
analysis. 
Plans 
. Continue ATIS development. 
* Expand effort in consistency modeling and training 
on larger data sets. 
* Continue development of software for easy porting 
to new vocabularies. 
• Begin effort to detect out-of-vocabulary items and 
design and evaluate an appropriate interface. 
• Collect and analyze more human-machine data un- 
der various conditions in order to guide the design 
and evaluation of the SRI SLS. 
424 
