USABLE, REAL-TIME, INTERACTIVE 
SPOKEN LANGUAGE SYSTEMS 
John Makhoul and Madeleine Bates 
makhoul@bbn.com, bates@bbn.com 
BBN Systems and Technologies 
70 Fawcett St. 
Cambridge, MA 02138 
1. PROJECT GOALS 
The primary objective of this project is to develop a robust, 
high-performance, domain-independent spoken language 
system. This system, termed HARC (Hear and Respond to 
Continuous speech), is composed of the BYBLOS speech 
recognition system and the DELPHI natural language 
understanding system. The goal is to develop systems that 
exhibit the following advances: high-accuracy speech 
understanding with a vocabulary of up to 10,000 words; a 
highly interactive user interface; adaptability to new users; 
easy portability to new applications; a system implementable 
in real-time on cost-effective COTS (commercial, off-the- 
shelf) hardware. 
2. RECENT RESULTS 
During the last year, we have: 
Designed and developed an initial implementation of, and 
began experimenting with, HUM, a Hidden Understanding 
Model. HUM is a new statistical model of semantics which 
will enable us to carry out studies in the alignment of 
corpora to semantic interpretations. The ultimate goal of 
HUM is to achieve understanding comparable or superior to 
other approaches, but using an automatic process to 
correlate words with meanings (instead of depending on 
hand-crafted semantic rules). 
Made a number of improvements sin DELPHI, including 
expanding the coverage of the Semantic Linker, 
improvements to quantification module, the 
implementation of sub-grammars, improvements in the 
handling of conjunctions, and the production of 
interpretations in the language FMRL (Flattened MRL) 
instead of older-style MRL. 
• Ported our ATIS development system to the HP735 machine 
to allow faster turn-around time for development. 
• Collected additional ATIS data for the general community, 
using our real-time system and the new, expanded database. 
• Delivered a version of our interactive ATIS system to NIST 
which they subsequently used for successful data collection. 
Improved the recognition accuracy for new words that were 
not in the speech training, and made the speech language 
model more robust to different phraseology. 
Made improvements to our real-time ATIS demo system and 
ATIS speech models, to improve usability and robustness. 
Improved the discourse component of DELPHI, making it 
better able to handle the sorts of task constraints which 
speakers impose in spoken interactions. 
Participated in discussions aimed at defining and 
developing semantic evaluation metrics. We made a 
proposal for the predicate-argument structure component of 
SemEval that is under consideration by the SemEval 
committee. 
Made plans to port HARC to a new domain, involving an 
unclassified database similar to one in use at Rome 
Laboratory. 
Participated in the December 1993 ARPA speech, NL, and 
SLS evaluations. 
Participated in various APRA committees. 
3. PLANS FOR THE COMING YEAR 
Continue the development of the Hidden Understanding 
Model. 
Development of aids to technology transfer. 
Improvements to search algorithms for speech recognition 
which allow fighter integration of knowledge sources (such 
as higher-order grammars) to improve real-time 
recognition accuracy without sacrificing speed. 
Port HARC to the Rome Laboratory domain. 
Participate in the planning and development of the ATIS4 
task and new evaluation methodology. 
447 
