USABLE, REAL-TIME, INTERACTIVE SPOKEN LANGUAGE 
SYSTEMS 
John Makhoul and Madeleine Bates, Principal Investigators 
BBN Systems and Technologies 
70 Fawcett St. 
Cambridge, MA 02138 
PROJECT GOALS • 
The primary objective of this project is develop a ro- 
bust, high-performance, domain-independent spoken lan- 
guage system. The system, termed HARC (Hear And Re- 
spond to Continuous speech), is composed of the BYBLOS 
speech recognition and the DELPHI natural language under- 
standing system. The goal is to develop systems that exhibit 
the following advances: high-accuracy speech understanding • 
with vocabulary of up to 10,000 words; a highly interactive 
user interface capable of mixed-initiative dialogue and other 
types of system feedback; transparent adaptability to new 
users; easy portability to new applications; a system imple- 
mentable in real-time on cost-effective COTS (commercial, 
off-the-shelf) hardware. 
RECENT RESULTS 
• Developed the first complete real-time spoken language • 
system on a workstation, without the use of additional 
hardware. The system integrated our real-time speech 
recognition and NL systems through the N-best inter- 
face. The real-time system was demonstrated in the 
ATIS domain at the DARPA 1992 and 1993 workshops 
on speech and natural language processing. The system 
was developed on a Silicon Graphics Indigo worksta- 
tion and has been ported to other Unix-based worksta- 
tions. • 
• Built a new semantic intmgretation component that 
enables a cleaner separation between the domain- 
independent knowledge stored in a general grammar of 
English and the domain-dependent knowledge associ- 
ated with particular lexical items, thus enhancing porta- 
bility and domain-independence. The new component 
is based on the notion of translating from grammatical 
relations to semantic relations. 
• Developed a hybrid approach to semantic representa- 
tion that combines the expressive power of logic with 
the reasoning power of frames. This allows the sys- 
tem to detect contradictions (e.g., "flights to denver ... 
uhh to Boston") without being forced into the simple 
frame-slot approach that is inadequate for many types 
of natural language phenomena. 
Developed the Semantic Linker, a new fallback un- 
derstanding component using probabilities that reduces 
the understanding error rate by 30%. The Semantic 
Linker makes extensive use of the hybrid representa- 
tion and interpretation component above to take over 
when DELPHI's regular parser cannot understand an 
input utterance. 
Laid out an initial design and developed an initial repre- 
sentation for a new statistical model of understanding, 
the Hidden Undersatanding Model. The new model 
will enable us to carry out studies in the alignment of 
corpora to semantic interpretations, with the goal of 
automatically deriving semantic rules. 
Participated in the November 1992 speech, NL and SLS 
evaluations, and in the end-to-end log-file evaluation 
dry-run experiment. 
Chaired the organi7atiorl of the March 1993 DARPA 
Human Language Technology workshop. 
PLANS FOR THE COMING YEAR 
Our plans for the coming year center on extending the work 
above in several areas. In particular, we plan to: 
Integrate the Semantic Linker and the parser, to increase 
the capabilities of both. 
Improve our discourse understanding component, mak- 
ing it more domain-independent. 
Develop ways of extending our grammar for differ- 
eat domains. This will allow us to keep our domain- 
independent knowledge base, but to tailor it for domain- 
specific syntactic constructions. 
Continue the development of the Hidden Understanding 
Model and test it by deriving the semantic rules for a 
real domain. 
Collect additional ATIS data for the general commu- 
nity, using our real-time system and the new, expanded 
drllaha$~, 
387 
