SESSION 2: INVITED OVERVIEWS 
Madeleine Bates 
BBN Systems & Technologies 
70 Fawcett Street 
Cambridge, MA 02138 
By adopting the name Human Language Technology for one 
of its flagship programs, ARPA intended to included every- 
thing that is involved in understanding and/or generating 
natural human language. In the Information Age, more and 
more resources are being devoted to collecting, retrieving, 
and using information (text and speech) in digital form. The 
time is ripe to multiply the power of human-machine prob- 
lem solving systems by automating ways to cope with the 
information explosion we are all experiencing. 
The scope of this workshop, expanded from a previous con- 
centration on spoken language systems to include all spoken 
and written language work at ARPA, meant that many of 
the attendees were coming into contact for the first time 
with researchers from related but different disciplines. Sev- 
eral invited overview presentations helped to set the context 
for what was to come later in the workshop. 
Beth Sundheim of the Naval Command, Control & Ocean 
Surveillance Center described the series of four MUCs (Mes- 
sage Understanding Conferences) that have taken place since 
1987. These conferences, each involving a highly structured 
evaluation of message processing systems, have served to 
quantify the state of the art in this field, and to provide a 
forum for studying and modifying the evaluation methodol- 
ogy. Participation in the MUCs has risen from 6 sites to 17, 
demonstrating increased interest in message processing as a 
task, and increased willingness to evaluate systems using a 
common task for which training data is available. 
Donna Harman of the National Institute of Science and Tech- 
nology (N/ST) presented the background leading up to the 
first Text REtrieval Conference (TREC-1), and the results 
of that conference. The TREC participants were eager to 
test a variety of retrieval techniques in a common, challeng- 
ing evaluation. Approaches ranging from pattern matching 
to term weighting to natural language processing competed 
in what was probably the best modern information retrieval 
test, and certainly the largest. Papers by several of the TREC 
participants appear in tiffs proceedings (particularly in Ses- 
sion 12, Information Retrieval). 
Thomas Crystal from ARPA presented an overview of the 
TIPSTER program, including both the message detection 
and data extraction tasks. By "detection" is meant two vari- 
ants of information retrieval: retrospective retrieval and rout-. 
ing. The detection problem is to process queries in the form 
of user-need statements and finding messages that meet the 
needs (e.g, are on a particular topic) from among a huge 
set of messages, some of which are similar but not on the 
desired topic, and some of which are completely irrelevant. 
By "extraction" is meant extracting specific types of infor- 
mation from messages that are likely to be relevant to the 
particular topic. Extraction systems typically apply text un- 
derstanding techniques to process the text, and then produce 
database fill from the results of that understanding. 
Part of the challenge of the TIPSTER program has been to 
provide a harder problem than has been worked on before, 
one that can be solved only by applying very advanced tech- 
nology; to this end, TIPSTER detection and extraction work 
has been pursued in both English and Japanese. The current 
TIPSTER contractors' work is represented throughout this 
proceedings. Because Tom Crystal will be leaving ARPA 
shortly, questions about this program should be directed to 
George Doddington. 
George Doddington of ARPA provided an overview of the 
Spoken Language Systems (SLS) program, which is also 
well-represented in this proceedings, particularly in Sessions 
1 and 3. The SLS progrmn is concerned with all aspects of 
human-machine communication by voice, and has been us- 
ing ATIS (Air Travel Information System) as a common do- 
main for development and evalulation. The paper by David 
Pallett gives a summary of the most recent benchmark eval- 
uations in this program. 
Inquires about the SLS program or any other aspect of the 
ARPA HLT program should be addressed to: 
Dr. George Doddington 
ARPA / SISTO Room 744 
3701 N. Fairfax Drive 
Arlington, VA 22303-1714 
gdoddington@ darpa.mil 
703-696-2259 
55 
