TECHNOLOGY TRANSFER: 
PROBLEMS AND PROSPECTS 
Jesse w. Fussell 
Department of Defense 
9800 Savage Road 
Fort Meade, MD 20755 
1. INTRODUCTION 
For at least the last twenty years, DARPA 
and other government organizations have been 
sponsoring language processing research. The goal 
of this research is to develop processes to automate 
or semi-automate manu~ speech or text operations 
and to create new, capabilities to make it easier for 
humans to work with speech and textual data. 
During that time, a substantial amount of money 
has been provided to the nation's best organizations 
which have employed the most innovative and 
intelligent individuals. However, while there is a 
large body of objective evidence which 
demonstrates continuous progress toward the 
various technical goals of the programs, the fact 
remains there has been virtually no success to date 
in transferring any of the research effort into day- 
to-day operational use by any of the government 
sponsors. The purpose of this paper is to explore 
the reasons for the dearth of technology transfer in 
this technical area in the past, to forecast prospects 
for technology transfer in the future, and to suggest 
some ideas for stimulating the process. 
2. TECHNOLOGY TRANSFER 
PROBLEMS 
Many reasons can be cited for lack of tech- 
nology transfer successes in the past. The one 
which typically is of greatest interest to the research 
community is that the algorithms developed in the 
past were not considered adequate for operational 
use. In the speech recognition area, word 
recognition error rates were intolerably high unless 
the speech source was limited vocabulary, isolated 
word, wide bandwidth and free of background 
noise. In addition, applications were limited to the 
simplest uses because there was little 
understanding of how to apply language 
processing results to speech. In the text 
processing area, problems being addressed were 
often severely constrained in terms of vocabulary, 
grammar and application. But it was vitrblally 
impossible to find real-world problems that were 
constrained enough to allow application of the 
newly developed techniques. The substantial 
progress that has been made by the research 
community has changed this situation. The 
current state-of-the-art, while clearly inadequate 
for very general problems, has great prospects for 
certain moderately specialized applications. 
An additional problem that has impeded 
technology transfer has been the high cost and 
limited capability of the data and signal processing 
equipment needed to implement the algorithms 
researchers developed. In the speech processing 
area, researchers used to be restricted to general- 
purpose hardware which was tens or hundreds of 
time slower than real time, or of using special, 
fixed-point signal processors with very limited 
memory and dynamic range which were 
programmed in assembly language. Use of more 
modern computer languages such as LISP allowed 
faster design and implementation of experiments, 
but even with machines specially designed for 
that environment, processing was very slow. In 
addition, the graphics needed for a good user 
interface was limited in display speed and 
resolution. The net result of these limitations was 
that, except for the simplest algorithms, it was 
impractical to transfer research successes into 
operational use. However, while the thirst of 
researchers for ever greater computing, storage, 
networking and display capability still exists, it is 
greatly slackened. The explosive progress of 
computing over the past decade, led by the 
development of modern, high-performance, low- 
295 
cost, graphics workstations, now allows real-time 
or faster operation of fairly computational and 
memory intensive algorithms on floating point 
processors programmed in high-level languages. 
Another fundamental problem has been that 
the customer processing environment did not 
contain the infrastucture needed for the introduction 
of the techniques being developed. New digital 
speech processing systems could not be easily 
introduced into systems in which speech signals 
were being processed as analog signals and stored 
on analog tape. Text processing systems could not 
be made to operate in an environment in which 
textual information was still being handled on paper 
and data bases were on file cards. Again, thanks to 
major advances in data networks, low-cost 
terminals, optical character recognition, data base 
software, A/D and D/A conversion, and other 
technologies, many offices now contain the 
backbone of the system needed to make use of new 
text or speech processing techniques. In addition, 
the adoption of standards for programming 
languages, operating systems, windowing systems 
and network protocols often is resulting in the 
users obtaining computing systems which are 
compatible or which can easily be made compatible 
with the systems on which the processes are being 
developed. Thus, standardization will result in 
technology transfer being an easier job. 
The problems described so far may be classed 
as technical problems. And, as stated above, these 
technical problems are rapidly decreasing in 
importance. However there is one technical 
problem that still must be overcome before 
technology transfer can occur on a widespread 
basis. And unfortunately this remaining major 
technical problem has received little attention to 
date, probably because of the way in which the 
research problem has been structured. That is, the 
problems reported on by most researchers at this 
conference are primarily defined by two things: the 
selected (or created) and marked training and test 
data, or corpus, and a criterion or criteria for testing 
the researchers system against that corpus. This 
standardization of problem domain, goals and test 
criteria has been a powerful tool of the research 
managers. It has focussed researchers" attention 
onto relatively specific goals and objectives and, 
since the performance of the algorithms produced 
by different researchers can be directly compared, it 
has simultaneously created an environment of 
constructive competition. However this narrow 
focus has resulted in the neglect of one critical area 
of work: the study of the process of converting an 
algorithm or process from one domain, that is 
supported by a well defined and documented corpus, 
to a different, operational domain, which may 
have little marked training or testing data. The 
result is there has been little work to even 
formally define the process necessary to convert 
from one domain to another and from one 
objective function to another one, much less to 
automate or semi-automate that process. In other 
words, there has been insufficient emphasis on or 
efforts to achieve technology transfer. 
There are also a number of political, 
managerial or psychological problems which need 
to be addressed ff the transfer of language 
processing techniques into operational use is to be 
successful. All of these nontechnical problems are 
associated with the potential customers of the 
research. And, it is important to recognize that in 
general, the customers are not the sponsors of the 
research. The sponsors are usually other 
researchers or research managers who are 
supposed to represent the customers. The true 
customers are the people or organizations that will 
be the end user of any new product or capability 
that results from the research effort. 
The first problem involving the potential 
customer is that the customer usually not directly 
involved in the research efforts of the human 
language technology program. Or to put the shoe 
on the other foot, most researchers are not 
sufficiently familiar with the customers" needs and 
operating procedures to know whether a test 
corpus is truly representative of "real" data and 
whether the research goals will solve any "real" 
problem. 
Government customers also tend to be 
relatively conservative and sometimes even 
suspicious of new ideas and new technology. 
They frequently have developed a well understood 
routine and procedure for doing their jobs and are 
reluctant to change. In other words, there is 
often a lot of inertia which may only be overcome 
through the use of force. In this context, one 
such force which comes from demonstrated 
success. But since the voice and text processing 
technology is (and probably always will be) 
imperfect, "success", like "beauty" will be defined 
in the eye of the beholder. Since customers will 
be the ultimate judge of our products, we need to 
ensure that those products are demonstrated in 
296 
the best possible manner. Another force for change 
results from budget reductions. The conflicting 
pressures of less resources and desire to maintain 
capability tends to make some customers much 
more open to accepting new, imperfect technology. 
3. RECOMMENDATIONS 
First and foremost, the current program of 
research must be continued. Over the last few 
years, there has been substantial progress in 
understanding the fundamental problems of 
language processing and in developing ever better 
techniques for addressing those problems. This 
progress must continue if we are to have any hope 
of success against the less constrained problems. 
However, the goal of this research program 
should be broadened to explicitly include 
technology transfer. The task of adapting speech 
and text understanding processes from one domain 
to another needs to be specified as one of the 
program goals. Algorithms need to be developed 
which can be easily converted from one task to 
another without requiring years of additional work 
by highly trained scientists. Techniques need to be 
developed which will allow supervised adaptation 
to new situations. 
One implication of broadening the program to 
extend across domains is the need to increase the 
dimensionality of the research corpora so they also 
extend across domains. Instead of having data for 
a single situation which is segmented into portions 
for training, test and validation, this single set 
should be considered as the training portion for one 
domain, with another similar set for a different, but 
related domain which could be used to test how 
well the algorithm performs on the new domain, 
and possibly a third set of data from yet a third 
domain to validate the results of the domain 
transfer test. 
There also needs to be additional efforts by 
researchers and research managers to find potential 
customers in the government, to educate them on the 
goals and results of the program and to solicit their 
inputs into those goals. In addition, researchers 
need to spend more time working with customers to 
better understand the manual processes currently 
being used so they can better understand what is 
needed in order to produce a "successful" language 
processing capability. 
Finally, the overall goal of the program 
should be broadened, and funding provided to 
produce pilot or prototype systems which have 
been designed to be moved into operational 
situations and used for extended periods of time. 
Experimental operational prototypes are an 
absolutely necessary step in any long-term 
research effort. The trick is to determine the time 
for that step. The time is now. 
297 
