Session 11: Natural Language II 
Deborah A. Dahl, Chair 
Unisys Defense Systems 
Center for Advanced Information Technology 
Paoli, PA 19301 
Three of the four papers in this session described work 
aimed at reducing the amount of labor-intensive "hand- 
crafting" of natural language systems. The fourth paper 
described the current need for machine translation and 
machine-assisted translation and advocated an interlingual 
approach involving a coordinated effort by several sites. 
It is extremely encouraging to see that new ideas are 
being explored for reducing the effort required to develop 
new applications of natural language understanding. These 
ideas will clearly be needed to enable natural language 
technology to be applied to real problems. The first three 
papers described three different approaches to this problem 
- using statistical automatic training techniques (BBN), 
developing "generic" text processing capabilities (GE), and 
the use of analogical reasoning to hypothesize new word 
senses (Berkeley). The work described in these papers was 
generally exploratory in nature. It will be exciting to see 
how these approaches continue to develop and become in- 
tegrated into full-scale systems. 
The first paper, by Ayuso, et. al. from BBN, described 
three experiments in statistical approaches - part of speech 
tagging, probabilistic parsing, and acquisition of lexical 
syntax. The most intriguing aspect of this work was the 
potential for synergy among the various tools described. 
For example, one suggestion was that the probabilisfic 
parsing could be used to control the ambiguity inherent in 
dealing with unknown words. 
The second paper, by Jacobs, et. al., from GE, described 
an approach to handling large amounts of unrestricted text 
which involves developing generic text processing 
capabilities. This paper reports on some of the tools which 
underlie this approach, including the development of a 
10,000 word lexicon and various text preprocessing tools. 
These tools are used to produce a text tagged with word 
senses. This result is interesting because it shows that it is 
possible to produce at least part of a semantic analysis 
(word sense tagging) for arbitrary text. 
Robert Wilensky's paper on lexical acquisition was the 
third paper in this session. This work uses lexical sub- 
regularities to extend the meaning of words used in new 
senses. This paper lists a number of lexical subregularities 
and outlines a preliminary procedure for lexical acquisition. 
This work is in its early stages and it should be very inter- 
esting to see how it develops. This capability should be 
quite useful for making natural language systems more in- 
dependent of their developers. 
The final paper in the session (Wilks, et. al.) was a 
proposal for a new effort in machine translation, based on 
an interlingua approach. It makes a strong case for the need 
for such work, since currently available machine translation 
systems are based on very old technology, although the 
need for translation assistance is so great that they are still 
being used. 
Several interesting points were made in the discussion 
of the Wilks paper. One criticism was made that the 
proposal offered no striking new ideas. The response to this 
was that the project would pull together existing skills. In 
addition, it should also be noted that there are many current 
ideas in areas such as parsing that have not yet been ex- 
ploited in machine translation systems. Other discussions 
concerned the relative merits of transfer vs. interlingua ap- 
proaches, and statistical vs. knowledge-based techniques. 
353 
