PROGRESS REPORT 
Active Knowledge Structures in Natural Language Understanding 
Computing Research Laboratory 
New Mexico State University 
Contact: Yorick Wilks 
yorick@nmsu.edu 
(505) 646-1835 
Work on the project has concentrated in the last six months on (a) reimplemenfing our 
two semantics-based message parsers and (b) integrating them with the KR formalism 
we use: a form of conceptual graphs (CG) embedded in the MGR (Model Generative 
Reasoning) framework. 
The PREMO (Preference Machine Organization) parser has been reimplemented, 
using a technique to adjust the weighting of word and structure preferences such that it 
can be trained to systematicities in the special (--"ill-formed") syntax of a given mes- 
sage type. 
In the case of the other semantics-based parser PM, Jerry Ball took two of the 
texts from the Navy message database and added the vocabulary from those messages 
to the parser's lexicon. After a small amount of modification, the parser was able to 
parse about 80% of the sentences in those two messages into reasonable representations. 
With some additional work this percentage can be improved. Given the lexically driven 
nature of the parser, extending the system to cover a larger subset of the Navy message 
database rests primarily on expansion of the lexicon. For demonstration purposes, an X 
Windows Interface to the parser was developed. 
We are also integrating the ViewGen belief manipulation system with the concep- 
tual graph + MGR knowledge representation, so as to provide a system that can both 
guide the message parsers and represent the results of message extraction. The immedi- 
ate goal is re-implementing ViewGen in conceptual graphs. This has led to the follow- 
ing recent developments: 
(1) Investigating the benefits of using CG to express ViewGen. 
(2) Representing the environments, lambda formulas, and propositions of ViewGen in 
CG. 
(3) Using examples from Navy messages to specify belief ascription in CG. 
421 
