University of California, Berkeley 
Principal Investigator: Robert Wilensky 
The objective of our intelligent language systems work is to conduct basic research in the 
areas of language understanding, common-sense inference, and knowledge representation. 
For specifically, we are interested in: 
(1) Producing better user interfaces. 
(2) Enabling the automatic processing of natural language text by computer. (3) Building 
autonomous planning agents that can operate in complex environments. 
Much of our work has focused on the design and implementation of a Unix Consultant (UC) 
program. This system carries out a dialog in English to answer user's queries about the UNIX 
operating system. UC comprises a natural language parsing and generation system, a goal 
analyzer that hypothesizes user's goals in different contexts, a user model component that 
allows UC to tailor answers to the user's level of expertise, a conversational planner that 
allows flexible reaction to user requests, a response formation mechanism that enables from 
of response to be concise and appropriate, and (iv) a knowledge-intensive planner that 
allows UC produce complex plans and warnings of potential plan failures 
We have also been interested in developing a general theory of inference for text 
understanding. Our previous work involved the development of a theory of inference for 
text understanding which identifies six classes of inference, and uses a highly-parallel 
marker-passing mechanism to identify potential inferences. We designed and implemented 
a system which uses this theory to correctly make inferences which were previously possible 
only with a much larger number of highly-specialized rules. 
Recent Accomplishments 
- Developed the details of a theory of the incremental acquisition of new metaphoric word 
senses. Two distinct kinds of learning were explored: the first is based on a theory of 
metaphorical similarity, the second on a theory of the hierarchical preservation of structure 
from the source to the target domains of conventional metaphors. Implemented sixteen 
distinct metaphor types with several hundred senses for 22 of the most common verbs. 
- Started implementation of a knowledge acquisition system that can augment UC's 
knowledge base by reading the on-line UNIX man pages. 
- Developed extensions to the KODIAK representation language and incorporated learning 
techniques into UC Teacher, the knowledge acquisition component of our UNIX Consultant 
(UC) system. 
- Addressed the principle theoretical issues necessary to advance our theory of inference, in 
particular a notion of sufficient explanation and focus. 
- Implemented a new grammar of a portion of English, which emphasizes the relation 
between grammar and meaning. The grammar extends commonly-used unification 
techniques for parsing and representing grammatical rules. 
- Analyzed how various properties of "operationality" theories may effect the efficiency of 
Explanation Based Learning algorithms. Focused on problems relating to the behaviors of 
operationality boundary conditions. Developed and implementated polynomial-time 
algorithm for EBL, where previous algorithms were exponential. 
263 
Plans 
Over the next year, we anticipate working on the following problems: 
- Augment the capabilities of our knowledge representation system to better handle problems 
of change over time, quantified statements, and hypothetical statements. 
- Build prototype version of a new language understanding system that combines inference, 
interpretation, and parsing. 
- Introduce a notion of level of importance into our inference algorithm, enabling the 
understander to focus attention. 
- Begin reimplementation UC to better integrate components and take advantage of new 
technology. 
264 
