Natural Language, Knowledge Representation and Discourse 
James F Allen and Lenhart K Schubert 
Department of Computer Science, 
University of Rochester, 
Rochester, NY 14627 
PROJECT GOALS 
The principal objective of this project is to develop a 
system for representing and reasoning about the 
discourse context in extended man-machine dialogs. 
Current focus areas include the development of a general 
theory of multi-agent planning to account for the 
structure of natural-language dialog, the development of 
a general knowledge representation for capturing a wide 
range of natural language semantics, and the 
development of a general, error-tolerant parser and 
semantic interpreter for English that can be guided by 
discourse information. Specifically, we are developing a 
model of discourse plans that includes actions such as 
introducing a new topic, as well as the actions of 
clarifying, correcting or acknowledging parts of the 
previous dialog. We are exploring how far the planning 
approach can be extended, and how the "traditional" 
language components, i.e. parsing, semantic 
interpretation and discourse processing, relate to the 
planning component. 
RECENT RESULTS 
The major initiative this year is the development of a 
prototype NL dialog system, TRAINS, operating in a 
simple, but realistic task domain requiring considerable 
man-machine interaction. This domain is one of 
scheduling transportation actions in a complex 
(simulated) world where only partial knowledge of the 
world state can ever be obtained. The system's 
responsibility is assist the human in constructing, 
monitoring and debugging the transportation plans. The 
first demo of this system, operating in an initial small 
subset of the domain was produced in September 1990. 
We are continuing work on extending the prototype 
system. We are completing a GPSG-style parser, which 
will show human-like preference-seeking behavior and 
error tolerance. Episodic logic, as described in last SNLP 
proceedings, is well developed at this point, and 
addressing such issues as causal connections between 
events, propositional attitudes of agents (goals, beliefs, 
etc.), and "defeasible" generalizations. The process of 
deriving a meaning representation from input text, 
which previously involved direct generation of episodic 
variables at the level of initial logical form, has been 
reformulated so that episodic variables are now 
introduced into the preliminary logical form in a top- 
down, context-dependent manner, using context 
structures called tense trees. 
TRAINS currently has a simple domain-plan reasoning 
system and a dialog model. We are now making a 
substantial effort to better characterize the capabilities 
needed of the discourse components, as reported in the 
paper in these proceedings. 
In summary, the major recent results are: 
• the completion and demonstration of a prototype 
dialog system, including parsing, semantic and 
contextual interpretation, domain plan reasoning and a 
discourse model; 
• an implemented detailed theory of tense and aspect, 
with emphasis on context-dependent, compositional 
semantics and on ease of computational realization. 
which addresses well-known problems in the systematic 
determination of "reference times" for tense-aspect 
constructs; and 
• the specification and construction of a dialog database, 
including the data collection, the development of a 
taxonomy or discourse acts, and initial prosodic 
labelling. 
PLANS FOR THE COMING YEAR 
Our plans for the coming year include the following: 
• continue to collect data and construct the dialog data- 
base, refining the taxonomy of discourse acts as 
necessary to maintain complete coverage of the corpus; 
• extend the discourse model to handle additional acts 
from the taxonomy, concentrating on the acts relating to 
maintaining the shared discourse state (e.g. 
acknowledgements, corrections, clarifications, etc); 
• develop a morphological model including a 
compositional semantics based on morpheme meaning 
in a way that addresses the alleged "bracketing 
paradoxes"; 
• implement a model of syntactic and lexical 
disarnbiguation based on information derived from the 
habituation of rules, explanation-based interpretation, 
and semantic priming; 
• perform some statistical studies on the dialog-database 
to examine the relationship between discourse acts and 
prosodic cues. 
431 
