NATURAL LANGUAGE PLANNING DIALOGUE FOR INTELLIGENT 
APPLICATIONS 
James F. Allen and Len Schubert 
Department of Computer Science 
University of Rochester 
Rochester, NY 14627 
PROJECT GOALS 
The goal of this project is to develop the underlying 
technologies for spoken dialogue systems that serve as 
interfaces to complex, state-of-the-art reasoning systems. 
Most current speech and natural language projects are 
focusing on applications that involve very little 
intelligent reasoning, such as data-base query and form- 
filling. However, the great promise for speech and natural 
language interfaces is in providing useful interfaces to 
complex AI reasoning systems such as planning systems 
and expert systems. 
We are developing a domain-independent general model of 
semantic representation together with an integrated plan- 
based representation both for problem solving in the 
domain and for managing the dialogue itself. The 
semantic representation is episodic logic, a rich 
representation language, based on the notion of episodes, 
that can representation of the meaning of a wide range of 
linguistic constructs. The plan-based model provides the 
necessary infrastructure needed to integrate syntactic and 
semantic processing, discourse structure, and domain 
reasoning to create an effective dialogue system. 
RECENT RESULTS 
1) We developed a plan reasoning system that can handle 
a wide range of interpretation problems that arise in the 
TRAINS dialogs, including suggestions of courses of 
action, of objects to use in the plan, of goals, and of 
other constraints on the plan. In addition, it supports 
purpose clause qualifications on any of these suggestions. 
2) We completed a new version of the discourse reasoner 
and tested it on some sample dialogs. The new system 
maintains what each agent believes, what each agent has 
suggested about the plan, and what parts of the plan have 
been agreed on so far. It can filter possible speech act 
interpretations using knowledge of the agents' beliefs and 
knowledge of the current plan. 
3) We developed and implemented a scope disambiguation 
algorithm. Operators are assigned their scope depending 
on their syntactic, semantic and pragmatic properties in 
conjunction with reference and tense interpretation. This 
work includes a model of contextual reasoning 
appropriate for interpreting definite descriptions. 
4) Chung Hee Hwang completed her doctoral dissertation: 
"A Logical Approach to Narrative Understanding", which 
fully specifies a general semantic representation for 
language, episodic logic. This work is especially notable 
for the breadth of its semantic coverage, including detailed 
analyses of tense constructs and of adverbial modification. 
5) We completed the design of a system that, when 
presented with a word it has never seen before, creates a 
new lexical entry with meaning postulates that represent 
a partial semantic definition of the word by considering 
specific word formation processes (e.g., affixation, 
argument structure altemations, compounding etc.). 
6) We set up a dialogue lab this year so that dialogues 
can be collected in a more controlled setting. With the 
dialogs collected so far, we are marking intonational 
features, annotating repairs, and producing an aligned 
transcription using specially developed tools built on top 
of the WAVES system. We are developing standards for 
annotating higher level discourse phenomena, such as 
segmentation, co-reference and speech act analysis. 
PLANS FOR THE COMING 
YEAR 
1) Complete a new domain plan reasoning system based 
on plan graphs and test it extensively on data from a wide 
range of TRAINS dialogues. 
2) Develop new versions of each module in the TRAINS 
system to support incremental interpretation, so the 
system be based on intonational phrases rather than 
complete sentences. 
3) Implement the lexical reasoning system described 
above to derive partial meanings of new words. 
4) Complete annotation schemes for segmentation, co- 
reference, speech acts and discourse acts, and test by 
annotating several hours of dialogue. 
5) Complete new versions of our work on scope 
disambiguation and reference, and on the dialogue 
reasoner. These will be described in two Ph.D. 
dissertations by David Traum and Massimo Poesio. 
420 
