SESSION 5: DISCOURSE 
Jerry R. Hobbs, Chair 
Artificial Intelligence Center 
SRI International 
Menlo Park, California 94025 
The fundamental problem in discourse is "What struc- 
ture is there in discourse above the level of the sentence?" 
This question can be asked in terms of text versus dia- 
logue. Is there a kind of structure that is exhibited in 
text but not in dialogue, or in dialogue but not in text? 
What are the appropriate structural descriptions of text 
and of dialogue? 
The question can also be asked from the perspective 
of recognition and from the perspective of generation. 
There is the structure that the speaker puts there and 
the structure the hearer discovers there. Are these the 
same? Does one have primacy over the other? For ex- 
ample, does the speaker impose the structure so that it 
is the job of the hearer to discover it, no matter what 
it is, or is it necessary for the speaker to design his dis- 
course in a way that makes it as easy as possible for the 
hearer to interpret? 
The four papers in this session discuss various aspects of 
this family of questions. 
The first paper, by Liddy and her colleagues at Syra- 
cuse University, is concerned with text rather than dia- 
logue, and recognition rather than generation. Structure 
in text can be studied from a general point of view or 
from a genre-specific point of view. This paper examines 
the structure specific to the genre of newspaper articles. 
The structure found can be in the form of a hierarchy, a 
tree-like structure, much like the syntactic structure of 
sentences, or it can consist of a division of the text into 
segments of various kinds, performing various functions, 
such as the Lead, the Main Body, and so on. Liddy and 
her colleagues argue that the latter kind of structure is 
more appropriate for newspaper articles. 
Among the questions addressed in this paper are 
• What are the structural elements or possibilities? 
• How can they be recognized, computationally? 
• Once recognized, how can they be used? 
The remaining papers are concerned with dialogue. The 
key idea in investigations of discourse from an artificial 
intelligence perspective is that the structure of a dialogue 
is, or is at least derived from, the structure of the par- 
ticipants' plans. "Plan" here is meant in the AI sense of 
a hierarchical structure of causal relations, decomposing 
goals into subgoals, and these subgoals into further sub- 
goals, and so on. An utterance in a dialogue is an action 
in a larger plan to achieve some goal. Generation is a 
matter of finding the right such actions. Recognition is 
a matter of discovering the role that action plays in the 
overall plan. 
Moore's paper examines a problem that arises in gener- 
ation. We don't like to be told too much. If we already 
know something, we don't like to hear it again. If the 
speaker does not take this into account, the result is 
what my children used to refer to as "talking to me like 
I'm a retard." It is also an inefficient way of conveying a 
plan or some other structured body of information. The 
problem that Moore addressees is "How should we use 
knowledge gained from the previous discourse to convey 
a plan as efficiently as possible?" 
The paper by Ferguson and Allen take the perspective 
of recognition. Part of understanding an utterance is 
discovering the role it plays in the speaker's larger plan. 
But the speaker's plan, or intentions, are inaccessible 
to us. All we have to go on is what he or she said. 
From that we have to hypothesize a plan in which the 
observable utterances would make the most sense. This 
is the problem that Ferguson and Allen address: "How 
can we recognize the speaker's plan, given only sparse 
information about it?" 
In stretches of dialogue larger than a single utterance, 
neither participant is exclusively a speaker or exclusively 
a hearer. They are each executing their own plan, and 
their plans change a8 they are impacted by the action~ 
of the other participant. In very focused, collaborative, 
problem-solving dialogues, they are working together to 
come up with a single plan for solving the problem. How 
can a single shared plan arise out of the interaction? This 
is the problem addressed in the paper by Biermann and 
157 
his colleagues at Duke University. Each participant has 
pieces of a plan; how can they be combined into a single, 
agreed--upon plan? The particular situation described is 
a very common one. One participant, in this case, the tu- 
toring system, has control over general knowledge about 
the domain. The other participant, here, the user, has 
the knowledge of the specific problematic situation. How 
can these be combined into a solution of the problem? 
158 
