Intentions iI1 Bilingual Dialogue Processing 
Susann LuperFoy 
The MITRE Corpora,tion 
susann@st arbase, mitre, org 
1 am currently leading a small research project on discourse-level processing for a Japanese-English 
interpreting telephone systern. The interpreting telephone (IT) would allow monolingual speakers 
of different languages to carry on a spoken conversation. The dialogue management task faced by 
the IT's discourse component defines mfique constraints on t.he overall IT software system. 
The dialogue manager must model five distinct dialogue interactions that arise during operation of 
the larger IT system. 
(1) the huma.u-hmnan (bilingual) dialogue that coitstitutes the primary data for tile system, 
(2),(3) two human-maclfine (monolingual) dialogues, each of which alternates input speech of one 
speaker with tile translated output of tile other speaker's input, and 
(4),(5) two human-nmchine dialogues that occur whenever the interpreting telephone system itself 
(or a component of it., such as the dialogue manager) engages either of the humans in a recta-dialogue 
to address any disruption for the flow of the primary (hunaan-human) dialogue. 
The role of intentions in the IT system varies with the role of the dialogue manager. In processing 
dialogues of Types 4 and 5, the IT is acting as an agent in a private dialogue with one of the users. 
It must discern the user's plan and (dynamically) construct and execute its own plan for repairing 
the dialogue error condition in collaboration with the user. The IT system can influence the user's 
behavior and intentions through its own behavior. 
In contrast, in dialogues of Types 2 and 3 the IT dialogue manager acts ms a third-party observer 
of l.he hmnan-hunmn bilingual dialogue. The dialogue manager's task is to recognize the evolving 
plans of the two users in order to analyze source-language utterances adequately for carrying out 
translation, and in order to make decisions during generation of target-language utterances tlmt 
reflect the source-language speaker's intentions. 
The two users may have shared experience that extends beyond the current conversation. They may, 
therefore, have intentions that are mutually accessible but not available to the dialogue manager. 
For instance, these two users may have spoken on the previous day and may now make context- 
dependent, reference to the extended context begun in an earlier dialogue, "Lets go back to the 
second suggestion we discussed yesterday." This requires a local definition of intentions at the level 
of l, lle immediate dialogue segment that the dialogue manager has access to. 
68 
Tile ultimate success of the IT system is measured by its ability to create the experience, for each 
of the two users, of a monolingual dialogue ill their own language. Intentions ill translation could 
be viewed as an automatic side-effect of a good translation, l propose instead that, for tile purposes 
of bilingual telephone dialogues, intentions are what get preserved under translation. Using two 
intention recognizers, one for Japanese and one for English, the dialogue manager will examine 
a set of alternative translations and filter those that appear not to realize the intentions of the 
source-language speaker. 
These five dialogue types can be represented using an adaptation of the three-tiered discourse reprc- 
sentation (defined iu my dissertation) that distinguishes three types of information accessed by the 
discourse interpreter during dialogue processing: linguistic, discourse, and belief information. The 
framework models the five IT dialogues in parallel, but minimizes redundancy and computational 
effort through the use of shared structures. A prototype dialogue manager module is currently undcr 
development in our group. The sample dialogues for the immediate system development involve a 
highly restricted domain (conl~rence registration) which defines a very small library of tasks and 
goals from which to choose. My primary interest in this workshop is in tim evidence for intentions 
for inachine translation and a classification of intention types for the various IT dialogues. 
