Modeling interpretation as
intention
Language is notoriously ambiguous, vague and
context-dependent. As a result, systems in computational
linguistics and models in cognitive science inevitably call for
representations which make meaning sufficiently explicit for
the purposes of language use. We usually refer to these
representations as interpretations, in an informal way. But
my current research suggest that surprising, deep and useful ideas
can come from appealing to first principles in rethinking and
formalizing what interpretations are.
The starting point is the philosopher H. P. Grice's work on
speaker meaning (Philosophical Review, 1957). He analyzes
the circumstances in which people attribute meaning to one
anothers' utterances and concludes that an interpretation
represents the intention that a speaker manifests in using an
utterance and addressees recognize from attending to the utterance
in context.
Now, Grice's attempts at formalizing the content of these
intentions (in the William James Lectures for example)
appeal to complicated and poorly-motivated assumptions about the
dynamics of human deliberation and action (as do similar
formalizations by Searle); they make sense only in the context of
Grice's program (quixotic on a Chomskyan understanding of
linguistic competence) to reduce knowledge of linguistic meaning to
the world knowledge we use generally to reason about others'
actions and mental states. Nevertheless, well-known computational
formalizations of interpretation as intentions by Perrault, Cohen,
Allen and others follow Grice's analysis. The result: difficult
formalism, implausible inferences and an all-but-unimplementable
theory.
On my view, the problem is not the analysis of interpretation
as intention; it remains the best analysis we have. Instead, the
difficulty comes from the many auxiliary assumptions that are
required to make sense of the analysis. But it has now been twenty
years since computer scientists first tried to formalize the
analysis, and research in the cognitive science of language use
offers a range of new insights to draw on:
-
Plans and intentions are resources for deliberation and
collaboration (Bratman, Pollack, Agre). To fill this role,
intentions should distill the real causal dynamics of the world
into its abstract, even heuristic, essentials.
-
Linguistic structures can be described abstractly but
precisely as combinations of independently-meaningful
elements assembled by simple primitive operations (Joshi's
adjunction, Steedman's combinators, Chomsky's merge). With these
descriptions, search through linguistic structures sets up a space
of natural choices that can be linked directly to interpretations.
-
Linguistic meanings can be described abstractly but precisely
as specifications for context change (Kamp, Heim, and many
others). These specifications cast utterances as actions by which
a speaker makes a relevant contribution to conversation by drawing
explicitly and intentionally on information given by the context.
-
Language use is a complex collaborative interaction
(Clark, Lewis, Brennan). This interaction doesn't just involve
signalling and recognizing intentions; it also involves processes
such as acknowledgment, grounding, accommodation, clarification,
and correction. Interpretations must support collaborative
language use, but they can also rely on it and abstract away from
it.
In my research, I hypothesize that these ideas can be
harmonized, so as to offer a streamlined analysis of communicative
intention that defers to, exploits and ties together research in
philosophy, linguistics and psychology. Look for the analysis to
make it possible to describe phenomena such as entrainment,
accommodation, and vagueness, which fall at the intersection of
these fields, with a new rigor that complements rather than
obscures our intuitions.
At the same time, these ideas open up ways to use the best
ideas from computer science to implement pratical dialogue systems
with sophisticated competence in language use.
-
Representations of interpretations can draw on
computational logic - starting with the typed data elements and
bound variables of the Curry-Howard isomorphism between
propositions and types, and between proofs and programs - to
describe inferences about coordinated deliberation, choice and
action.
-
Characterizations of processes in conversation can
appeal to probabilistic and decision-theoretic reasoning about
interpretations, and can thereby reconcile high-level
specifications of linguistic behavior with empirical facts about
people's language use.
New work and work in
progress
-
Representing communicative
intentions in collaborative conversational agents.
Matthew Stone.
AAAI Fall Symposium on Intent Inference for Collaborative
Tasks.
A formal investigation of how intention representations
might support conversational process, with the example of entrainment.
-
Communicative intentions and
conversational processes in human-human and human-computer
dialogue.
Matthew Stone.
To appear in John Trueswell and Michael Tanenhaus, eds,
World Situated Language Use, MIT 2002.
An informal investigation of how intention representations
might support conversational process, with an example of
disambiguation of referring expressions based on domain plan
recognition.
-
Microplanning with communicative
intentions: The SPUD system.
Matthew Stone, Christine Doran, Bonnie Webber, Tonia Bleam and
Martha Palmer.
RuCCS TR 65, under review.
Natural language generation described as a process of
constructing and manifesting a communicative intention - and
implemented effectively using modern models of linguistic structure
and meaning.
-
Indefinite information in modal logic
programming.
Matthew Stone.
Revision of RuCCS TR 56, under review.
An investigation into automated reasoning for planning,
reasoning in context, and language use. Draws on First-order multi-modal deduction. (Matthew
Stone, RuCCS TR 55).
-
Foundations of language interaction.
Matthew Stone.
Outline for a course on formal and computational pragmatics
(including sample programs and meeting handouts) run in Summer,
2001.