 |
13th International Conference
on Functional Grammar
|
| Using Functional Discourse Grammar in an
Agent-based Dialogue System |
Nieske Vergunst, Reinier
Lamers & Frank Dignum, Universiteit Utrecht, Utrecht, The
Netherlands
We are attempting to build a spoken
dialogue system that can assist users in selecting recipes and
in cooking. For automatically interpreting the input of our
dialogue system we are planning to use Functional Discourse
Grammar (FDG, Hengeveld & Mackenzie 2008), because it integrates into the grammar all
the discourse theory that is necessary to explain the form of
linguistic utterances. We will focus on a computational
approach to parsing spontaneously spoken language to a
structure in FDG. This involves a way to parse utterances to a
morphosyntactic representation, an algorithm to compute the
structure at the representational level given the structure at
the morphosyntactic level, and an algorithm to compute the
structure at the interpersonal level given the structures at
the representational and morphosyntactic levels.
However, building such a parser is not
an easy task. FDG is a relatively new grammar theory and no
parsers that generate FDG structures have been made yet. FDG
is usually described “quasi-productively”, describing how
a communicative intention is transformed into an utterance via
four levels of representation. We want to do the reverse:
automatically transform a transcribed utterance to a
communicative intention, going through the four levels of
representation in reverse order. A representation that
dissects the text into morphosyntactic constituents can
probably be achieved by any off-the-shelf parsing technology,
with small modifications to generate FDG structures instead of
another hierarchical representation of the syntax. The
difficulties will be in deriving the representational and
interpersonal level from the morphosyntactic level and the
input utterance.
The translation from a communicative
intention in the Conceptual Component to a representation on
the interpersonal and representational levels is done through
the process of Formulation. For interpreting input and
inferring the speaker's intentions from it, we need to focus
on reversing this process: eliciting the communicative
intention from the speaker's utterance. This is difficult, as
there is no one-to-one mapping from a speech act to a
communicative intention and back. However, there are various
assumptions that can help us to elicit the speaker's intention.
In a cooperative dialogue situation, both dialogue partners
usually act with a considerable amount of predictability. In
order to use these ‘hints’, we need to keep a user
model, a formalization of the Conceptual Component of the
user. We also need to formalize the system's own Conceptual
Component. For both of these, we are planning to use the
Beliefs-Desires-Intentions (BDI) model (Cohen & Levesque
1990), a theory
pertaining to the internals of rational agents. With this
model, some expectations about the FDG structures can already
be formed by the system. When receiving input from the user,
the system will attempt to match the input from the user with
the (partial) structures that can already be formed at the FDG
level based on the BDI user model.
Our
poster will focus on the difficulties that we encounter in
tackling these problems.
|
|
|
|
| References: |
-
Cohen,
Philip R., and Levesque Hector J. 1990 . Intention is choice with
commitment. Artificial Intelligence,
42(2–3):213–261.
-
Hengeveld, Kees and Mackenzie J. Lachlan. 2008. Functional
Discourse Grammar.
Oxford: Oxford University Press.
|
|
|
