Obviously, the meaning of a text is more than just its explicit
meaning. On the basis of what we hear, of our general linguistic and
non-linguistic knowledge and of the situational context, we draw a
number of inferences and thereby reconstruct the (more or less) full
meaning of the message intended by our interlocutor. We use inference
for instance to: resolve anaphors and ellipses, establish relational
links between discourse segments and recover implicit
implicatures. Clearly, we draw these inferences very efficiently and
the question naturally arises of how to model such efficiency in
natural language processing systems.
In the field of theorem proving on the other hand, much progress has
been made in the last decades, and theorem provers are now available
which perform very efficiently on at least certain classes of
mathematical problems.
The goal of the LISA project is to develop representation formalisms
and inference techniques that are tailored to the linguistic task of
discourse comprehension and generation.
In the first phase, we have concentrated on higher-order
representation formalisms and inference procedures to account for the
reconstruction of discourse material that is not linguistically
realized (verb-phrase ellipses are an example of this) and more
generally, of material that is semantically redundant
(e.g. deaccenting). We have shown in particular that certain forms of
Higher-Order Unification (HOCU, HOUTP) developed in the theorem
proving community are more appropriate than others for the natural
language understanding task.
In the second phase, we experiment with more general inference
procedures and their relevance for natural language processing. In
particular, we investigate the usability of general purpose
theorem provers for the generation task; we use model generation and
abduction to incorporate world knowledge into the semantic analysis
process; and we attempt to extend these findings to discourse
representation formalisms.
|