Team LeD

Members
Overall Objectives
Scientific Foundations
Application Domains
Software
New Results
Other Grants and Activities
Dissemination
Bibliography

Section: Scientific Foundations

Semantics and Inference

Over the next decade, progress in natural language semantics will likely depend on obtaining a deeper understanding of the role played by inference. One of the simplest levels at which inference enters natural language is as a disambiguation mechanism. Utterances in natural language are typically highly ambiguous: inference allows human beings to (seemingly effortlessly) eliminate the irrelevant possibilities and isolate the intended meaning. But inference can be used in many other processes, for example, in the integration of new information into a known context. This is important when generating natural language utterances. For this task we need to be sure that the utterance we generate is suitable for the person being addressed. That is, we need to be sure that the generated representations fit in well with the recipient's knowledge and expectations of the world, and it is inference which guides us in achieving this.

Much recent semantic research actively addresses such problems by systematically integrating inference as a key element. This is an interesting development, as such work redefines the boundary between semantics and pragmatics. For example, van der Sandt's algorithm for presupposition resolution (a classic problem of pragmatics) uses inference to guarantee that new information is integrated in a coherent way with the old information.

The LED team investigates such semantic/pragmatic problems from various angles (for example, from generation and discourse analysis perspectives) and tries to combine the insights offered by different approaches. For example, for some applications (e.g., the textual entailment task) shallow syntactic parsing combined with fast inference in description logic may be the most suitable approach. In other cases, deep analysis of utterances or sentences and the use of a first-order inference engine may be better. Our aim is to explore these approaches and their limitations.


previous
next

Logo Inria