Using Answer Set Programming in an Inference-Based approach to Natural Language Semantics

The traditional tri-partition syntax/semantics/pragmatics is commonly used in most of the computer systems that aim at the simulation of the human understanding of Natural Language (NL). This conception does not reflect the flexible and creative manner that humans use in reality to interpret texts. Generally speaking, formal NL semantics is referential i.e. it assumes that it is possible to create a static discourse universe and to equate the objects of this universe to the (static) meanings of words. The meaning of a sentence is then built from the meanings of the words in a compositional process and the semantic interpretation of a sentence is reduced to its logical interpretation based on the truth conditions. The very difficult task of adapting the meaning of a sentence to its context is often left to the pragmatic level, and this task requires to use a huge amount of common sense knowledge about the domain. This approach is seriously challenged (see for example [4][14]). It has been showed that the above tri-partition is very artificial because linguistic as well as extra-linguistic knowledge interact in the same global process to provide the necessary elements for understanding. Linguistic phenomena such as polysemy, plurals, metaphors and shifts in meaning create real difficulties to the referential approach of the NL semantics discussed above. As an alternative solution to these problems, [4] proposes an inferential approach to the NL semantics in which words trigger inferences depending on the context of their apparition. In the same spirit we claim that understanding a NL text is a reasoning process based on our knowledge about the norms1 of its domain i.e. what we generally expect to happen in normal situations. But what kind of reasoning is needed for natural language semantics?