Challenging the Principle of Compositionality in Interpreting Natural Language Texts

Abstract

The main assumption of many contemporary semantic theories, from Montague grammars to the most recent papers in journals of linguistics semantics, is and remains the principle of compositionality. It is also adopted by more computation-oriented traditions (Artificial Intelligence or Natural Language Processing - henceforth, NLP). Although its adequacy to ”real” language clearly raises several difficulties, it is almost never challenged and the presuppositions on which it relies are seldom questioned. Rather than proposing some more arrangements to overcome some of its drawbacks, our point, in this paper, will actually consist in a more radical critique which addresses first the very notion of meaning. For us, a semantic theory should not focus on finding ”the meaning of a complex expression” but rather on drawing adequate inferences. This shift of focus allows us to present an approach which fulfills the tasks that the notion of meaning is supposed to carry out, without an explicit (and compositional) construction of the meaning of parts and wholes.