When Compositionality Fails to Predict Systematicity

In this paper we argue that compositionality alone isn't sufficient to derive systematicity clauses for meanings (meaning = truth-functional content of an utterance). In a seminal paper by Fodor & Pylyshyn (1988) the 'systematicity ' of adjectival modification has been derived from the stipulation that the syntactic operation of modification relates to the semantic operation of conjunction. We argue that this assumption fails already in the case of so-called 'absolute ' adjectives. For instance, in order to determine the truth-functional content of a phrase like red apple it must be specified which parts of the apple are red, and this cannot be described by conjoining the adjective with the noun meaning. Our answer to this and related puzzles is to take the meaning of the adjective as a function that applies to the meaning of the noun. Though compositional in the formal sense, this operation fails to predict any kind of systematicity for adjectival modification. In order to account for certain regularities we have to pose restrictions on the set of possible functions that count as adjective meanings. We are able to pose the right restrictions using a connectionist network that respects the role of encyclopedic knowledge. We argue that this model also is able to derive the prototype effects discussed in connection with adjectival modification. Copyright © 2004, American Association for Artificial Intelligenc