On Modeling Meaning Shifts by Relaxing Underspecified Semantic Representations

Author: Kristina Striegnitz

Editor:

The context in which a word appears in natural language often influences its interpretation in such a way that the base meaning of the word is changed or made more specific. Polysemy and metonymy are examples for this phenomenon. These meaning shifts of words can be modeled by augmenting the semantic representation of a natural language utterance with the information that is missing to make the shift in meaning explicit. This information can be provided by linguistic or non-linguistic sources or an interaction of both. Recently, Egg (1999) has suggested an account of meaning shifts which exploits underspecification methods to yield a monotonic augmentation process. The main idea is to have semantic construction derive a sufficiently relaxed (i.e. made less specific) semantic representation, so that adding the missing information is simply further specification of this representation. This thesis will examine a treatment of meaning shifts due to systematic polysemy or metonymy within Egg's framework. We will present a syntax/semantics interface which derives appropriately relaxed semantic representations. To account for meaning shifts these representations can be augmented monotonically with additional information. We will point out a potential problem for this approach: making underspecified semantic representations less specific may cause overgeneration. However, as we will show, for our applications relaxation is safe, i.e. there is no danger of overgeneration. The underspecification formalism that we will use throughout this thesis is in the class of tree description languages subsuming dominance constraints. We will distinguish a novel class of subconstraints with a certain structure which powerfully support the type of inferences on dominance and disjointness which we have to make for proving safety of relaxation.