Feature Integration in Metonymy Resolution

Cognitive Science
Human Communication Research Centre, Edinburgh

Extended Abstract

1. Regarding recognition, selectional restriction violations are most commonly used. Here, selectional restrictions are regarded as constraints predicates impose on their arguments, e.g. "read" imposes on its object that it is a kind of document. Many researchers regard it as mandatory that metonymies violate such selectional restrictions (as Example (1) and (2)) and can be recognised by noticing such violations. In these approaches, metonymy resolution will only be triggered if the literal reading violates selectional restrictions.

   Example (1): "I read Shakespeare."
   Example (2) : "The ham sandwich is waiting to pay."

2. For interpretation, a combination of selectional restrictions and world knowledge or lexical knowledge is common. Thus, in the above example, we know that the type constraint "wait" imposes on its argument is Document, and that thus Shakespeare stands metonymically for a document. Knowledge of the world then tells us that Shakespeare was an author and wrote plays and that therefore Shakespeare stands metonymically for one of its plays. In cases of ambiguous metonymic readings, so-called schemata are employed, preferring conventional metonymies over unconventional ones. Conventionality is often identified with belonging to a schema: in Example (1) the metonymy belongs to an Author-for-Works schema, whereas Example (2) is unconventional.

I will argue in my talk that this model is far too restrictive. In particular I make the following claims:

1. Selectional restriction violations are not sufficient for triggering metonymy resolution. This has the following reasons:
   1. Selectional restrictions are just one of a wide range of semantic constraints. Other semantic constraints are sometimes hard to identify and are also often defeasible (see Example (3)).
   2. Many metonymies without any violations exist (see Example (3) and (4)).

   Example (3): "Shakespeare is interesting."
   Example (4): "He threw the book on the floor. Now, the poor author was lying there. (Bonhomme, 1987).

2. In many cases (see Example (3) and (4)) only the incorporation of the additional features of world knowledge and discourse constraints allows metonymy recognition and interpretation.
3. The preference for schematic metonymies can also be overridden by discourse constraints (see Example (5)).

   Example (5): "(uttered while looking at T-shirts with heads of English writers) I particularly liked the Shakespeare."

I will then focus on the contribution discourse, especially referential constraints (see Example (4) and (5)) make to metonymy resolution. In the ensuing model, literal and metonymic interpretations are computed on a par and disambiguated with the help of discourse constraints. Readings allowing for anaphoric readings are preferred.

This model yields too many ambiguities, however. It is farther constrained by including another feature into disambiguation, which is hardly ever noticed. I claim that metonymic relationships are constraint by aptness features, embodied in world knowledge. This concerns the question why e.g. some parts are more likely to stand for a whole than another. So "screen' is a more apt metonymy for a TV than "transistor", disregarding specific discourse contexts. This causes schemata exceptions. I claim that these aptness considerations capture the typicality of relationships and will present a formalisation of these constraints, which is rooted in psychological research by Tversky and Hemenway as well as Blasko and Connine.

As a summary, my talk will show that individual features are all inconclusive and that only an integration of as many features as possible can adequately explain metonymic usage. This includes two other features not used in my model yet. Firstly, language-specific constraints, accounting for schemata exceptions, that cannot be explained by aptness or world knowledge. Thus, in English "I eat pig" is ungrammatical as the lexeme "pork" already exists, whereas "I eat lamb" is a normal metonymic usage (Blocking). Secondly, conceptual inference that can explain that Example (3) is unlikely to be used literally when uttered by a speaker in our time.

I will finish with an outlook on probabilistic models for metonymy resolutions. This is motivated by the fact that symbolic models are hard-pressed to find an integration weighing the influence of all these features, whereas frequency considerations and conditional probabilities might solve this problem.

to SFB-Homepage

created by: Anke Weinberger (2000-08-30). maintained by: Anke Weinberger (2000-10-04).