MetoTalk: Understanding Metonymies

Dr. Katja Markert, Language Technology Group, University of Edinburgh

Metonymy is the figure of speech where "we are using one entity to refer to another that is related to it", according to a well-known definition by Lakoff and Johnson (1980). Typical examples are "The ham sandwich is waiting for his check" or "I love Shakespeare". In my talk I will present an automatic method for recognising and resolving metonymies --- MetoTalk. MetoTalk is a knowledge-based and discourse-oriented algorithm. It takes cognitive results on the understanding of figurative language as well as data from naturally-occurring texts into account.

In particular, I make the following claims:

  1. Relying on the violation of selectional restrictions for recognising figurative language is in general inadequate for metonymy resolution. It contradicts both cognitive evidence and text data.
  2. Previous discourse is a major factor in disambiguating between several literal and metonymic readings. MetoTalk supports a strong interaction between anaphora and metonymy resolution, improving the resolution rates of both significantly.
  3. World knowledge can make a contribution to the resolution of metonymies that goes beyond relation identification. In MetoTalk's disambiguation component, we use it for judging the aptness of metonymic readings by judging e.g. the typicality of the underlying semantic relations.
In the last part of the talk, I will discuss the advantages and shortcomings of the presented model and talk about ideas for a statistical model of metonymy resolution.