Understanding Metonymies in Discourse

Earlier Approaches to Computational Metonymy Resolution

• Recognize syntactic and semantic irregularities that are associated with metonyms
• Syntactic: feature agreement (“The french fries is getting impatient.”)
• Syntactic: phrase structure (“He read the Shakespeare.”)
• Semantic: selection restriction violations (“Denise drank the bottle.”)

Earlier Approaches (contd.)

• Recognize two types of metonymy
• Referential: The metonymic noun phrase has an intended referent related to what the metonym literally refers to.
• (Ex. “The ham sandwich is waiting to pay.”
• “He (*it) is impatient.” “ham sandwich” <=>
• “customer”)
• Predicative: The intended referent of the metonymic noun phrase is the same as what the metonym literally refers to.
• (Ex. “Nixon bombed Hanoi.” “He (*they) wanted to force the Communists to negotiate.” “Nixon” <=> “Nixon”)

Earlier Approaches (contd.)

• Recognize patterns in the relationships between metonyms and the intended referents.
• Producer-for-product (“He bought a Ford.”)
• Part-for-whole (“I liked the laser….”)
• Place-for-institution (“The White House said….”)
• Container-for-contents (“….drank the bottle”)
• Artist-for-artworks (“He bought a Picasso.”)

Earlier Approaches: Metonymy Resolution Algorithm

• Prefer literal interpretation first
• If a selection restriction violation is detected, consider possible metonymy and attempt resolution as follows:
• Determine what types the selection restriction permits
• Identify concepts related to the supposed metonym that are of the permitted types
• Constrain set of candidate concepts (and the related words and phrases) when possible using known relationship patterns and syntactic and referential/predicative metonymy information
• Coerce the metonym to a type permitted by the selection restriction by making it refer to the chosen related concept.