HLT — Metaphor Identification (overview)

This is a lexical note — a longer, lecture-style write-up that contains several distinct ideas. The atomic zettels mipvu-procedure, conceptual-metaphor-theory, and metaphor-as-classification are extracted from it.

What is metaphor identification, computationally?

Two traditions sit underneath the field:

1. MIPVU (“Metaphor Identification Procedure VU”) — the linguistic annotation procedure developed at the VU. It walks each lexical unit and asks whether its contextual meaning differs from a more basic meaning, in which case it is marked as metaphor-related (see mipvu-procedure).

2. Conceptual Metaphor Theory — the cognitive-linguistic frame from Lakoff & Johnson, in which metaphors map source domains onto target domains. ARGUMENT IS WAR. LIFE IS A JOURNEY. The cognitive level, not the lexical one (see conceptual-metaphor-theory).

These are different operationalisations of the same phenomenon, and a lot of recent NLP work is essentially a tug-of-war between them.

Computational systems

In NLP, metaphor identification has been framed as token-level binary classification: each token is or is not metaphor-related. Early systems used hand-crafted features (concreteness, imageability) on top of WordNet relations; modern systems use contextual embeddings. See metaphor-as-classification.

Open problems

Annotation cost is high. Disagreement between annotators is high. Cross-domain generalisation is poor. The relationship between MIPVU and Conceptual Metaphor Theory is unresolved at the dataset level.