# Metaphor collision as a detection signal > Source-anchored research for Metaphor collision as a detection signal. Run status: COMPLETED. Parent: [Mixed metaphors and catachresis as signs of reflexive figurative language](https://llms-explorer.com/tree/ai-isms-mixed-metaphors/) · 2 facets · 3 facts · page: https://llms-explorer.com/tree/metaphor-collision-as-a-detection-signal/ ## How it works - MIPVU labels a word metaphor-related when its contextual meaning contrasts with a more basic meaning and can be understood by comparison with it. It is a per-lexical-unit procedure; two colliding domains would be found only by labelling each word separately and then comparing their source domains. (medium confidence) — [source](https://pmc.ncbi.nlm.nih.gov/articles/PMC10666762/) - Incongruity-based metaphor detection is invalid for conventional metaphors, because the surrounding context is usual for the target word. A collision signal is therefore strongest for novel or freshly chosen figures and weak for dead ones. (medium confidence) — [source](https://arxiv.org/pdf/2212.08395) ## Facts and statements - Cross-dataset evaluation finds LLM metaphor detection tracks lexical overlap and sentence length, and accuracy on literal paraphrases often drops or paraphrases keep the metaphor. An LLM-based collision detector should not be trusted without a check against a source-domain lexicon. (low confidence) (verified as of 2025-07-21) — [source](https://arxiv.org/html/2507.15357v1)