Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Measuring Internal Representational Alignment between Conceptual Metaphor Domains

Afjal Chowdhury¹, Zachary Land²; ¹Massachusetts Institute of Technology, ²Imperial College London

Presenter: Afjal Chowdhury

The shared structure exhibited between neural representations is commonly quantified using representational similarity analysis (RSA). A notion of shared structure arises in Conceptual Metaphor Theory, where a "conceptual domain" is understood through the relational structure of another. We propose using RSA to compare conceptual domain word embeddings to detect metaphor in a *single* model's representation space. We introduce a graph-based RSA metric, Graph Edit Distance (GED), to extend methods like CKA by comparing the topology of sparse graphs versus dense similarity matrices. We demonstrate that both methods effectively detect metaphor, with each identifying cases that the other missed, suggesting that they both offer complementary value when measuring alignment within a single representation, which we dub *intra-representational alignment*.

Topic Area: Methods, Tools, Theory & Neural Coding