Concept Representations in Vector Space: Quantifying Sharedness and Idiosyncrasy of Meaning

Key points

• Concepts are the basic units through which we navigate the world and communicate meaning.1
• But do you and I really mean the same thing when we think or speak about a concept?1
• Short personal definitions, elicited from $N = 125$ participants, are encoded as sentence embeddings and submitted pairwise to a crossed random effects model yielding two concept-level quantities: sharedness ($Sb$), the expected dissimilarity between two randomly drawn individuals, and idiosyncrasy ($Ib$), the person-specific share of total dyadic variance.1
• Two validity tests support the approach: the embedding space fully recovers concept and domain organization from participant definitions even with the target word excluded, and a parameter-free classifier achieves 0.630 accuracy in a within-participant extensional prediction task against a chance baseline of 0.500 and a consistency ceiling of 0.893.1

Sources (1)

• [1]Concept Representations in Vector Space: Quantifying Sharedness and Idiosyncrasy of Meaning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 11, 09:52 AM

  The present study develops a measurement framework for modeling and analyzing interindividual variation in concept representations through sentence embeddings, dense vector representations mapping entire utterances into a shared geometric space.

  Concepts are the basic units through which we navigate the world and communicate meaning.

Extractive summary: sentences quoted from the sources.