Do Vision and Language Models Share Concepts? A Vector Space Alignment Study

Jiaang Li, Yova Kementchedjhieva, Constanza Fierro, Anders Søgaard


Abstract
Large-scale pretrained language models (LMs) are said to “lack the ability to connect utterances to the world” (Bender and Koller, 2020), because they do not have “mental models of the world” (Mitchell and Krakauer, 2023). If so, one would expect LM representations to be unrelated to representations induced by vision models. We present an empirical evaluation across four families of LMs (BERT, GPT-2, OPT, and LLaMA-2) and three vision model architectures (ResNet, SegFormer, and MAE). Our experiments show that LMs partially converge towards representations isomorphic to those of vision models, subject to dispersion, polysemy, and frequency. This has important implications for both multi-modal processing and the LM understanding debate (Mitchell and Krakauer, 2023).1
Anthology ID:
2024.tacl-1.68
Volume:
Transactions of the Association for Computational Linguistics, Volume 12
Month:
Year:
2024
Address:
Cambridge, MA
Venue:
TACL
SIG:
Publisher:
MIT Press
Note:
Pages:
1232–1249
Language:
URL:
https://preview.aclanthology.org/icon-24-ingestion/2024.tacl-1.68/
DOI:
10.1162/tacl_a_00698
Bibkey:
Cite (ACL):
Jiaang Li, Yova Kementchedjhieva, Constanza Fierro, and Anders Søgaard. 2024. Do Vision and Language Models Share Concepts? A Vector Space Alignment Study. Transactions of the Association for Computational Linguistics, 12:1232–1249.
Cite (Informal):
Do Vision and Language Models Share Concepts? A Vector Space Alignment Study (Li et al., TACL 2024)
Copy Citation:
PDF:
https://preview.aclanthology.org/icon-24-ingestion/2024.tacl-1.68.pdf