%0 Conference Paper %1 donmez-etal-2023-hnc %A Dönmez, Esra %A Tilli, Pascal %A Yang, Hsiu-Yu %A Vu, Ngoc Thang %A Silberer, Carina %B Proceedings of the 27th Conference on Computational Natural Language Learning (CoNLL) %C Singapore %D 2023 %E Jiang, Jing %E Reitter, David %E Deng, Shumin %I Association for Computational Linguistics %K EXC2075 PN6-5 PN6-5(II) selected %P 364--388 %R 10.18653/v1/2023.conll-1.24 %T HNC: Leveraging Hard Negative Captions towards Models with Fine-Grained Visual-Linguistic Comprehension Capabilities %U https://aclanthology.org/2023.conll-1.24 %X Image-Text-Matching (ITM) is one of the defacto methods of learning generalized representations from a large corpus in Vision and Language (VL). However, due to the weak association between the web-collected image--text pairs, models fail to show fine-grained understanding of the combined semantics of these modalities. To this end, we propose Hard Negative Captions (HNC): an automatically created dataset containing foiled hard negative captions for ITM training towards achieving fine-grained cross-modal comprehension in VL. Additionally, we provide a challenging manually-created test set for benchmarking models on a fine-grained cross-modal mismatch with varying levels of compositional complexity. Our results show the effectiveness of training on HNC by improving the models' zero-shot capabilities in detecting mismatches on diagnostic tasks and performing robustly under noisy visual input scenarios. Also, we demonstrate that HNC models yield a comparable or better initialization for fine-tuning. Our code and data are publicly available.

@inproceedings{donmez-etal-2023-hnc, abstract = {Image-Text-Matching (ITM) is one of the defacto methods of learning generalized representations from a large corpus in Vision and Language (VL). However, due to the weak association between the web-collected image{--}text pairs, models fail to show fine-grained understanding of the combined semantics of these modalities. To this end, we propose Hard Negative Captions (HNC): an automatically created dataset containing foiled hard negative captions for ITM training towards achieving fine-grained cross-modal comprehension in VL. Additionally, we provide a challenging manually-created test set for benchmarking models on a fine-grained cross-modal mismatch with varying levels of compositional complexity. Our results show the effectiveness of training on HNC by improving the models{'} zero-shot capabilities in detecting mismatches on diagnostic tasks and performing robustly under noisy visual input scenarios. Also, we demonstrate that HNC models yield a comparable or better initialization for fine-tuning. Our code and data are publicly available.}, added-at = {2025-01-27T13:09:00.000+0100}, address = {Singapore}, author = {D{\"o}nmez, Esra and Tilli, Pascal and Yang, Hsiu-Yu and Vu, Ngoc Thang and Silberer, Carina}, biburl = {https://puma.ub.uni-stuttgart.de/bibtex/25944e92567502f815dd268d2a8973b21/testusersimtech}, booktitle = {Proceedings of the 27th Conference on Computational Natural Language Learning (CoNLL)}, doi = {10.18653/v1/2023.conll-1.24}, editor = {Jiang, Jing and Reitter, David and Deng, Shumin}, interhash = {c8bde8198bf201783db70355cc494958}, intrahash = {5944e92567502f815dd268d2a8973b21}, keywords = {EXC2075 PN6-5 PN6-5(II) selected}, month = {12}, pages = {364--388}, publisher = {Association for Computational Linguistics}, timestamp = {2025-01-27T13:09:00.000+0100}, title = {{HNC}: Leveraging Hard Negative Captions towards Models with Fine-Grained Visual-Linguistic Comprehension Capabilities}, url = {https://aclanthology.org/2023.conll-1.24}, year = 2023 }