Why is Winoground Hard? Investigating Failures in Visuolinguistic Compositionality

Authors: Anuj Diwan, Layne Berry, Eunsol Choi, David Harwath, Kyle Mahowald
EMNLP 2022 Paper

❗ This code release is still a work-in-progress: please raise an issue or send an email to anuj.diwan@utexas.edu and layne.berry@utexas.edu for questions.

Setup

conda create -n winoground python=3.6.9
conda activate winoground
pip install -r requirements.txt

Next, follow the installation instructions at https://github.com/GEM-benchmark/NL-Augmenter to install NL-Augmenter.

Download the Winoground dataset from https://huggingface.co/datasets/facebook/winoground. Place examples.jsonl and the extracted directory images inside dataset.

Generating Augmented Examples

Run

python augmentations/text_augmentations.py

to find the generated file at examples_augmented.jsonl.

UNITER

For all experiments, when running UNITER code, first start a Docker container using the following code, then proceed as normal. All absolute paths /path are now /slash/path

bash UNITER/launch_container_simpler.sh
bash UNITER/run_init_docker.sh

Section 3.1 Reproduction

First, run get_MODEL_scores.py for the model you're investigating to collect pairwise similarity scores for all 800x800 image-text combinations in Winoground.

Then, modify the variable "model" in recall_at_k.py to specify which model you are testing. Running recall_at_k.py will then output the R@1,2,5,10 scores for the I2T and T2I directions for the specified model.

Section 3.2 Reproduction

First, run get_MODEL_feats.py for the model you're investigating to collect the embeddings at each layer for all 400x4 possible combinations of inputs within a Winoground set (I0+T0, I0+T1, I1+T0, and I1+T1).

Next, edit the "file_to_split" variable in split_train_test.py and run it to generate the stratified train and test splits we used.

Finally, run finetune_over_pooled_outputs/run_fused_embeddings_probe.py to perform the test. You can use the --train_path and --test_path command line parameters to specify the dataset to probe. A number of other command line parameters are available to configure the size of the probe, layer being probed, method for generating embeddings from a layer (i.e., CLS, Mean- or Max-Pooling), etc.

Section 4 Reproduction

The Winoground sets assigned each newly introduced tag are provided in the file new_tag_assignments.json as a dictionary.

Section 5.1 Reproduction

First, generate the augmented captions as specified above. Use get_MODEL_feats.py to generate embeddings of the augmentations.

Next, edit "file_to_split" in split_train_test.py and run it to generate the stratified train and test splits we used.

Finally, run SVC_linear_separability_by_example/probe_MODEL.py, where MODEL is UNITER, LXMERT, or CLIP. Use the command line arguments (visible via "--help" or at the top of the file) to configure the probe.

Section 5.3 Reproduction

Generate augmented captions, embeddings of those captions, and the stratified train and test splits as for 5.1. Then, run full_dataset_separability/MODEL/run_unimodal_text_variants_probe.py to train and test a probe. Use the command line arguments (visible via "--help" or at the top of the file) to configure the probe.

Citations and Contact

Please cite our paper if you use our paper, code, finegrained Winoground tags or the augmented Winoground examples in your work:

@inproceedings{diwan-etal-2022-winoground,
    title = "Why is Winoground Hard? Investigating Failures in Visuolinguistic Compositionality",
    author = "Diwan, Anuj  and
      Berry, Layne  and
      Choi, Eunsol  and
      Harwath, David  and
      Mahowald, Kyle",
    booktitle = "Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing",
    month = dec,
    year = "2022",
    address = "Abu Dhabi, United Arab Emirates",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2022.emnlp-main.143",
    pages = "2236--2250",
    abstract = "Recent visuolinguistic pre-trained models show promising progress on various end tasks such as image retrieval and video captioning. Yet, they fail miserably on the recently proposed Winoground dataset, which challenges models to match paired images and English captions, with items constructed to overlap lexically but differ in meaning (e.g., {``}there is a mug in some grass{''} vs. {``}there is some grass in a mug{''}). By annotating the dataset using new fine-grained tags, we show that solving the Winoground task requires not just compositional language understanding, but a host of other abilities like commonsense reasoning or locating small, out-of-focus objects in low-resolution images. In this paper, we identify the dataset{'}s main challenges through a suite of experiments on related tasks (probing task, image retrieval task), data augmentation, and manual inspection of the dataset. Our analysis suggests that a main challenge in visuolinguistic models may lie in fusing visual and textual representations, rather than in compositional language understanding. We release our annotation and code at https://github.com/ajd12342/why-winoground-hard.",
}

Please also cite the wonderful paper that introduces the Winoground dataset:

@InProceedings{Thrush_2022_CVPR,
    author    = {Thrush, Tristan and Jiang, Ryan and Bartolo, Max and Singh, Amanpreet and Williams, Adina and Kiela, Douwe and Ross, Candace},
    title     = {Winoground: Probing Vision and Language Models for Visio-Linguistic Compositionality},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {5238-5248}
}

Feel free to contact anuj.diwan@utexas.edu and layne.berry@utexas.edu with any questions!