Freda Shi*, Jiayuan Mao*, Kevin Gimpel, Karen Livescu

ACL 2019   Best Paper Nominee [paper] [project page] [bib]

PyTorch >= 1.0.1

See also env/conda_env.txt for detailed (but maybe not necessary) environment setup.

Download our pre-processed data here and unzip it to the data folder.

To run our demo, the data folder should be organized as follows

data
├── mscoco
│   ├── dev_caps.txt
│   ├── dev_ims.npy
│   ├── test_caps.txt
│   ├── test_ground-truth.txt
│   ├── train_caps.txt
│   ├── train_ims.npy
│   └── vocab.pkl
...

Note: The demos are NOT used to reproduce the numbers reported in the ACL paper. To reproduce the results, we need to train multiple models and do self F1 driven model selection (see Section 4.7 for details). To get roughly similar numbers to the ones reported in the paper, take the 6th or 7th checkpoint of VG-NSL and the checkpoints after 20 epochs for VG-NSL+HI. Thanks to Nori for bringing this into our attention.

VG-NSL

bash demos/demo_train.sh

VG-NSL+HI

bash demos/demo_train_head-initial.sh

After training a model, we can test it by running

bash demos/demo_test.sh

If you found the codebase useful, please consider citing

@inproceedings{shi2019visually,
    Title = {Visually Grounded Neural Syntax Acquisition},
    Author = {Shi, Haoyue and Mao, Jiayuan and Gimpel, Kevin and Livescu, Karen},
    Booktitle = {Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics},
    Year = {2019}
}

The visual-semantic embedding part is adapted from the codebase of VSE++ (Faghri et al., BMVC 2018) and VSE-C (Shi et al., COLING 2018). Part of the basic code is adapated from Jiayuan's personal Python toolkits Jacinle and Freda's toolkits Witter. We also thank Victor Silva for providing the original concreteness estimation codebase (Hessel et al., NAACL-HLT 2018).

MIT