Exploiting BERT for Multimodal Target Sentiment Classification Through Input Space Translation

Running The Code

Data

Download the Twitter-17 dataset here, and the Twitter-15 dataset here. You don't need the images to run the code in the repo, but if you want to download them, there are instructions in the TomBERT repo here.

Generating Captions

The captions used in the paper are provided in the captions/ directory. If you want to generate your own captions, clone the repo with git clone --recurse-submodules and move the file caption_multiple.py into the cloned CATR repository. Make sure all the requirements for CATR are installed, and run caption_multiple.py.

Training & Evaluation

The training/eval scripts are very straightforward, and follow the same structure. Looking at EF_CapBERT_Tw15.py as a concrete example, all you need to do is edit the following lines:

train_tsv = "/path/to/the/file"
dev_tsv = "/path/to/the/file" 
test_tsv = "/path/to/the/file" 
captions_json = "/path/to/the/file"

to match where you've placed the files.

Citations

If you found this paper useful, citing the paper and dataset would be greatly appreciated.

@inproceedings{khanExploitingBERTTranslation,
  title = {Exploiting {{BERT}} for {{Multimodal Target Sentiment Classification Through Input Space Translation}}},
  booktitle = {{{MM}} '21: {{The}} 29th {{ACM}} International Conference on Multimedia, Virtual Event / China October 20-24, 2021},
  author = {Khan, Zaid and Fu, Yun},
  year = {2021},
  publisher = {{ACM}},
  doi = {10.1145/3474085.3475692},
}


@inproceedings{yuAdaptingBERTTargetOriented2019,
  title = {Adapting {{BERT}} for {{Target}}-{{Oriented Multimodal Sentiment Classification}}},
  booktitle = {Proceedings of the {{Twenty}}-{{Eighth International Joint Conference}} on {{Artificial Intelligence}}},
  author = {Yu, Jianfei and Jiang, Jing},
  year = {2019},
  month = aug,
  pages = {5408--5414},
  publisher = {{International Joint Conferences on Artificial Intelligence Organization}}
}

codezakh / exploiting-BERT-thru-translation