Documentation | Colab Notebook | Blog Post

A toolkit for incorporating multimodal data on top of text data for classification and regression tasks. It uses HuggingFace transformers as the base model for text features. The toolkit adds a combining module that takes the outputs of the transformer in addition to categorical and numerical features to produce rich multimodal features for downstream classification/regression layers. Given a pretrained transformer, the parameters of the combining module and transformer are trained based on the supervised task. For a brief literature review, check out the accompanying blog post on Georgian's Impact Blog.

The code was developed in Python 3.7 with PyTorch and Transformers 4.26.1. The multimodal specific code is in multimodal_transformers folder.

pip install multimodal-transformers

The following Hugging Face Transformers are supported to handle tabular data. See the documentation here.

• BERT from Devlin et al.: BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (ACL 2019)
• ALBERT from Lan et al.: ALBERT: A Lite BERT for Self-supervised Learning of Language Representations (ICLR 2020)
• DistilBERT from Sanh et al.: DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter (NeurIPS 2019)
• RoBERTa from Liu et al.: RoBERTa: A Robustly Optimized BERT Pretraining Approach
• XLM from Lample et al.: Cross-lingual Language Model Pretraining (NeurIPS 2019)
• XLNET from Yang et al.: XLNet: Generalized Autoregressive Pretraining for Language Understanding (NeurIPS 2019)
• XLM-RoBERTa from Conneau et al.: Unsupervised Cross-lingual Representation Learning at Scale (ACL 2020)

This repository also includes two kaggle datasets which contain text data and rich tabular features

• Women's Clothing E-Commerce Reviews for Recommendation Prediction (Classification)
• Melbourne Airbnb Open Data for Price Prediction (Regression)
• PetFindermy Adoption Prediction for Pet Adoption Speed Prediction (Multiclass Classification)

To quickly see these models in action on say one of the above datasets with preset configurations

$ python main.py ./datasets/Melbourne_Airbnb_Open_Data/train_config.json

Or if you prefer command line arguments run

$ python main.py \
    --output_dir=./logs/test \
    --task=classification \
    --combine_feat_method=individual_mlps_on_cat_and_numerical_feats_then_concat \
    --do_train \
    --model_name_or_path=distilbert-base-uncased \
    --data_path=./datasets/Womens_Clothing_E-Commerce_Reviews \
    --column_info_path=./datasets/Womens_Clothing_E-Commerce_Reviews/column_info.json

main.py expects a json file detailing which columns in a dataset contain text, categorical, or numerical input features. It also expects a path to the folder where the data is stored as train.csv, and test.csv(and if given val.csv).For more details on the arguments see multimodal_exp_args.py.

To see the modules come together in a notebook:

In practice, taking the categorical and numerical features as they are and just tokenizing them and just concatenating them to the text columns as extra text sentences is a strong baseline. To do that here, just specify all the categorical and numerical columns as text columns and set combine_feat_method to text_only. For example for each of the included sample datasets in ./datasets, in train_config.json change combine_feat_method to text_only and column_info_path to ./datasets/{dataset}/column_info_all_text.json.

In the experiments below this baseline corresponds to Combine Feat Method being unimodal.

The following tables shows the results on the two included datasets's respective test sets, by running main.py Non specified parameters are the default.

Specific training parameters can be seen in datasets/Womens_Clothing_E-Commerce_Reviews/train_config.json.

There are 2 text columns, 3 categorical columns, and 3 numerical columns.

Specific training parameters can be seen in datasets/Melbourne_Airbnb_Open_Data/train_config.json.

There are 3 text columns, 74 categorical columns, and 15 numerical columns.

Specific training parameters can be seen in datasets/PetFindermy_Adoption_Prediction There are 2 text columns, 14 categorical columns, and 5 numerical columns.

We now have a paper you can cite for the Multimodal-Toolkit.

@inproceedings{gu-budhkar-2021-package,
    title = "A Package for Learning on Tabular and Text Data with Transformers",
    author = "Gu, Ken  and
      Budhkar, Akshay",
    booktitle = "Proceedings of the Third Workshop on Multimodal Artificial Intelligence",
    month = jun,
    year = "2021",
    address = "Mexico City, Mexico",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2021.maiworkshop-1.10",
    doi = "10.18653/v1/2021.maiworkshop-1.10",
    pages = "69--73",
}