Multimodal Transformers | Transformers with Tabular Data

---

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.

Installation

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

Supported 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)

Included Datasets

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)

Working Examples

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.

Notebook Introduction

To see the modules come together in a notebook:

Included Methods

combine feat method	description	requires both cat and num features
text_only	Uses just the text columns as processed by a HuggingFace transformer before final classifier layer(s). Essentially equivalent to HuggingFace's ForSequenceClassification models	False
concat	Concatenate transformer output, numerical feats, and categorical feats all at once before final classifier layer(s)	False
mlp_on_categorical_then_concat	MLP on categorical feats then concat transformer output, numerical feats, and processed categorical feats before final classifier layer(s)	False (Requires cat feats)
individual_mlps_on_cat_and_numerical_feats_then_concat	Separate MLPs on categorical feats and numerical feats then concatenation of transformer output, with processed numerical feats, and processed categorical feats before final classifier layer(s).	False
mlp_on_concatenated_cat_and_numerical_feats_then_concat	MLP on concatenated categorical and numerical feat then concatenated with transformer output before final classifier layer(s)	True
attention_on_cat_and_numerical_feats	Attention based summation of transformer outputs, numerical feats, and categorical feats queried by transformer outputs before final classifier layer(s).	False
gating_on_cat_and_num_feats_then_sum	Gated summation of transformer outputs, numerical feats, and categorical feats before final classifier layer(s). Inspired by Integrating Multimodal Information in Large Pretrained Transformers which performs the mechanism for each token.	False
weighted_feature_sum_on_transformer_cat_and_numerical_feats	Learnable weighted feature-wise sum of transformer outputs, numerical feats and categorical feats for each feature dimension before final classifier layer(s)	False

Simple baseline model

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.

Results

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.

Review Prediction

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.

Model	Combine Feat Method	F1	PR AUC
Bert Base Uncased	text_only	0.957	0.992
Bert Base Uncased	unimodal	0.968	0.995
Bert Base Uncased	concat	0.958	0.992
Bert Base Uncased	individual_mlps_on_cat_and_numerical_feats_then_concat	0.959	0.992
Bert Base Uncased	attention_on_cat_and_numerical_feats	0.959	0.992
Bert Base Uncased	gating_on_cat_and_num_feats_then_sum	0.961	0.994
Bert Base Uncased	weighted_feature_sum_on_transformer_cat_and_numerical_feats	0.962	0.994

Pricing Prediction

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.

Model	Combine Feat Method	MAE	RMSE
Bert Base Multilingual Uncased	text_only	82.74	254.0
Bert Base Multilingual Uncased	unimodal	79.34	245.2
Bert Base Uncased	concat	65.68	239.3
Bert Base Multilingual Uncased	individual_mlps_on_cat_and_numerical_feats_then_concat	66.73	237.3
Bert Base Multilingual Uncased	attention_on_cat_and_numerical_feats	74.72	246.3
Bert Base Multilingual Uncased	gating_on_cat_and_num_feats_then_sum	66.64	237.8
Bert Base Multilingual Uncased	weighted_feature_sum_on_transformer_cat_and_numerical_feats	71.19	245.2

Pet Adoption Prediction

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

Model	Combine Feat Method	F1_macro	F1_micro
Bert Base Multilingual Uncased	text_only	0.088	0.281
Bert Base Multilingual Uncased	unimodal	0.089	0.283
Bert Base Uncased	concat	0.199	0.362
Bert Base Multilingual Uncased	individual_mlps_on_cat_and_numerical_feats_then_concat	0.244	0.352
Bert Base Multilingual Uncased	attention_on_cat_and_numerical_feats	0.254	0.375
Bert Base Multilingual Uncased	gating_on_cat_and_num_feats_then_sum	0.275	0.375
Bert Base Multilingual Uncased	weighted_feature_sum_on_transformer_cat_and_numerical_feats	0.266	0.380

Citation

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",
}