*** UPDATE March 28th, 2021: GAN-BERT in PyTorch ***
You can find a brand new implementation of GANBERT in PyTorch (compatible with the HuggingFace framework) at the following repository:
https://github.com/crux82/ganbert-pytorch
This PyTorch implementation is just a porting of the Tensorflow one so any bug fixes or suggestions would be greatly appreciated.
This repository contains the code for the paper GAN-BERT: Generative Adversarial Learning for Robust Text Classification with a Bunch of Labeled Examples published in the ACL 2020 - short paper by Danilo Croce (Tor Vergata, University of Rome), Giuseppe Castellucci (Amazon) and Roberto Basili (Tor Vergata, University of Rome). The paper can be found here.
GAN-BERT is an extension of BERT which uses a Generative Adversarial setting to implement an effective semi-supervised learning schema. It allows training BERT with datasets composed of a limited amount of labeled examples and larger subsets of unlabeled material. GAN-BERT can be used in sequence classification tasks (also involving text pairs).
This code runs the GAN-BERT experiment over the TREC dataset for the fine-grained Question Classification task. We provide in this package the code as well as the data for running an experiment by using 2% of the labeled material (109 examples) and 5343 unlabeled examples. The test set is composed of 500 annotated examples.
As a result, BERT trained over 109 examples (in a classification task involving 50 classes) achieves an accuracy of ~13% while GAN-BERT achieves an accuracy of ~42%.
GAN-BERT is an extension of the BERT model within the Generative Adversarial Network (GAN) framework (Goodfellow et al, 2014). In particular, the Semi-Supervised GAN (Salimans et al, 2016) is used to make the BERT fine-tuning robust in such training scenarios where obtaining annotated material is problematic. In fact, when fine-tuned with very few labeled examples the BERT model is not able to provide sufficient performances. With GAN-BERT we extend the fine-tuning stage by introducing a Discriminator-Generator setting, where:
- the Generator G is devoted to producing "fake" vector representations of sentences;
- the Discriminator D is a BERT-based classifier over k+1 categories.
D has the role of classifying an example with respect to the k categories of the task of interest, and it should recognize the examples that are generated by G (the k+1 category). G, instead, must produce representations as much similar as possible to the ones produced by the model for the "real" examples. G is penalized when D correctly classifies an example as fake.
In this context, the model is trained on both labeled and unlabeled examples. The labeled examples contribute to the computation of the loss function with respect to the task k categories. The unlabeled examples contribute to the computation of the loss functions as they should not be incorrectly classified as belonging to the k+1 category (i.e., the fake category).
The resulting model is demonstrated to learn text classification tasks starting from very few labeled examples (50-60 examples) and to outperform the classical BERT fine-tuned models by a large margin in this setting.
In the following plots, the performances of GAN-BERT are reported for different tasks at a different percentages of labeled examples. We measured the accuracy (or F1) of the model for the following tasks: Topic Classification on the 20News (20N) dataset; Question Classification (QC) on the TREC dataset; Sentiment Analysis on the SST dataset (SST-5); Natural Language Inference over the MNLI dataset (MNLI).
The code is a modification of the original Tensorflow code for BERT (https://github.com/google-research/bert). It has been tested with Tensorflow 1.14 over a single Nvidia V100 GPU. The code should be compatible with TPUs, but it has not been tested on such architecture or on multiple GPUs. Moreover, it uses tf_metrics (https://github.com/guillaumegenthial/tf_metrics) to compute some performance measures.
It is suggested to use a python 3.6 environment to run the experiment. If you're using conda, create a new environment with:
conda create --name ganbert python=3.6
Activate the newly created environment with:
conda activate ganbert
And install the required packages by:
pip install -r requirements.txt
This should install both Tensorflow and tf_metrics.
The run_experiment.sh script contains the necessary steps to run an experiment with both BERT and GANBERT.
The script can be launched with:
sh run_experiment.sh
The script will first download the BERT-based model, and then it will run the experiments both with GANBERT and with BERT.
After some time (on an Nvidia Tesla V100 it takes about 5 minutes) there will be two files in the output: qc-fine_statistics_BERT0.02.txt and qc-fine_statistics_GANBERT0.02.txt. These two contain the performance measures of BERT and GANBERT, respectively.
After training a traditional BERT and GAN-BERT on only 109 labeled examples in a classification task involving 50 classes, the following results are obtained:
BERT
eval_accuracy = 0.136 eval_f1_macro = 0.010 eval_f1_micro = 0.136 eval_loss = 3.764 eval_precision = 0.136 eval_recall = 0.136
GAN-BERT
eval_accuracy = 0.418 eval_f1_macro = 0.057 eval_f1_micro = 0.418 eval_loss = 2.744 eval_precision = 0.418 eval_recall = 0.418
As the code is based on the original BERT Tensorflow code and that it starts from the BERT-base model, the same batch size and sequence length restrictions apply here based on the GPU that is used to run an experiment.
Please, refer to the BERT Github page (https://github.com/google-research/bert#out-of-memory-issues) to find the suggested batch size and sequence length given the amount of GPU memory available.
To cite the paper, please use the following:
@inproceedings{croce-etal-2020-gan,
title = "{GAN}-{BERT}: Generative Adversarial Learning for Robust Text Classification with a Bunch of Labeled Examples",
author = "Croce, Danilo and
Castellucci, Giuseppe and
Basili, Roberto",
booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
month = jul,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.acl-main.191",
pages = "2114--2119"
}- Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza,Bing Xu, David Warde-Farley, Sherjil Ozair, AaronCourville and Yoshua Bengio. 2014. Generative Adversarial Nets. In Z. Ghahramani, M. Welling, C. Cortes, N. D. Lawrence, and K. Q. Weinberger, editors, Advances in Neural Information Processing Systems 27, pages 2672–2680. Curran Associates, Inc.
- Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, Xi Chen, and Xi Chen. 2016. Improved techniques for training gans. In D. D. Lee, M. Sugiyama, U. V. Luxburg, I. Guyon, and R. Garnett, editors, Advances in Neural Information Processing Systems 29, pages 2234–2242. Curran Associates, Inc.