UST or Uncertainty-aware Self-Training is a method of task-specific training of pre-trainined language models (e.g., BERT, Electra, GPT) with only a few-labeled examples for the target classification task and large amounts of unlabeled data.
Our academic paper published as a spotlight presentation at NeurIPS 2020 describes the framework in details here: Uncertainy-aware Self-training for Few-shot Text Classification
With only 20-30 labeled examples for each class for each task and large amounts of task-specific unlabeled data, UST performs within 3% accuracy of fully supervised pre-trained language models fine-tuned on thousands of labeled examples with an aggregate accuracy of 91% and improvement of upto 12% over baselines (e.g., BERT) for text classification on benchmark datasets. It does not use any auxiliary resources like paraphrases or backtranslations.
The following table reports text classification results over 5 benchmark datasets averaged over over multiple runs.
| BERT (30 labels) | UDA SSL (30 labels) | Classic ST (30 labels) | UST (30 labels) | BERT (Supervised ~150K labels) | |
|---|---|---|---|---|---|
| SST | 69.79 | 83.58 | 84.81 | 87.69 | 92.12 |
| IMDB | 73.03 | 89.3 | 78.97 | 89.21 | 91.7 |
| Elec | 82.92 | 89.64 | 89.92 | 91.27 | 93.46 |
| AG News | 80.74 | 85.92 | 84.62 | 88.19 | 92.12 |
| Macro Average | 80.85 | 89.06 | 87.34 | 91.00 | 93.73 |
UST is a semi-supervised learning method that leverages pre-trained language models with stochastic regularization techniques and iterative self-training with student-teacher models. Specifically, it extends traditional self-training with three core components, namely:
(i) Masked model dropout for uncertainty estimation. We adopt MC dropout (Gal and Ghahramani, 2016) as a technique to obtain uncertainty estimates from the pre-trained language model. In this, we apply stochastic dropouts after different hidden layers in the neural network model and approximate the model output as a random sample from the posterior distribution. This allows us to compute the model uncertainty in terms of the stochastic mean and variance of the samples with a few stochastic forward passes through the network.
(ii) Sample selection. Given the above uncertainty estimates for a sample, we employ entropy-based measures to select samples that the teacher is most or least confused about to infuse for self-training corresponding to easy- and hard-entropy-aware example mining.
(iii) Confident learning. In this, we train the student model to explicitly account for the teacher confidence by emphasizing on the low variance examples. All of the above components are jointly used for end-to-end learning.
For the few-shot learning setting with limited training labels, continued pre-training on task-specific unlabeled data starting from available pre-trained checkpoints is an effective mechanism to obtain a good base encoder to initialize the teacher model for UST. Code for continued pre-trainining with masked language modeling objective can be found in the original BERT repo here: https://github.com/google-research/bert. This requires invoking the create_pretraining_data.py and the run_pretraining.py scripts from the repo with additional instructions therein. This produces a new tensorflow checkpoint that can be used as the pre-trained checkpoint for UST.
You can use transformers-cli from https://huggingface.co/transformers/converting_tensorflow_models.html to convert tensorflow checkpoints (ckpt) to compatible checkpoints (bin) for HuggingFace transformers.
Note that this continued pre-training step in optional for UST, but required to reproduce the results in the paper. In absence of this step, UST uses the default pre-trained checkpoints for any pre-trained langauge model which also works well in practise. The continued pre-trained checkpoints for the tasks in the paper are available here that can be used to initialize UST for the respective tasks.
UST is integrated with HuggingFace Transformers which makes it possible to use any supported pre-trained language model as a base encoder.
UST requires 3 input files train.tsv and test.tsv with tab-separated (i) instances (e.g., SST and IMDB) or pairs of instances (e.g., MRPC and MNLI) and (ii) labels; and transfer.txt for the unlabeled instances of the corresponding task (all line-separated) in the data directory.
The code has been tested with Tensorflow 2.3.1, Transformers 3.3.1 and Python 3.6.9. Install all the required dependencies with pip install -r requirements.txt.
These are some standard set of arugments to run UST for the few-shot setting. Refer to run_ust.py for all the optional arugments and descriptions.
PYTHONHASHSEED=42 python run_ust.py
--task $DATA_DIR
--model_dir $OUTPUT_DIR
--seq_len 128
--sample_scheme easy_bald_class_conf
--sup_labels 60
--valid_split 0.5
--pt_teacher TFBertModel
--pt_teacher_checkpoint bert-base-uncased
--N_base 5
--sup_batch_size 4
Classification tasks: Set --do_pairwise for pairwise classification tasks like MRPC and MNLI.
Sampling schemes: Supported sample scheme: uniform, easy_bald_class_conf (sampling easy examples with uncertainty given by Eqn. 7 in paper), dif_bald_class_conf (sampling difficult examples given by Eqn. 8). conf enables confident learning, whereas class enables class dependent exploration. Additionally, you can append soft to the above sampling scheme (e.g., easy_bald_class_conf_soft) for leveraging majority predictions from T stochastic forward passes that works well for settings involving many classes / labels.
HuggingFace Transformers: To use different pre-trained language models from HuggingFace, set pt_teacher and pt_teacher_checkpoint to corresponding model versions available from https://huggingface.co/transformers/pretrained_models.html. A default set of pre-trained language models is available at ``huggingface_utils.py`.
Training and validation: sup_labels denote the total number of available labeled examples for each class for each task, where valid_split uses a fraction of those labels as validation set for early stopping. Set sup_labels to -1 to use all training labels. Set valid_split to -1 to use the available test data as the validation set.
Initializing the teacher model: To start with a good base encoder for the few-shot setting with very few labeled examples, UST uses different random seeds to initialize and fine-tune the teacher model N_base times and selects the best one to start the self-training process. This is not required when large number of labeled examples are available (correspondingly set N_base to 1).
Fine-tuning batch size: Set sup_batch_size to a small number for few-shot fine-tuning (e.g., 4) of the teacher model. In case of many training labels, set sup_batch_size to a higher value for faster training (e.g., 32).
Self-training works for both low-data and high-data regime. For example, UST obtains 0.5 accuracy improvement for MNLI (mismatched) using all the available labeled examples (393K) to use for both training as well as the transfer set without using any additional unlabeled data.
Standard set of arugments to run UST with all labeled examples (e.g., MNLI).
PYTHONHASHSEED=42 python run_ust.py
--task $DATA_DIR/MNLI
--model_dir $OUTPUT_DIR
--seq_len 128
--sample_scheme easy_bald_class_conf_soft
--sup_labels -1
--valid_split -1
--sup_batch_size 32
--do_pairwise
--N_base 1
Dropouts are the key for stochastic regularization and obtaining uncertainty estimates. However, too small values lead to less perturbation; whereas, too large values distort the pre-trained model attention mechanism. Good values of dropouts: BERT --hidden_dropout_prob 0.3 --attention_probs_dropout_prob 0.3, Electra/Roberta --hidden_dropout_prob 0.2 --attention_probs_dropout_prob 0.2
Examples of using other pre-trained language models (defined in huggingface_utils.py): Electra --pt_teacher TFElectraModel --pt_teacher_checkpoint google/electra-base-discriminator and Roberta --pt_teacher TFRobertaModel --pt_teacher_checkpoint roberta-base
If you use this code, please cite:
@inproceedings{mukherjee-awadallah-2020-ust,
title = "Uncertainty-aware Self-training for Few-shot Text Classification",
author = "Mukherjee, Subhabrata and
Hassan Awadallah, Ahmed",
booktitle = "Advances in Neural Information Processing Systems (NeurIPS 2020)",
year = "2020",
address = "Online",
url = "https://papers.nips.cc/paper/2020/file/f23d125da1e29e34c552f448610ff25f-Paper.pdf",
}
Code is released under MIT license.
![microsoft-github-operations[bot] avatar](https://avatars.githubusercontent.com/in/41902?v=4 "microsoft-github-operations[bot]")
![microsoft-github-policy-service[bot] avatar](https://avatars.githubusercontent.com/in/95686?v=4 "microsoft-github-policy-service[bot]")
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent