On the Sentence Embeddings from Pre-trained Language Models

This is a TensorFlow implementation of the following paper:

On the Sentence Embeddings from Pre-trained Language Models
Bohan Li, Hao Zhou, Junxian He, Mingxuan Wang, Yiming Yang, Lei Li
EMNLP 2020

Model	Spearman's rho
BERT-large-NLI	77.80
BERT-large-NLI-last2avg	78.45
BERT-large-NLI-flow (target, train only)	80.54
BERT-large-NLI-flow (target, train+dev+test)	81.18

Please contact bohanl1@cs.cmu.edu if you have any questions.

Requirements

Python >= 3.6
TensorFlow >= 1.14

Preparation

Pretrained BERT models

export BERT_PREMODELS="../bert_premodels"
mkdir ${BERT_PREMODELS}; cd ${BERT_PREMODELS}

# then download the pre-trained BERT models from https://github.com/google-research/bert
curl -O https://storage.googleapis.com/bert_models/2018_10_18/uncased_L-12_H-768_A-12.zip
curl -O https://storage.googleapis.com/bert_models/2018_10_18/uncased_L-24_H-1024_A-16.zip

ls ${BERT_PREMODELS}/uncased_L-12_H-768_A-12 # base
ls ${BERT_PREMODELS}/uncased_L-24_H-1024_A-16 # large

GLUE

export GLUE_DIR="../glue_data"
python download_glue_data.py --data_dir=${GLUE_DIR}

# then download the labeled test set of STS-B
cd ../glue_data/STS-B
curl -O https://raw.githubusercontent.com/kawine/usif/master/STSBenchmark/sts-test.csv

SentEval

cd ..
git clone https://github.com/facebookresearch/SentEval

Usage

Fine-tune BERT with NLI supervision (optional)

export OUTPUT_PARENT_DIR="../exp"
export CACHED_DIR=${OUTPUT_PARENT_DIR}/cached_data
mkdir ${CACHED_DIR}

export RANDOM_SEED=1234
export CUDA_VISIBLE_DEVICES=0
export BERT_NAME="large"
export TASK_NAME="ALLNLI"
unset INIT_CKPT
bash scripts/train_siamese.sh train \
"--exp_name=exp_${BERT_NAME}_${RANDOM_SEED} \
--num_train_epochs=1.0 \
--learning_rate=2e-5 \
--train_batch_size=16 \
--cached_dir=${CACHED_DIR}"


# evaluation
export RANDOM_SEED=1234
export CUDA_VISIBLE_DEVICES=0
export TASK_NAME=STS-B
export BERT_NAME=large
export OUTPUT_PARENT_DIR="../exp"
export INIT_CKPT=${OUTPUT_PARENT_DIR}/exp_${BERT_NAME}_${RANDOM_SEED}/model.ckpt-60108
export CACHED_DIR=${OUTPUT_PARENT_DIR}/cached_data
export EXP_NAME=exp_${BERT_NAME}_${RANDOM_SEED}_eval
bash scripts/train_siamese.sh predict \
"--exp_name=${EXP_NAME} \
 --cached_dir=${CACHED_DIR} \
 --sentence_embedding_type=avg \
 --flow=0 --flow_loss=0 \
 --num_examples=0 \
 --num_train_epochs=1e-10"

Note: You may want to add --use_xla to speed up the BERT fine-tuning.

Unsupervised learning of flow-based generative models

export CUDA_VISIBLE_DEVICES=0
export TASK_NAME=STS-B
export BERT_NAME=large
export OUTPUT_PARENT_DIR="../exp"
export INIT_CKPT=${OUTPUT_PARENT_DIR}/exp_large_1234/model.ckpt-60108
export CACHED_DIR=${OUTPUT_PARENT_DIR}/cached_data
bash scripts/train_siamese.sh train \
"--exp_name_prefix=exp \
 --cached_dir=${CACHED_DIR} \
 --sentence_embedding_type=avg-last-2 \
 --flow=1 --flow_loss=1 \
 --num_examples=0 \
 --num_train_epochs=1.0 \
 --flow_learning_rate=1e-3 \
 --use_full_for_training=1"

# evaluation
export CUDA_VISIBLE_DEVICES=0
export TASK_NAME=STS-B
export BERT_NAME=large
export OUTPUT_PARENT_DIR="../exp"
export INIT_CKPT=${OUTPUT_PARENT_DIR}/exp_large_1234/model.ckpt-60108
export CACHED_DIR=${OUTPUT_PARENT_DIR}/cached_data
export EXP_NAME=exp_t_STS-B_ep_1.00_lr_5.00e-05_e_avg-last-2_f_11_1.00e-03_allsplits
bash scripts/train_siamese.sh predict \
"--exp_name=${EXP_NAME} \
 --cached_dir=${CACHED_DIR} \
 --sentence_embedding_type=avg-last-2 \
 --flow=1 --flow_loss=1 \
 --num_examples=0 \
 --num_train_epochs=1.0 \
 --flow_learning_rate=1e-3 \
 --use_full_for_training=1"

Fit flow with only the training set of STS-B

export CUDA_VISIBLE_DEVICES=0
export TASK_NAME=STS-B
export BERT_NAME=large
export OUTPUT_PARENT_DIR="../exp"
export INIT_CKPT=${OUTPUT_PARENT_DIR}/exp_large_1234/model.ckpt-60108
export CACHED_DIR=${OUTPUT_PARENT_DIR}/cached_data
bash scripts/train_siamese.sh train \
"--exp_name_prefix=exp \
 --cached_dir=${CACHED_DIR} \
 --sentence_embedding_type=avg-last-2 \
 --flow=1 --flow_loss=1 \
 --num_examples=0 \
 --num_train_epochs=1.0 \
 --flow_learning_rate=1e-3 \
 --use_full_for_training=0"

# evaluation
export CUDA_VISIBLE_DEVICES=0
export TASK_NAME=STS-B
export BERT_NAME=large
export OUTPUT_PARENT_DIR="../exp"
export INIT_CKPT=${OUTPUT_PARENT_DIR}/exp_large_1234/model.ckpt-60108
export CACHED_DIR=${OUTPUT_PARENT_DIR}/cached_data
export EXP_NAME=exp_t_STS-B_ep_1.00_lr_5.00e-05_e_avg-last-2_f_11_1.00e-03
bash scripts/train_siamese.sh predict \
"--exp_name=${EXP_NAME} \
 --cached_dir=${CACHED_DIR} \
 --sentence_embedding_type=avg-last-2 \
 --flow=1 --flow_loss=1 \
 --num_examples=0 \
 --num_train_epochs=1.0 \
 --flow_learning_rate=1e-3 \
 --use_full_for_training=1"

Download our models

Our models are available at https://drive.google.com/file/d/1-vO47t5SPFfzZPKkkhSe4tXhn8u--KLR/view?usp=sharing

Reference

@inproceedings{li2020emnlp,
    title = {On the Sentence Embeddings from Pre-trained Language Models},
    author = {Bohan Li and Hao Zhou and Junxian He and Mingxuan Wang and Yiming Yang and Lei Li},
    booktitle = {Conference on Empirical Methods in Natural Language Processing (EMNLP)},
    month = {November},
    year = {2020}
}

Acknowledgements

A large portion of this repo is borrowed from the following projects:

bohanli/BERT-flow