whaleloops/phrase-bert

This is the official repository for the EMNLP 2021 long paper Phrase-BERT: Improved Phrase Embeddings from BERT with an Application to Corpus Exploration. We provide code for training and evaluating Phrase-BERT in addition to the datasets used in the paper.

Usage (Sentence-Transformers)

Using this model becomes easy when you have sentence-transformers installed:

pip install -U sentence-transformers

Our model is tested on pytorch=1.9.0, tranformers=4.8.1, sentence-tranformers = 2.1.0 TODO

Then you can use the model like this:

from sentence_transformers import SentenceTransformer
phrase_list = [ 'play an active role', 'participate actively', 'active lifestyle']

model = SentenceTransformer('whaleloops/phrase-bert')
phrase_embs = model.encode( phrase_list )
[p1, p2, p3] = phrase_embs

As in sentence-BERT, the default output is a list of numpy arrays:

for phrase, embedding in zip(phrase_list, phrase_embs):
    print("Phrase:", phrase)
    print("Embedding:", embedding)
    print("")

An example of computing the dot product of phrase embeddings:

import numpy as np
print(f'The dot product between phrase 1 and 2 is: {np.dot(p1, p2)}')
print(f'The dot product between phrase 1 and 3 is: {np.dot(p1, p3)}')
print(f'The dot product between phrase 2 and 3 is: {np.dot(p2, p3)}')

An example of computing cosine similarity of phrase embeddings:

import torch 
from torch import nn
cos_sim = nn.CosineSimilarity(dim=0)
print(f'The cosine similarity between phrase 1 and 2 is: {cos_sim( torch.tensor(p1), torch.tensor(p2))}')
print(f'The cosine similarity between phrase 1 and 3 is: {cos_sim( torch.tensor(p1), torch.tensor(p3))}')
print(f'The cosine similarity between phrase 2 and 3 is: {cos_sim( torch.tensor(p2), torch.tensor(p3))}')

The output should look like:

The dot product between phrase 1 and 2 is: 218.43600463867188
The dot product between phrase 1 and 3 is: 165.48483276367188
The dot product between phrase 2 and 3 is: 160.51708984375
The cosine similarity between phrase 1 and 2 is: 0.8142536282539368
The cosine similarity between phrase 1 and 3 is: 0.6130303144454956
The cosine similarity between phrase 2 and 3 is: 0.584893524646759

Evaluation

Given the lack of a unified phrase embedding evaluation benchmark, we collect the following five phrase semantics evaluation tasks, which are described further in our paper:

• Turney [Download ]
• BiRD [Download]
• PPDB [Download]
• PPDB-filtered [Download]
• PAWS-short [Download Train-split ] [Download Dev-split ] [Download Test-split ]

Change config/model_path.py with the model path according to your directories and

• For evaluation on Turney, run python eval_turney.py

• For evaluation on BiRD, run python eval_bird.py

• for evaluation on PPDB / PPDB-filtered / PAWS-short, run eval_ppdb_paws.py with:

  nohup python  -u eval_ppdb_paws.py \
      --full_run_mode \
      --task <task-name> \
      --data_dir <input-data-dir> \
      --result_dir <result-storage-dr> \
      >./output.txt 2>&1 &
  

Train your own Phrase-BERT

If you would like to go beyond using the pre-trained Phrase-BERT model, you may train your own Phrase-BERT using data from the domain you are interested in. Please refer to phrase-bert/phrase_bert_finetune.py

The datasets we used to fine-tune Phrase-BERT are here: training data csv file and validation data csv file.

To re-produce the trained Phrase-BERT, please run:

export INPUT_DATA_PATH=<directory-of-phrasebert-finetuning-data>
export TRAIN_DATA_FILE=<training-data-filename.csv>
export VALID_DATA_FILE=<validation-data-filename.csv>
export INPUT_MODEL_PATH=bert-base-nli-stsb-mean-tokens 
export OUTPUT_MODEL_PATH=<directory-of-saved-model>


python -u phrase_bert_finetune.py \
    --input_data_path $INPUT_DATA_PATH \
    --train_data_file $TRAIN_DATA_FILE \
    --valid_data_file $VALID_DATA_FILE \
    --input_model_path $INPUT_MODEL_PATH \
    --output_model_path $OUTPUT_MODEL_PATH

Citation:

Please cite us if you find this useful:

@inproceedings{phrasebertwang2021,
    author={Shufan Wang and Laure Thompson and Mohit Iyyer},
    Booktitle = {Empirical Methods in Natural Language Processing},
    Year = "2021",
    Title={Phrase-BERT: Improved Phrase Embeddings from BERT with an Application to Corpus Exploration}
}