Deep Keyphrase Extraction using BERT

I have used BERT Token Classification Model to extract keywords from a sentence. Feel free to clone and use it. If you face any problems, kindly post it on issues section.

Special credits to BERT authors: Jacob Devlin, Ming-Wei Chang, Kenton Lee, Kristina Toutanova, original repo and Huggingface for PyTorch version original repo.

Requirements

You need:

pytorch 1.0
python 3.6
pytorch-pretrained-bert 0.4.0

Usage

The keyword-extractor.py script can be used to extract keywords from a sentence and accepts the following arguments:

optional arguments:
  -h, --help         show this help message and exit
  --sentence SEN        sentence to extract keywords
  --path LOAD        path to load model from

Example:

python keyword-extractor.py --sentence "BERT is a great model." --path "model.pt"           

Training

You can also train it from scratch using BERT's pre-trained model. The main.py script can be utilized for training and accepts the following arguments:

optional arguments:
  -h, --help         show this help message and exit
  --data DATA        location of the data corpus
  --lr LR            initial learning rate
  --epochs EPOCHS    upper epoch limit
  --batch_size N     batch size
  --seq_len N        sequence length
  --save SAVE        path to save the final model

Example:

python main.py --data "maui-semeval2010-train" --lr 2e-5 --batch_size 32 --save "model.pt" --epochs 3      

This model has been trained on SemEval 2010 dataset (scientific publications). You can swap this with your own custom dataset.

Code explanations

I have provided the explanation of keyphrase extraction in the form of python notebook which you can view here

Hyper-parameter Tuning

I ran ablation experiments according to the BERT paper and these are the results. I suggest to use parameters in line 4. All training was done on batch size of 32.

Learning Rate	Number of Epochs	Validation loss	Validation Accuracy	F1-Score
3.00E-05	3	0.05294724515	98.30%	0.5318559557
5.00E-05	3	0.04899719357	98.47%	0.56218628
2.00E-05	3	0.05733459462	98.15%	0.4390547264
3.00E-05	4	0.05020467712	98.48%	0.5528169014
5.00E-05	4	0.05194576555	98.43%	0.5780836421
2.00E-05	4	0.05373481681	98.25%	0.5019740553