BERT for Keyphrase Extraction (PyTorch)

This repository provides the code of the paper Capturing Global Informativeness in Open Domain Keyphrase Extraction.

In this paper, we conduct an empirical study of 5 keyphrase extraction models with 3 BERT variants, and then propose a multi-task model BERT-JointKPE. Experiments on two KPE benchmarks, OpenKP with Bing web pages and KP20K demonstrate JointKPE’s state-of-the-art and robust effectiveness. Our further analyses also show that JointKPE has advantages in predicting long keyphrases and non-entity keyphrases, which were challenging for previous KPE techniques.

Please cite our paper if our experimental results, analysis conclusions or the code are helpful to you ~ 😊

@article{sun2020joint,
    title={Joint Keyphrase Chunking and Salience Ranking with BERT},
    author={Si Sun, Zhenghao Liu, Chenyan Xiong, Zhiyuan Liu and Jie Bao},
    year={2020},
    eprint={2004.13639},
    archivePrefix={arXiv},
    primaryClass={cs.CL}
}

CONTACT

For any question, feel free to create an issue, and we will try our best to solve.
If the problem is more urgent, you can send an email to me at the same time (I check email almost everyday 😉).

NAME: Si Sun
EMAIL: [email protected]

🤠 What's New ?

• 2020/9/5

  Compared with the OpenKP dataset we downloaded from MS MARCO in October of 2019 (all our experiments are based on this version of the dataset), we found that the dataset has been updated. We remind you to download the latest data from the official website. For comparison, we also provide the data version we use. (The dataset version issue was raised by Yansen Wang et al from CMU, thank them ! )

  ~DownLoad from Here or ~Email [email protected] for Data

• 2021/12/7

  Our repo now adds Multilingual-KPE and FP16 Training Mode. Thanks, Amit Chaulwar! Amit also shared their zero-shot results on the Wikinews (French), Cacic (Spanish), Pak2018 (Polish), wicc (spanish), 110-PT-BN-KP (Portugese).

Spported Model Classes

BERT Variants Tested

• BERT
• SpanBERT
• RoBERTa
• DistilBERT

Requirements

python 3.8
pytorch 1.9.0
pip install -r pip-requirements.txt

QUICKSTART

1/ Download

• First download and decompress our data folder to this repo, the folder includes benchmark datasets and pre-trained BERT variants.

  • Data Download Link

• We also provide 15 checkpoints (5 KPE models * 3 BERT variants) trained on OpenKP training dataset.

  • Checkpoint Download Link

2/ Preprocess

• To preprocess the source datasets using preprocess.sh in the preprocess folder:

  source preprocess.sh
  

• Optional arguments:

  --dataset_class         choices=['openkp', 'kp20k', 'multidata]
  --source_dataset_dir    The path to the source dataset
  --output_path           The dir to save preprocess data; default: ../data/prepro_dataset
  

• To preprocess the multilingual dataset, download respective datasets from https://github.com/LIAAD/KeywordExtractor-Datasets and use scripts jsonify_multidata.py to preprocess the datasets. The dataset can be split into train, dev, and test sets using split_json.py.

3/ Train Models

• To train a new model from scratch using train.sh in the scripts folder:

  source train.sh
  

  PS. Running the training script for the first time will take some time to perform preprocess such as tokenization, and by default, the processed features will be saved under ../data/cached_features, which can be directly loaded next time.

• Optional arguments:

  --dataset_class         choices=['openkp', 'kp20k', 'multidata']
  --model_class           choices=['bert2span', 'bert2tag', 'bert2chunk', 'bert2rank', 'bert2joint']
  --pretrain_model_type   choices=['bert-base-cased', 'spanbert-base-cased', 'roberta-base', 'distilbert-base-cased']
  

  Complete optional arguments can be seen in config.py in the scripts folder.

• Training Parameters:

  We always keep the following settings in all our experiments:

  args.warmup_proportion = 0.1
  args.max_train_steps = 20810 (openkp) , 73430 (kp20k)
  args.per_gpu_train_batch_size * max(1, args.n_gpu) * args.gradient_accumulation_steps = 64
  

• Distributed Training

  We recommend using DistributedDataParallel to train models on multiple GPUs (It's faster than DataParallel, but it will take up more memory)

  CUDA_VISIBLE_DEVICES=0,1 OMP_NUM_THREADS=2 python -m torch.distributed.launch --nproc_per_node=2 --master_port=1234 train.py
  # if you use DataParallel rather than DistributedDataParallel, remember to set --local_rank=-1
  

4/ Inference

• To evaluate models using trained checkpoints using test.sh in the scripts folder:

  source test.sh
  

• Optional arguments:

  --dataset_class         choices=['openkp', 'kp20k', 'multidata']
  --model_class           choices=['bert2span', 'bert2tag', 'bert2chunk', 'bert2rank', 'bert2joint']
  --pretrain_model_type   choices=['bert-base-cased', 'spanbert-base-cased', 'roberta-base', 'distilbert-base-cased']
  --eval_checkpoint       The checkpoint file to be evaluated
  

5/ Re-produce evaluation results using our checkpoints

• Run test.sh, and change the eval_checkpoint to the checkpoint files we provided to reproduce the following results.

  --dataset_class         openkp
  --eval_checkpoint       The filepath of our provided checkpoint
  

* RESULTS

The following results are ranked by F1@3 on OpenKP Dev dataset, the eval results can be seen in the OpenKP Leaderboard.

* BERT (Base)

* SpanBERT (Base)

* RoBERTa (Base)

MODEL OVERVIEW

* BERT-JointKPE, RankKPE, ChunkKPE (See Paper)

* BERT-TagKPE (See Code)

• Word-Level Representations : We encode an input document into a sequence of WordPiece tokens' vectors with a pretrained BERT (or its variants), and then we pick up the first sub-token vector of each word to represent the input in word-level.

• Phrase-Level Representations : We perform a soft-select method to decode phrase from word-level vector instead of hard-select used in the standard sequence tagging task .

  The word-level representation is feed into an classification layer to obtain the tag probabilities of each word on 5 classes (O, B, I, E, U) , and then we employ different tag patterns for extracting different n-grams ( 1 ≤ n ≤ 5 ) over the whole sequence.

  Last there are a collect of n-gram candidates, each word of the n-gram just has one score.

  Soft-select Example : considering all 3-grams (B I E) on the L-length document, we can extract (L-3+1) 3-grams sequentially like sliding window. In each 3-gram, we only keep B score for the first word, I score for the middle word, and E score for the last word, etc.

  O : Non Keyphrase ; B : Begin word of the keyprase ; I : Middle word of the keyphrase ; E : End word of keyprhase ; U : Uni-word keyphrase

• Document-Level Keyphrase : At the Last stage, the recovering from phrase-level n-grams to document-level keyphrases can be naturally formulated as a ranking task.

  Incorporating with term frequency, we employ Min Pooling to get the final score of each n-gram (we called it Buckets Effect: No matter how high a bucket, it depends on the height of the water in which the lowest piece of wood) . Based on the final scores, we extract 5 top ranked keyprhase candidates for each document.

* BERT-SpanKPE (See Code)

• Word-Level Representations : Same as BERT-TagKPE

• Phrase-Level Representations : Traditional span extraction model could not extract multiple important keyphrase spans for the same document. Therefore, we propose an self-attention span extraction model.

  Given the token representations {t1, t2, ..., tn}, we first calculate the probability that the token is the starting word Ps(ti), and then apply the single-head self-attention layer to calculate the ending word probability of all j>=i tokens Pe(tj).

• Document-Level Keyphrase : We select the spans with the highest probability P = Ps(ti) * Pe(tj) as the keyphrase spans.