Unsupervised Keyphrase Extraction

This is code for EMNLP 2021 paper: Unsupervised Keyphrase Extraction by Jointly Modeling Local and Global Context.

Update 2022-09-02: We release the data preprocess script and chinese keyphrase extraction code in https://github.com/xnliang98/CKE-ZH.

requirements

• transformers==3.0.2
• nltk
• pytorch
• tqdm

We employ StanfordCoreNLP Tools to preprocess the data.

Runing

Step 1: obtain embeddings of candidate phrases and the whole document.

python src/get_embedding.py --file_path [data_path] --file_name [file_name] --model_name [pretrained model name/path]

Step 2: extract keyphrases

python src/ranker.py [data_path] [model_name]

Comments & TODO

The middle layer representation of BERT model may get better performance.