Skip to content

README.md

Latest commit

 

History

History
72 lines (54 loc) · 2.01 KB

README.md

File metadata and controls

72 lines (54 loc) · 2.01 KB

PromptRank

This is code of our paper PromptRank: Unsupervised Keyphrase Extraction using Prompt.

Environment

StanfordCoreNLP 3.9.1.1
Python 3.8
torch 1.9.1
nltk 3.7
transformers 4.23

Usage

First, you should download Stanford CoreNLP and modify the file path in data.py.

StanfordCoreNLP_path = '../../stanford-corenlp-full-2018-02-27'

We use run.sh script to run PromptRank for a specfic dataset.

bash run.sh

We use run_all.sh script to run PromptRank for all six datasets and the results will be summarized in result.txt by summary.py.

bash run_all.sh

The settings of PromptRank can be modified in main.py as follows:

def get_setting_dict():
    setting_dict = {}
    setting_dict["max_len"] = 512
    setting_dict["temp_en"] = "Book:"
    setting_dict["temp_de"] = "This book mainly talks about "
    setting_dict["model"] = "base"
    setting_dict["enable_filter"] = False
    setting_dict["enable_pos"] = True
    setting_dict["position_factor"] = 1.2e8
    setting_dict["length_factor"] = 0.6
    return setting_dict

The settings of running can be modified in run.sh or run_all.sh.

Performance

The performance of PromptRank is shown as follows. See the performance of ${\rm PromptRank}_{filter}$ and other baselines in the paper.

F1@K Inspec SemEval2017 SemEval2010 DUC2001 Krapivin NUS AVG
5 31.73 27.14 17.24 27.39 16.11 17.24 22.81
10 37.88 37.76 20.66 31.59 16.71 20.13 27.46
15 38.17 41.57 21.35 31.01 16.02 20.12 28.04

Citation

@article{kong2023promptrank,
  title={PromptRank: Unsupervised Keyphrase Extraction Using Prompt},
  author={Kong, Aobo and Zhao, Shiwan and Chen, Hao and Li, Qicheng and Qin, Yong and Sun, Ruiqi and Bai, Xiaoyan},
  journal={arXiv preprint arXiv:2305.04490},
  year={2023}
}