One2Set
This repository contains the code for our ACL 2021 paper “One2Set: Generating Diverse Keyphrases as a Set”.
Our implementation is built on the source code from keyphrase-generation-rl and fastNLP. Thanks for their work.
If you use this code, please cite our paper:
@inproceedings{ye2021one2set,
title={One2Set: Generating Diverse Keyphrases as a Set},
author={Ye, Jiacheng and Gui, Tao and Luo, Yichao and Xu, Yige and Zhang, Qi},
booktitle={Proceedings of ACL},
year={2021}
}
Dependency
- python 3.5+
- pytorch 1.0+
Dataset
The datasets can be downloaded from here, which are the tokenized version of the datasets provided by Ken Chen:
- The
testsetsdirectory contains the five datasets for testing (i.e., inspec, krapivin, nus, and semeval and kp20k), where each of the datasets containstest_src.txtandtest_trg.txt. - The
kp20k_separateddirectory contains the training and validation files (i.e.,train_src.txt,train_trg.txt,valid_src.txtandvalid_trg.txt). - Each line of the
*_src.txtfile is the source document, which contains the tokenized words oftitle <eos> abstract. - Each line of the
*_trg.txtfile contains the target keyphrases separated by an;character. The<peos>is used to mark the end of present ground-truth keyphrases and train a separate set loss forSetTransmodel. For example, each line can be likepresent keyphrase one;present keyphrase two;<peos>;absent keyprhase one;absent keyphrase two.
Quick Start
The whole process includes the following steps:
Preprocessing: Thepreprocess.pyscript numericalizes thetrain_src.txt,train_trg.txt,valid_src.txtandvalid_trg.txtfiles, and producestrain.one2many.pt,valid.one2many.ptandvocab.pt.Training: Thetrain.pyscript loads thetrain.one2many.pt,valid.one2many.ptandvocab.ptfile and performs training. We evaluate the model every 8000 batches on the valid set, and the model will be saved if the valid loss is lower than the previous one.Decoding: Thepredict.pyscript loads the trained model and performs decoding on the five test datasets. The prediction file will be saved, which is likepredicted keyphrase one;predicted keyphrase two;…. ForSetTrans, we ignore the $\varnothing$ predictions that represent the meaning of “no corresponding keyphrase”.Evaluation: Theevaluate_prediction.pyscript loads the ground-truth and predicted keyphrases, and calculates the $F_1@5$ and $F_1@M$ metrics.
For the sake of simplicity, we provide an one-click script in the script directory. You can run the following command to run the whole process with SetTrans model under One2Set paradigm:
bash scripts/run_one2set.sh
You can also run the baseline Transformer model under One2Seq paradigm with the following command:
bash scripts/run_one2seq.sh
Note:
- Please download and unzip the datasets in the
./datadirectory first. - To run all the bash files smoothly, you may need to specify the correct
home_dir(i.e., the absolute path tokg_one2setdictionary) and the gpu id forCUDA_VISIBLE_DEVICES. We provide a small amount of data to quickly test whether your running environment is correct. You can test by running the following command:
bash scripts/run_small_one2set.sh
Resources
You can download our trained model here. We also provide raw predictions and corresponding evaluation results of three runs with different random seeds here, which contains the following files:
test
├── Full_One2set_Copy_Seed27_Dropout0.1_LR0.0001_BS12_MaxLen6_MaxNum20_LossScalePre0.2_LossScaleAb0.1_Step2_SetLoss
│ ├── inspec
│ │ ├── predictions.txt
│ │ └── results_log_5_M_5_M_5_M.txt
│ ├── kp20k
│ │ ├── predictions.txt
│ │ └── results_log_5_M_5_M_5_M.txt
│ ├── krapivin
│ │ ├── predictions.txt
│ │ └── results_log_5_M_5_M_5_M.txt
│ ├── nus
│ │ ├── predictions.txt
│ │ └── results_log_5_M_5_M_5_M.txt
│ └── semeval
│ ├── predictions.txt
│ └── results_log_5_M_5_M_5_M.txt
├── Full_One2set_Copy_Seed527_Dropout0.1_LR0.0001_BS12_MaxLen6_MaxNum20_LossScalePre0.2_LossScaleAb0.1_Step2_SetLoss
│ ├── ...
└── Full_One2set_Copy_Seed9527_Dropout0.1_LR0.0001_BS12_MaxLen6_MaxNum20_LossScalePre0.2_LossScaleAb0.1_Step2_SetLoss
├── ...
21 May 18, 2022
34 Dec 21, 2022
![[NAACL & ACL 2021] SapBERT: Self-alignment pretraining for BERT.](https://github.com/cambridgeltl/sapbert/raw/main/sapbert_front_graphs_v6.png?raw=true)
104 Dec 7, 2022
51 Oct 8, 2022
25 Oct 30, 2022
14 Aug 24, 2022
52 Dec 27, 2022
31 Oct 25, 2022
71 Dec 8, 2022
![a bug about [dec_layers]](https://avatars.githubusercontent.com/u/13014170?v=4)
150 Dec 12, 2022
5 Jun 16, 2022
68 Dec 15, 2022
190 Jan 3, 2023
81 Dec 27, 2022
75 Nov 6, 2022
47 Jan 1, 2023
22 Dec 12, 2022
124 Dec 27, 2022
10 Dec 6, 2022