Code and datasets for the WWW2022 paper KnowPrompt: Knowledge-aware Prompt-tuning with Synergistic Optimization for Relation Extraction.

• ❗NOTE: We provide a paper-list at PromptKG and open-source KnowLM, a knowledgeable large language model framework with pre-training and instruction fine-tuning code (supports multi-machine multi-GPU setup).

• Our follow-up paper on schema-aware prompting KGC Schema-aware Reference as Prompt Improves Data-Efficient Knowledge Graph Construction has been accepted by SIGIR2023. The project address is visible at RAP.

• We release a checkpoint for SemEval dataset at huggingface model hub

• Our follow-up paper on retrieval-augmented prompting NLP Decoupling Knowledge from Memorization: Retrieval-augmented Prompt Learning has been accepted by NeurIPS2022. The project address is visible at RetroPrompt.

• Our follow-up paper on retrieval-enhanced prompting RE Relation Extraction as Open-book Examination: Retrieval-enhanced Prompt Tuning has been accepted by SIGIR2022. The project address is visible at RetrievalRE.

• Our paper KnowPrompt: Knowledge-aware Prompt-tuning with Synergistic Optimization for Relation Extraction has been accepted by WWW2022.

It is recommended to use a virtual environment to run KnowPrompt.

conda create -n knowprompt python=3.8

conda activate knowprompt

To install requirements:

pip install -r requirements.txt

We provide all the datasets and prompts used in our experiments.

• [SEMEVAL]

• [DialogRE]

• [TACRED-Revisit]

• [Re-TACRED]

• [TACRED]

The expected structure of files is:

knowprompt
 |-- dataset
 |    |-- semeval
 |    |    |-- train.txt       
 |    |    |-- dev.txt
 |    |    |-- test.txt
 |    |    |-- temp.txt
 |    |    |-- rel2id.json
 |    |-- dialogue
 |    |    |-- train.json       
 |    |    |-- dev.json
 |    |    |-- test.json
 |    |    |-- rel2id.json
 |    |-- tacred
 |    |    |-- train.txt       
 |    |    |-- dev.txt
 |    |    |-- test.txt
 |    |    |-- temp.txt
 |    |    |-- rel2id.json
 |    |-- tacrev
 |    |    |-- train.txt       
 |    |    |-- dev.txt
 |    |    |-- test.txt
 |    |    |-- temp.txt
 |    |    |-- rel2id.json
 |    |-- retacred
 |    |    |-- train.txt       
 |    |    |-- dev.txt
 |    |    |-- test.txt
 |    |    |-- temp.txt
 |    |    |-- rel2id.json
 |-- scripts
 |    |-- semeval.sh
 |    |-- dialogue.sh
 |    |-- ...
 

Use the comand below to get the answer words to use in the training.

python get_label_word.py --model_name_or_path bert-large-uncased  --dataset_name semeval

The {answer_words}.ptwill be saved in the dataset, you need to assign the model_name_or_path and dataset_name in the get_label_word.py.

Download the data first, and put it to dataset folder. Run the comand below, and get the few shot dataset.

python generate_k_shot.py --data_dir ./dataset --k 8 --dataset semeval
cd dataset
cd semeval
cp rel2id.json val.txt test.txt ./k-shot/8-1

You need to modify the k and dataset to assign k-shot and dataset. Here we default seed as 1,2,3,4,5 to split each k-shot, you can revise it in the generate_k_shot.py

Our script code can automatically run the experiments in 8-shot, 16-shot, 32-shot and standard supervised settings with both the procedures of train, eval and test. We just choose the random seed to be 1 as an example in our code. Actually you can perform multiple experments with different seeds.

Train the KonwPrompt model on SEMEVAL with the following command:

>> bash scripts/semeval.sh  # for roberta-large

As the scripts for TACRED-Revist, Re-TACRED, Wiki80 included in our paper are also provided, you just need to run it like above example.

As the data format of DialogRE is very different from other dataset, Class of processor is also different. Train the KonwPrompt model on DialogRE with the following command:

>> bash scripts/dialogue.sh  # for roberta-base

We report emperical results on more datasets in the EMNLP 2022 (Findings) paper "Towards Realistic Low-resource Relation Extraction: A Benchmark with Empirical Baseline Study" [code].

Part of our code is borrowed from code of PTR: Prompt Tuning with Rules for Text Classification, many thanks.

If you use the code, please cite the following paper:

@inproceedings{DBLP:conf/www/ChenZXDYTHSC22,
  author    = {Xiang Chen and
               Ningyu Zhang and
               Xin Xie and
               Shumin Deng and
               Yunzhi Yao and
               Chuanqi Tan and
               Fei Huang and
               Luo Si and
               Huajun Chen},
  editor    = {Fr{\'{e}}d{\'{e}}rique Laforest and
               Rapha{\"{e}}l Troncy and
               Elena Simperl and
               Deepak Agarwal and
               Aristides Gionis and
               Ivan Herman and
               Lionel M{\'{e}}dini},
  title     = {KnowPrompt: Knowledge-aware Prompt-tuning with Synergistic Optimization
               for Relation Extraction},
  booktitle = {{WWW} '22: The {ACM} Web Conference 2022, Virtual Event, Lyon, France,
               April 25 - 29, 2022},
  pages     = {2778--2788},
  publisher = {{ACM}},
  year      = {2022},
  url       = {https://doi.org/10.1145/3485447.3511998},
  doi       = {10.1145/3485447.3511998},
  timestamp = {Tue, 26 Apr 2022 16:02:09 +0200},
  biburl    = {https://dblp.org/rec/conf/www/ChenZXDYTHSC22.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}