The code is the PyTorch reimplement and development of the paper "A Novel Cascade Binary Tagging Framework for Relational Triple Extraction" ACL2020. The official code was written in Keras.

I followed the previous work of Onion12138, longlongman and JuliaSun623, so I have to express sincere thanks to them.

Some changes are made in order to better apply to the Chinese Dataset and put the model into use. The changes contain:

• I used the tokenizer BertTokenizer, so Chinese sentences are tokenized by single character.
• I substituted the original pretrained model with 'bert-wwm-chinese'.
• I used fastNLP to build the datasets.
• I changed the encoding and decoding methods in order to fit the Chinese Dataset.
• I reconstructed the structure for readability.
• I added checkpoint saving for epoch with best performance.
• I implemented the Evaluation and Prediction interface for direct use.

• torch==1.8.0+cu111 (In fact there is no need to use such high edition, just choose according to the cudn of your machine)
• transformers==4.3.3
• fastNLP==0.6.0
• tqdm==4.59.0
• numpy==1.20.1

• conda create -n casrel python==3.7
• pip install torch==1.4.0 torchvision==0.5.0 -f https://download.pytorch.org/whl/cu100/torch_stable.html
• pip install transformers==4.3.3
• pip install fastNLP==0.6.0
• pip install tqdm==4.59.0

The training process only use the 'train' dataset, and evaluation/prediction process use only one dataset that you choose. ('test' dataset for the suggested one)

And then 'dev' dataset is for no use since callback does not influence the training process at all

I preprocessed the open-source dataset from Baidu. I did some cleaning, and the data given have 18 relation types. Some noisy data are eliminated.

The data are in form of json. Take one as an example:

{
    "text": "陶喆的一首《好好说再见》推荐给大家，希望你们能够喜欢",
    "spo_list": [
        {
            "predicate": "歌手",
            "object_type": "人物",
            "subject_type": "歌曲",
            "object": "陶喆",
            "subject": "好好说再见"
        }
    ]
}

In fact the field 'object_type' and 'subject_type' are not used.

Some basic statistics:

Baidu dataset is actually a part of DuIE, so if you want to get a better model, choose this larger dataset for training. The data given have 48 relation types. The data structure is the same as that in Baidu dataset.

I also did some statistics:

If you have your own data, you can organize your data in the same format.

keep encoding='utf-8' for best

python Run.py

I have already set the default value of the model, but you can still set your own configuration in model/config.py and arguments when running Run.py

python Evaluation.py

I have set five kinds of computing method for triples(s, r, o), relation, entity(s, o), subject and object, respectively. Choose one or more methods to evaluate a trained model.

python Prediction.py

Prediction is a basically implemented interface. Input a single sentence, and the corresponding output will be in form of json:

{
	"entity": ["ent1", "ent2", "ent3", "ent4"], 
	"relation": [
			["ent1", "ent2"],
			["ent3", "ent4"],
			["ent2", "ent1"]
		],
	"relation_type": ["rel1", "rel2", "rel3"]
}

Evaluation results for 15 epochs:

It may not reach its utmost since there is only 30 epochs.

• Learning rate 1e-5 seems a good choice. If you change the learning rate, the model will be dramatically affected.
• It is crucial to shuffle the datasets in order to avoid overfitting.
• Do not set the batch size too big/small, 8 is OK. Or the model will be affected to some extent/difficult to converge.
• There is no pytorch_model.bin file in bert-chinese-wwm directory since it is a little large, remember to download it.
• When training with GPUs, pay attention to the device setting in Run.py, model/data.py, model/evaluate.py.
• Remember to change the name of .pkl of checkpoints and choose the right one for loading.