This repository restore the sorce code and datasets for "Using Prior Knowledge to Guide BERT’s Attention in Semantic Textual Matching Tasks"

1.First, you can use the get_similarity.py file in the ESIM/scripts/preprocessing folder to obtain the similarity matrix of the msrp/sts/url dataset. For QQP dataset, you should use get_similarity_quora.py. You can also use the matrix we have built, then this step can be skipped.

3.Then run the preprocess_msrp.py file in the preprocessing folder to preprocess the data.

5.Run under the ESIM/scripts/training folder.

python main_msrp.py --proportion 0.1 --output 10

--proportion specifies the size of the dataset, --output determines the storage location of models trained on datasets of different sizes, where 10 is 10% of the data
You can run python main_msrp.py directly, and use the default 100% data at this time.

Here is the data we have processed, Please place it under the ESIM/data/dataset folder after downloading.

We use glove.840B.300d as embedding, which can be downloaded here, and then please put it in the ESIM\data\embeddings folder.

1.Download the data we have processed, and put it in the UER/datasets folder, Or you can use the get_similarity.py file in ESIM to preprocess the data.

2.The data we provide cannot be directly used in the BERT model, so further preprocessing is required to adapt to the structure of the BERT model, use get_similarity.py in UER folder to further preprocess the data.

3.Then use the following command to run the BERT model

python run.py --train_path datasets/msrp/train.tsv 
--dev_path datasets/msrp/dev.tsv 
--test_path datasets/msrp/test.tsv 
--output_model_path models/msrp_100.bin 
--proportion 1.0

The pretrained_model can be dowmloaded here.

If you use this code, please cite this paper:

@inproceedings{xia2021using,
  title={Using Prior Knowledge to Guide BERT’s Attention in Semantic Textual Matching Tasks},
  author={Xia, Tingyu and Wang, Yue and Tian, Yuan and Chang, Yi},
  booktitle={Proceedings of the Web Conference 2021},
  pages={2466--2475},
  year={2021}
}