Code for EMNLP 2020 paper: Adversarial Semantic Decoupling for Recognizing Open-Vocabulary Slots.
python==3.7.0
pytorch==1.3.1
allennlp==0.9.0
The BERT pretrained models is available at https://huggingface.co/models. To run experiments with BERT, you should download the bert-large-uncased model to local file system and modify the path ("pretrained_model": "/local/path/to/bert-large-uncased/") in the config file to your downloading path.
The datasets we use are publicly available at:
We provided dataset in /data dir, together with the statistics including:
- number of samples for each slot type
- list of intents where the slot type appears
- example with longest slot value (in number of tokens)
- number of tokens for longest slot value
- list of slot values
- number of slot value types
The original splits for MIT-restaurant only contain train set (7660 samples) and test set (1521 samples). We further split the original train set into train (6894 samples) and valid (766 samples) sets.
To train the model, run:
python3 train.py --config_path ./config/SETTING.json --output_dir ./outputs/SETTING/
where SETTING should be the specific setting in /config dir, e.g. rnn-snips, bert-mr or your customized configs.
To test the trained model, run:
python3 test.py --output_dir ./outputs/SETTING/
where ./outputs/SETTING/ should be the output_dir param same as that in the training command.
If you find the code useful, please consider citing our paper:
@inproceedings{yan-etal-2020-adversarial,
title = "Adversarial Semantic Decoupling for Recognizing Open-Vocabulary Slots",
author = "Yan, Yuanmeng and
He, Keqing and
Xu, Hong and
Liu, Sihong and
Meng, Fanyu and
Hu, Min and
Xu, Weiran",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.emnlp-main.490",
pages = "6070--6075",
}