ha-lins / maml-dst Goto Github PK

This repo implements the model-agnostic meta-learning (MAML) for dialogue state tracking (DST). It achieves better performance than TRADE in the few-shot setting. It also implements the multi-task learning (MTL-DST) as the baseline. This code has been written using PyTorch >= 1.0.

This project has a similar idea to MERET. Our project is preliminarily named recallable meta-learning for DST (RM-DST), which can not only fast adapt to a low-resource new domain, but also employ a recallable mechanism to maintain the high performance in old tasks. If you are interested in the work, you can contact me for the draft. The overview of RM-DST is as follows:

Download the MultiWOZ dataset and the processed dst version.

❱❱❱ python3 create_data.py

Check the packages needed or simply run the command

❱❱❱ pip install -r requirements.txt

If you run into an error related to Cython, try to upgrade it first.

❱❱❱ pip install --upgrade cython

Training

❱❱❱ python3 myTrain_maml_DND.py -dec=TRADE -bsz=32 -dr=0.2 -lr=0.001 -le=1 -exceptd=${domain}

Training

❱❱❱ python3 myTrain_MTL.py -dec=TRADE -bsz=32 -dr=0.2 -lr=0.001 -le=1 -exceptd=${domain} -add_name=MTL

• -exceptd: except domain selection, choose one from {hotel, train, attraction, restaurant, taxi}.

MAML-DST

❱❱❱ python3 fine_tune_dnd.py -bsz=8 -dr=0.2 -lr=0.001 -path=${save_path_except_domain} -exceptd=${except_domain}

MTL-DST

❱❱❱ python3 fine_tune.py -bsz=8 -dr=0.2 -lr=0.001 -path=${save_path_except_domain} -exceptd=${except_domain}

1. DAML: Domain Adaptive Dialog Generation via Meta Learning. [ACL 2019]
2. TRADE: Transferable Multi-Domain State Generator for Task-Oriented Dialogue Systems. [ACL 2019]
3. Meta-Reinforced Multi-Domain State Generator for Dialogue Systems. [ACL 2020]