Cross-Lingual-NBT

[News]: Just fixed the ontology problem, now the code should be able to run.

Code for EMNLP 2018 paper "XL-NBT: A Cross-lingual Neural Belief Tracking Framework"

In this paper, we propose the interesting cross-lingual state tracking problem and design a simple yet very efficient algorithm to tackle it. This repository contains all the experiment codes for both En->De and En->It transfer scenarios under two different parallel resource cases.

For more details, please check the latest version of the paper: https://128.84.21.199/abs/1808.06244.

The general architecture of the model is displayed as below:

We decompose the transfer into two components, language understanding level transfer and ontology transfer.

Prerequisites

Python 2.7
TensorFlow 1.4
cuda & cudnn

Code Structure

code/ contains all the python modules to support the algorithm
config/ contains all the parameter setting
data/ contains the dialog data for different languages and also the parallel translation data for XL-NBT-C
word-vectors/ contains the pre-trained multi-lingual word embedding and two bilingual dictionary for XL-NBT-D
models/ stores the saved models during training
train.sh sets up all the supervised learning and transfer learning

Config File

[train]
batch_size=256
batches_per_epoch=64
max_iteration=100000
restore=True
alpha=1.0

[model]
id=student-dict-bilingual
restore_id=teacher-bilingual
dataset_name=woz
language=english
foreign_language=german
value_specific_decoder=False
learn_belief_state_update=True
tau=0.1

Learning Curve

Usage

1. Setup

Clone this github repository recursively:

git clone https://github.com/wenhuchen/Cross-Lingual-NBT.git ./

Please Download the Bilingual embedding and put them into word-vectors/ folder

2. Supervised Learning

We use cross entropy loss to train the teacher model, please set "id=teacher-bilingual" and "restore=False" in the config file.

./train.sh train

Here, we already upload the pre-trained teacher models into the models/ folder, you can skip this step and jump to step 3 or 4 depending on differnt transfer scenarios.

3. Transfer Learning XL-NBT-C

We use teacher-student framework to transfer knowledge across language boundary, please set "restore_id=teacher-bilingual" and "restore=True" in the config file, and "id=student-corpus-bilingual".

./train.sh corpus_transfer

4. Transfer Learning XL-NBT-D

We use teacher-student framework to transfer knowledge across language boundary, please set "restore_id=teacher-bilingual" and "restore=True" in the config file, and "id=student-dict-bilingual".

./train.sh dict_transfer

If you find this paper useful, please this paper with the following bibtex

@inproceedings{chen2018xl,
  title={XL-NBT: A Cross-lingual Neural Belief Tracking Framework},
  author={Chen, Wenhu and Chen, Jianshu and Su, Yu and Wang, Xin and Yu, Dong and Yan, Xifeng and Wang, William Yang},
  booktitle={Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing},
  pages={414--424},
  year={2018}
}

Acknowledgement

Neural-Belief Tracker

Jiahuan-Pei/Cross-Lingual-NBT