Language-Adversarial Training for Cross-Lingual Text Classification

This repo contains the source code for our TACL journal paper:

Adversarial Deep Averaging Networks for Cross-Lingual Sentiment Classification
Xilun Chen, Yu Sun, Ben Athiwaratkun, Claire Cardie, Kilian Weinberger
Transactions of the Association for Computational Linguistics (TACL)
paper (arXiv), bibtex (arXiv), paper (TACL), bibtex, talk@EMNLP2018

Introduction

ADAN transfers the knowledge learned from labeled data on a resource-rich source language to low-resource languages where only unlabeled data exists. It achieves cross-lingual model transfer via learning language-invariant features extracted by Language-Adversarial Training.

Requirements

• Python 3.6
• PyTorch 0.4
• PyTorchNet (for confusion matrix)
• scipy
• tqdm (for progress bar)

File Structure

.
├── README.md
└── code
    ├── data_prep                       (data processing scripts)
    │   ├── chn_hotel_dataset.py        (processing the Chinese Hotel Review dataset)
    │   └── yelp_dataset.py             (processing the English Yelp Review dataset)
    ├── layers.py                       (lower-level helper modules)
    ├── models.py                       (higher-level modules)
    ├── options.py                      (hyper-parameters aka. all the knobs you may want to turn)
    ├── train.py                        (main file to train the model)
    ├── utils.py                        (helper functions)
    └── vocab.py                        (vocabulary)

Dataset

The datasets can be downloaded separately here.

To support new datasets, simply write a new script under data_prep similar to the current ones and update train.py to correctly load it.

Run Experiments

python train.py --model_save_file {path_to_save_the_model}

By default, the code uses CNN as the feature extractor. To use the LSTM (with dot attention) feature extractor:

python train.py --model lstm --F_layers 2 --model_save_file {path_to_save_the_model}