ccsasuke / man

Multinomial Adversarial Networks for Multi-Domain Text Classification (NAACL 2018)

Multinomial Adversarial Nets

This repo contains the source code for our paper:

Multinomial Adversarial Networks for Multi-Domain Text Classification
Xilun Chen, Claire Cardie
NAACL-HLT 2018
paper, bibtex

Requirements:

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

An Anaconda environment.yml file is also provided.

Dev version

The dev branch contains a full version of our code, including some options we experimented but did not include in the final version.

Follow-up Paper

You might also be interested in our follow-up ACL 2019 paper (source code available here) with an improved model and better performance.

Before Running

The pre-trained word embeddings file exceeds the 100MB limit of github, and is thus provided as a gzipped tar ball. Please run the following command to extract it first:

tar -xvf data/w2v/word2vec.tar.gz -C data/w2v/

Experiment 1: MDTC on the multi-domain Amazon dataset

cd code/
python train_man_exp1.py --dataset prep-amazon --model mlp

Experiment 2: Multi-Source Domain Adaptation

cd code/
# target domain: books
python train_man_exp2.py --dataset prep-amazon --model mlp --no_wgan_trick --domains dvd electronics kitchen --unlabeled_domains books --dev_domains books
# target domain: dvd
python train_man_exp2.py --dataset prep-amazon --model mlp --no_wgan_trick --domains books electronics kitchen --unlabeled_domains dvd --dev_domains dvd
# target domain: electronics
python train_man_exp2.py --dataset prep-amazon --model mlp --no_wgan_trick --domains books dvd kitchen --unlabeled_domains electronics --dev_domains electronics
# target domain: kitchen
python train_man_exp2.py --dataset prep-amazon --model mlp --no_wgan_trick --domains dvd electronics kitchen --unlabeled_domains kitchen --dev_domains kitchen

Experiment 3: MDTC on the FDU-MTL dataset

cd code/
python train_man_exp3.py --dataset fdu-mtl --model cnn --max_epoch 50

A larger batch size can also be used to reduce the training time.