man

KO | EN

gitlite — search

#javascript #python #react #go #android #machine-learning #rust #linux #docker #macos #awesome #vue #awesome-list #pytorch

(★ 81)

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

File Explorer

Download Latest Version (.zip)

- - fdu_mtl_dataset.py
  - folded_dataset.py
  - msda_preprocessed_amazon_dataset.py
- layers.py
- models.py
- options.py
- train_man_exp1.py
- train_man_exp2.py
- train_man_exp3.py
- utils.py
- vocab.py
- - apparel.task.test
  - apparel.task.train
  - apparel.task.unlabel
  - baby.task.test
  - baby.task.train
  - baby.task.unlabel
  - books.task.test
  - books.task.train
  - books.task.unlabel
  - camera_photo.task.test
  - camera_photo.task.train
  - camera_photo.task.unlabel
  - dvd.task.test
  - dvd.task.train
  - dvd.task.unlabel
  - electronics.task.test
  - electronics.task.train
  - electronics.task.unlabel
  - health_personal_care.task.test
  - health_personal_care.task.train
  - health_personal_care.task.unlabel
  - imdb.task.test
  - imdb.task.train
  - imdb.task.unlabel
  - kitchen_housewares.task.test
  - kitchen_housewares.task.train
  - kitchen_housewares.task.unlabel
  - magazines.task.test
  - magazines.task.train
  - magazines.task.unlabel
  - MR.task.test
  - MR.task.train
  - MR.task.unlabel
  - music.task.test
  - music.task.train
  - music.task.unlabel
  - software.task.test
  - software.task.train
  - software.task.unlabel
  - sports_outdoors.task.test
  - sports_outdoors.task.train
  - sports_outdoors.task.unlabel
  - toys_games.task.test
  - toys_games.task.train
  - toys_games.task.unlabel
  - video.task.test
  - video.task.train
  - video.task.unlabel
- - amazon-shuffle-indices.pkl
  - amazon.pkl
- - word2vec.tar.gz
.gitignore
environment.yml
LICENSE
README.md

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Multinomial Adversarial Nets
This repo contains the source code for our paper:

[**Multinomial Adversarial Networks for Multi-Domain Text Classification**](http://aclweb.org/anthology/N18-1111)
<br>
[Xilun Chen](http://www.cs.cornell.edu/~xlchen/),
[Claire Cardie](http://www.cs.cornell.edu/home/cardie/)
<br>
NAACL-HLT 2018
<br>
[paper](http://aclweb.org/anthology/N18-1111),
[bibtex](https://aclanthology.coli.uni-saarland.de/papers/N18-1111/n18-1111.bib)

## 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](https://www.aclweb.org/anthology/P19-1299) (source code available [here](https://github.com/microsoft/Multilingual-Model-Transfer)) 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

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

## Experiment 2: Multi-Source Domain Adaptation
```bash
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

```bash
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.

man

File Explorer

Today's picks

Was this content helpful?

Recent Feedback

man

File Explorer

Today's picks

Was this content helpful?

Recent Feedback

Download README