Emergent Translation in Multi-Agent Communication

PyTorch implementation of the models described in the paper Emergent Translation in Multi-Agent Communication.

We present code for training and decoding both word- and sentence-level models and baselines, as well as preprocessed datasets.

Dependencies

Python

• Python 2.7

• PyTorch 0.2

• Numpy

GPU

• CUDA (we recommend using the latest version. The version 8.0 was used in all our experiments.)

Related code

• For preprocessing, we used scripts from Moses and Subword-NMT.

Downloading Datasets

The original corpora can be downloaded from (Bergsma500, Multi30k, MS COCO). For the preprocessed corpora see below.

| | Dataset | | ------------- | ------------- | | Bergsma500 | Data | | Multi30k | Data | | MS COCO | Data |

Before you run the code

1. Download the datasets and place them in /data/word (Bergsma500) and /data/sentence (Multi30k and MS COCO)

2. Set correct path in scr_path() from /scr/word/util.py and scr_path(), multi30k_reorg_path() and coco_path() from /src/sentence/util.py

Word-level Models

Running nearest neighbour baselines

$ python word/bergsma_bli.py

Running our models

$ python word/train_word_joint.py --l1 <L1> --l2 <L2>

where <L1> and <L2> are any of {en, de, es, fr, it, nl}

Sentence-level Models

Baseline 1 : Nearest neighbour

$ python sentence/baseline_nn.py --dataset <DATASET> --task <TASK> --src <SRC> --trg <TRG>

Baseline 2 : NMT with neighbouring sentence pairs

$ python sentence/nmt.py --dataset <DATASET> --task <TASK> --src <SRC> --trg <TRG> --nn_baseline

Baseline 3 : Nakayama and Nishida, 2017

$ python sentence/train_naka_encdec.py --dataset <DATASET> --task <TASK> --src <SRC> --trg <TRG> --train_enc_how <ENC_HOW> --train_dec_how <DEC_HOW>

where <ENC_HOW> is either two or three, and <DEC_HOW> is either img, des, or both.

Our models :

$ python sentence/train_seq_joint.py --dataset <DATASET> --task <TASK>

Aligned NMT :

$ python sentence/nmt.py --dataset <DATASET> --task <TASK> --src <SRC> --trg <TRG>

where <DATASET> is multi30k or coco, and <TASK> is either 1 or 2 (only applicable for Multi30k).

Dataset & Related Code Attribution

• Moses is licensed under LGPL, and Subword-NMT is licensed under MIT License.

• MS COCO and Multi30k are licensed under Creative Commons.

Citation

If you find the resources in this repository useful, please consider citing:

@inproceedings{Lee:18,
  author    = {Jason Lee and Kyunghyun Cho and Jason Weston and Douwe Kiela},
  title     = {Emergent Translation in Multi-Agent Communication},
  year      = {2018},
  booktitle = {Proceedings of the International Conference on Learning Representations},
}