On the Word Boundaries of Emergent Languages Based on Harris's Articulation Scheme

The implementation code of "On the Word Boundaries of Emergent Languages Based on Harris's Articulation Scheme." ICLR 2023.

The word segmentation algorithm can be found at boundary_detection.EntropyCalculator.

This repo depends on EGG toolkit. So, first of all, install EGG toolkit:

$ pip install git+ssh://git@github.com/facebookresearch/EGG.git

For more detail, refer to https://github.com/facebookresearch/EGG.

For training agents,

$ python train.py [options] > output_file.txt

To check options,

$ python train.py --help

For applying boundary detection,

$ python boundary_detection.py --config [configuration_file]

Format & example of configuration file:

log_dirs: [
    output_file_1.txt
    ...
    output_file_n.txt
]
least_acc: 0.9  # successful language
trained_lang_mode: "max_epoch"  # Option["max_epoch", "min_epoch_to_reach_least_acc"]
img_dir: "./img_dir"  # Directory to save png files
tmp_dir: "./tmp_dir"  # Directory to save temporary files

Configuration files have to be made in yaml format.