This repo hosts the codebase for reproduce the results for the ACL paper Neural Decipherment via Minimum-Cost Flow: from Ugaritic to Linear B.

Data for linear B and Ugaritic are included in the data folder:

• uga-heb.no_spe.cog is the entire Ugaritic-Hebrew data obtained from Ben Snyder. no_spe stands for no special symbols since the original file contains special symbols that mark the morphological segmentations and affixes.
• uga-heb.small.no_spe.cog is the exact random subset of Ugaritic data I used in the paper for training the model. Around one tenth of the original file.
• linear_b-greek.cog is the linear B data used in the paper. notebooks/Linear_b_simplified.ipynb is the same notebook I used for preparing the linear B data.
• linear_b-greek.names.cog is the linear B data that only included names on the Greek side.

Note that you might need to install fonts in order to render Linear B scripts properly in your computer.

• Run git submodule init && git submodule update to obtain all the submodules.
• Install pytorch. Any version >= 1.3 should work.
• pip install -r requirements.txt for other libraries.
• Go to the three folders (editdistance, arglib, and dev_misc) and run pip install . in each folder to install the dependency.
• Run pip install . in the root folder.

The recommended way to run this program is to write down a configuration class first. See nd/config/decipher_config.py for an example, and accordingly you can use python nd/main.py --cfg UgaHebSmallNoSpe to start running.