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.cogis the entire Ugaritic-Hebrew data obtained from Ben Snyder.no_spestands for no special symbols since the original file contains special symbols that mark the morphological segmentations and affixes.uga-heb.small.no_spe.cogis 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.cogis the linear B data used in the paper.notebooks/Linear_b_simplified.ipynbis the same notebook I used for preparing the linear B data.linear_b-greek.names.cogis 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 updateto obtain all the submodules. - Install
pytorch. Any version>= 1.3should work. pip install -r requirements.txtfor other libraries.- Go to the three folders (
editdistance,arglib, anddev_misc) and runpip 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.