Benchmark Sound Law LSTM

Part of a project that tries to automatically derive sound laws from a list of cognates.

This project uses the ielex dataset as provided in Jäger et al. 2017, "Using support vector machines and state-of-the-art algorithms for phonetic alignment to ientify cognates in multi-lingual wordlists".

Prepare data

• Obtain NorthEuraLex dataset by running wget http://www.sfs.uni-tuebingen.de/~jdellert/northeuralex/0.9/northeuralex-0.9-forms.tsv.
• Obtain cognate set dataset and merge it with NorthEuraLex by using wikt_reader library. You would get a family file.
• Prepare input data by running

python scripts/process_data_wikt.py --data_path <path_to_family_file> --source <src> --targets <tgt_langs> --no_need_transcriber

For instance, for the Germanic language family, run

python scripts/process_data_wikt.py --data_path data/Germanic.tsv --source gem-pro --targets eng deu isl nor swe dan nld --no_need_transcriber

Dependencies

• various packages in requirements.txt. Run pip install -r requirements.txt.
• boost packages are needed. On Ubuntu, run sudo apt-get install libboost-all-dev.
• Install spdlog with the static lib version.