murawaki / latent-typology

Bayesian model for inducing latent representations of linguistic typological features

Now superseded by https://github.com/murawaki/lattyp

About

Yugo Murawaki. 2017. Diachrony-aware Induction of Binary Latent Representations from Typological Features. In Proceedings of the 8th International Joint Conference on Natural Language Processing (IJCNLP 2017), pp. 451-461. http://aclweb.org/anthology/I17-1046

Requirements

• Python2
  • numpy
  • scipy

The code also depends on the comp-typology package.

Inputs

• data/langs_full.json: languages taken from WALS (one language per line)
• data/flist.json: a subset of WALS features (single JSON object)

These files were generated using the comp-typology package.

Inference

Train the model

python train.py --seed=10 --initK=50 --maxanneal=100 --init_clusters --norm_sigma=10.0 --gamma_scale=1.0 --resume_if --output ../data/mda_K50.pkl ../data/langs_full.json ../data/flist.json 2>&1 | tee -a ../data/mda_K50.log

Collect samples while keeping W fixed

python sample_auto.py --seed=10 --iter=100 --a_repeat=5 ../data/mda_K50.pkl.final ../data/flist.json  ../data/mda_K50.xz.json 2>&1 | tee  ../data/mda_K50.xz.log
python convert_auto_xz.py --burnin=0 --update ../data/mda_K50.xz.json ../data/langs_full.json ../data/flist.json > ../data/mda_K50.xz.merged.json

Missing value imputation

make -j -f eval_mv.make all MODEL_PREFIX=mda TRAIN_OPTS="--init_clusters --maxanneal=100 --norm_sigma=10.0 --gamma_scale=1.0"