delmaksym/bigram_embedder

Unsupervised Compositional Modeling Framework for Phrases

Results replication guide

Prepare data for training

• cd data_gen_scripts
• run step0 script to download and preprocess wiki data
• run step1 script to extract potential bigrams we will work with
• run step2 script to glue some words to form bigrams in corpus from step0
• run step3 script to train fasttext skip-gram model on words & bigrams corpus from step2 (get fasttext from https://fasttext.cc/))
• run step4 script to extract actual wordpair_vec -> bigram_vec corpus needed to train prediction model
• run step5 script to create train, dev, and test splits

Train model

So by now you have the right data in data/step5 folder.
Next, set up your development environment by installing what is in requirements.txt, and spacy>=2.0. (as after step5 data generation step) and run:

python -m allennlp.run train experiments/bigram_embedder.json -s models/bigram_embedder_feedforw --include-package bigram_embedder

This project relies on python-3.6.