Skip to content

Latest commit

 

History

History
24 lines (22 loc) · 1.9 KB

README.md

File metadata and controls

24 lines (22 loc) · 1.9 KB

Training a ConvE based encoder for definitions

This folder contains a copy of the necessary files and modifications over https://github.com/TimDettmers/ConvE.

Steps to train embeddings

  1. Preprocess files 
    bash preprocess.sh
  2. Run ConvE to get initial embeddings for entities and relations. Model is saved at saved_models/WN18RR_conve_0.2_0.3_.model. When running for the first time, add the --preprocess flag. 
    CUDA_VISIBLE_DEVICES=0 python conve_main.py --model conve --data WN18RR  --input-drop 0.2 --hidden-drop 0.3 --feat-drop 0.2  --lr 0.001
  3. Train ConvE with a definition encoder. Model is saved at saved_models/WN18RR_conve_0.2_0.3__defn.model 
    CUDA_VISIBLE_DEVICES=0 python run_definition_encoder.py --model conve --data WN18RR  --input-drop 0.2 --hidden-drop 0.3 --feat-drop 0.2  --lr 0.0001 --batch-size 128 --test-batch-size 128  --epochs 500  --initialize  saved_models/WN18RR_conve_0.2_0.3_.model
  4. Generate representations for definitions using trained model. Embeddings are saved at saved_embeddings/embeddings.npz. These embeddings can now be used to train a WSD model following the instructions in the root directory. 
    CUDA_VISIBLE_DEVICES=0 python run_definition_encoder.py --model conve --data WN18RR  --input-drop 0.2 --hidden-drop 0.3 --feat-drop 0.2  --lr 0.0001 --batch-size 128 --test-batch-size 128  --epochs 0  --resume --represent saved_embeddings/embeddings.npz

LICENSE

Most of the files in this directory are copied with or without modifications from https://github.com/TimDettmers/ConvE. The corresponding LICENSE applies (also copied at LICENSE in this directory). The modifications are described in the correspoding files. An exception is definition_preprocessor.py which is an entirey original code written for this project and the LICENSE in the root directory applies.