Shift Reduce Dependency Parser for Japanese

Overview

A neural network based dependency parser (syntax tree parser) for Japanese.

TL;DR Summary:

1. Implemented a shift reduce parser using the arc-standard transition system.
2. Implemented a neural network using PyTorch to predict the next action given a parser state.
3. Trained neural network with UD Japanese GSD treebank and pretrained word embedding weights from Wikipedia2vec.

Example/Visualization

Example usage and dependency tree visualization in a Google Colab notebook here.

Training

Run python3 model.py to execute training loop.

Model weights are saved in PyTorch model state format to model.pth. Other relevant files such as model_lists.txt and embeddings/jawiki_gsd_word2vec.txt are also required to load the model.

The model should generally converge at approximately .96 LAS (Labelled Attachment Score), .97 UAS (Unlabelled Attachment Score). Increasing the hyperparameters (embed size, hidden size) beyond the specified defaults may marginally improve accuracy.

References

• http://www.cs.toronto.edu/~gpenn/csc485/a1.pdf
• http://www1.cs.columbia.edu/~stratos/research/sr_parsing.pdf
• https://universaldependencies.org/treebanks/ja_gsd/index.html
• https://wikipedia2vec.github.io/wikipedia2vec/pretrained/