biaffineparser's Introduction

biaffineparser: Deep Biaffine Attention Dependency Parser

biaffineparser is a PyTorch implementation of "Deep Biaffine Attention for Neural Dependency Parsing."

Installation

biaffineparser works on PyTorch.

$ git clone https://github.com/chantera/biaffineparser
$ cd biaffineparser
$ pip install -r requirements.txt

Usage

Training

usage: main.py train [-h] --train_file FILE [--eval_file FILE]
                     [--embed_file FILE] [--max_steps NUM]
                     [--eval_interval NUM] [--batch_size NUM]
                     [--learning_rate VALUE] [--cuda] [--save_dir DIR]
                     [--cache_dir DIR] [--seed VALUE]

Evaluation

usage: main.py evaluate [-h] --eval_file FILE --checkpoint_file FILE
                        --preprocessor_file FILE [--batch_size NUM] [--cuda]
                        [--verbose]

Example

$ mkdir models
$ python3 src/main.py train --train_file $DATA/train.conll --eval_file $DATA/dev.conll --embed_file $DATA/glove.6B.100d.txt --cuda --save_dir ./models
$ python3 src/main.py evaluate --eval_file $DATA/test.conll --ckpt ./models/step-[num].ckpt --proc ./models/preprocessor.pt --cuda

Performance

The model achieves UAS: 95.77 and LAS: 94.10 in wsj 23 (test set) in PTB-SD 3.3.0 with the reported hyperparameters.

References

Dozat, T., Manning, C. D., 2016. Deep Biaffine Attention for Neural Dependency Parsing. arXiv preprint arXiv:1611.01734. https://arxiv.org/abs/1611.01734