Graph-based dependency parsing using BiLSTM feature extractors

A PyTorch implementation of the BIST graph-based parser as described in the paper Simple and Accurate Dependency Parsing Using Bidirectional LSTM Feature Representations.

The parser acheives 94.38% UAS and 92.93% LAS on the standard Penn Treebank dataset (Standford Dependencies).

Requirements

• Python 3.7
• PyTorch 1.8.1
• NLTK 3.6.2

Data format

The software requires having a train.conll and dev.conll files formatted according to the CoNLL data format.

Train a parsing model

For training a graph-based parsing model, run:

python main.py --train_path data/train.conll --dev_path data/dev.conll --epochs 30 --lr 1e-3 --w_emb_dim 100 --pos_emb_dim 25 --lstm_hid_dim 125 --mlp_hid_dim 100 --n_lstm_layers 2

Parse data

For parsing data with a previously trained model, run:

python main.py --test_path data/test.conll --model_dir [model directory] --do_eval