This repo contains experiment code accompanying:

Tianze Shi, Adrian Benton, Igor Malioutov, and Ozan Irsoy. "Diversity-Aware Batch Active Learning for Dependency Parsing". NAACL. 2021.

How to Train

We provide an example training script in scripts/train.sh. Please edit paths to train/dev/test conllu files accordingly. There are toggles to indicate which quality and diversity measures to use through the variables QUALITY and DIVERSITY. COMBO_METHOD variable indicates the use of DPPs or a naive top-K strategy.

Example

Run:

# Download EWT
wget https://raw.githubusercontent.com/UniversalDependencies/UD_English-EWT/master/en_ewt-ud-train.conllu
wget https://raw.githubusercontent.com/UniversalDependencies/UD_English-EWT/master/en_ewt-ud-dev.conllu
wget https://raw.githubusercontent.com/UniversalDependencies/UD_English-EWT/master/en_ewt-ud-test.conllu

sh train_ewt_amp-dpp_sample.sh

to train and evaluate a dependency parser on the UD English Web Treebank using the AMP w/ DPP active learning strategy. It is assumed that you are training on a GPU machine with a CUDA-enabled torch installation.

Code Structure

The main entry point for the model definition is deppar/parser.py. Active learning sampling strategies are implemented in deppar/acquisition.py for generic sample strategies, and deppar/dpp.py contains functions to sample batches using DPPs.

LICENSE

All files are licensed by

Copyright 2021 Bloomberg Finance L.P. Apache 2.0 License

unless otherwise noted in CONTENTS.