mengf1 / PAL

Policy based Active Learning with DQN (EMNLP-2017)

Welcome to Learning how to Active Learn

Introduction

This source code is the basis of the following paper:

Learning how to Active Learn: A Deep Reinforcement Learning Approach, EMNLP 2017

Building

It's developed on TensorFlow.

• Install TensorFlow
• Install pycrfsuite

Code

• launcher_ner_bilingual: the starter of playing
• game_ner: the game
• robot: active learning policy
• tagger

How to run

For example, we train an active learning policy on English and then apply the policy to German.

python launcher_ner_bilingual.py --agent "CNNDQN" --episode 10000 --budget 1000 --train "en.train;en.testa;en.testb;en.emb;en.model.saved" --test "de.train;de.testa;de.testb;de.emb;de.model.saved"

Data resource

• CoNLL2002 (or in this url: https://www.clips.uantwerpen.be/conll2002/ner/)
• CoNLL2003 (or in this url:https://www.clips.uantwerpen.be/conll2003/ner/)
• Crosslingual word embeddings and corpora