KG-A2C

Goal driven language generation using knowledge graph A2C agents. This code accompanies the paper Graph Constrained Reinforcement Learning for Natural Language Action Spaces.

Bibtex

@inproceedings{
ammanabrolu2020graph,
title={Graph Constrained Reinforcement Learning for Natural Language Action Spaces},
author={Prithviraj Ammanabrolu and Matthew Hausknecht},
booktitle={International Conference on Learning Representations},
year={2020},
url={https://openreview.net/forum?id=B1x6w0EtwH}
}

Quickstart

Install Dependencies: Jericho, Redis, Pytorch >= 1.2

pip3 install --user jericho
pip3 install torch torchvision
sudo apt-get install redis-server

Download and extract Stanford CoreNLP then start the OpenIE server:

cd stanford-corenlp-full-2018-10-05/ && java -mx8g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 9000 -timeout 15000

Train KG-A2C

cd kga2c && python train.py --rom_file_path path_to_your_rom --openie_path path_to_your_openie_install --tsv_file ../data/rom_name_here