EASDRL

Code for the IJCAI-18 paper 'Extracting Action Sequences from Texts Based on Deep Reinforcement Learning'

requriements:

tensorflow / keras

wxpython

gensim

ipdb

...

• PS: There should be a folder named 'weights', but it was automatically removed by github since it is empty. Remember to add it if you want to train a new model.

• $ mkdir weights

Running

Training: All arguments are preset in main.py, so you can start training by:

$ python main.py

For trianing Argument Extractor, you can run:

$ python main.py --agent_mode arg

If you want to change the domain from 'cooking' to 'win2k' or 'wikihow', try:

$ python main.py --domain win2k

It may takes 2-4 hours for "win2k", 10-15 hours for "cooking" and 20-30 hours for "wikihow" in our computer with TITAN Xp GPU. Change the size of replay memory, GPU fraction or number of epochs according to your servers.

Human-agent Interaction: If you want to use the interacting environment, make sure you have installed the wxpython, and try:

$ python gui.py

It's the initial version, which is simple and maybe has some bugs. We have a latest version which adopts Active Learning for labeling data. It can be run by:

$ python guiActiveLearning.py

About the data

The following {domain} can be one of "cooking", "wikihow" and "win2k"

Labeled data

1. {domain}_labeled_text_data.pkl is the labeled data for action name extractor
2. refined_{domain}_data.pkl is the labeled data for action argument extractor

POS data

1. {domain}_dependency.pkl contains the part-of-speech data for action name extractor
2. {domain}_arg_pos.pkl contains the part-of-speech data for action argument extractor

Unlabeled data

1. home_and_garden_500_words_with_tile.pkl contains more than 15k unlabeled texts from WikiHow Home and Garden.

Others

1. wordvec_dim* is pre-trained word2vec
2. There are some simple text in ./data/online_test/. They are originally used for online interaction test.