FlorianPfisterer / scs-baselines

Baseline models for the paper: "Modeling Naive Psychology of Characters in Simple Commonsense Stories" by Hannah Rashkin, Antoine Bosselut, Maarten Sap, Kevin Knight and Yejin Choi

Story Commonsense Baselines

Classifiying common sense emotional and motivational states in stories

Installation

Starting from whatever directory you will be placing the scs-baselines repository, run the following commands:

git clone [e-mail]:atcbosselut/scs-baselines.git
cd scs-baselines
wget http://homes.cs.washington.edu/~antoineb/datasets/scs-baselines-data-noids.tar.gz
tar -xvzf scs-baselines-data.tar.gz
cd ..

These dependencies must also be installed. Apart from Jobman, they should be available from a typical package manager such as anaconda or pip:

• python2.7
• progressbar2
• pandas
• pytorch3.1
• nltk
• Jobman

Installing Jobman

Instructions for installing Jobman are a bit convoluted, so just run the following commands from your home directory and you should be fine:

git clone git://git.assembla.com/jobman.git Jobman
cd jobman
python setup.py install

Making Data

Run the following command from the working directory.

bash make_data.sh

Running experiments

Training a classification model

To run a classification model run the following command:

python src/main_class.py

This command will load the configuration settings in the config/class_config.json file and run a model according to these parameters. The src/config.py source file explains what each variable in this configuration file does.

Training a generation model

To run a generation model, run the following command:

python src/main_gen.py

This command will load the configuration settings in the config/gen_config.json file and run a model according to these parameters. The src/config.py source file explains what each variable in this configuration file does.

Training a classification model with a pretrained generation model

To run a classification model using a model pretrained on generation, do the following. First, initialize an entry in the config/pretrained_config.json JSON configuration file whose key is "load_model_${MODEL_TYPE}_${TASK}" where $TASK is one of motivation or emotion and ${MODEL_TYPE} is a class of model such as lstm, cnn, ren, or npn. An example is provided in the configuration file. Then, run the following command:

python src/main_pretrained.py

This command will load the configuration settings in the config/pretrained_config.json file and run a model according to these parameters. The src/config.py source file explains what each variable in this configuration file does.

Evaluating a model on the test set

Set the load_model_name value in the config/class_config.json file. Then run the following command:

python src/evaluate_test.py

Contact

Feel free to reach out with questions to [e-mail]