WriterForcing: Generating more interesting story endings

This repository has code for the published work at ACL Storytelling workshop 2019. Please cite this paper if you use any part of this code : https://www.aclweb.org/anthology/W19-3413

Note: You need to have a GPU (and CUDA) to exucute the code.

Setup instructions

Create a conda environment:

conda create -n writerForcing python=3.6

Activate the environment:

source activate writerForcing

Install a few required packages:

conda install pytorch torchvision -c pytorch 

pip install torchtext 

pip install rake_nltk

Download the Glove embeddings from here and copy the file named "glove.6B.200d.txt" in the data folder.

To train a model with "keyphrase addition", please run the following command:

python -m model.model_train_seq2seq --keyword_attention alpha_add

To train a model with "keyphrase attention loss", please run the following command:

python -m model.model_train_seq2seq --keyword_attention att_sum_mse

To train a model with "context concatenation", please run the following command:

python -m model.model_train_seq2seq --keyword_attention context_add

To train a model with "coverage loss", please run the following command:

python -m model.model_train_seq2seq --keyword_attention att_sum_coverage

Story cloze evaluation

As mentioned in the paper we have used BERT for story cloze evaluation. The model can be downloaded here and must be placed outside the git repo.

To install BERT follow the instructions in this repo: After training your model with flag--keyword_attention att_sum_mse_True the outputs would be written to att_sum_mse_True, then run the evalutation script as follows

cd evaluate_attributes
python evaluate_story_cloze.py --model1 outputs/ie.txt --model2 ../att_sum_mse_True/../att_sum_mse_True/keyadd_outputs_storys_

Distinct Metric :

cd outputs
python calculate_metrics.py --file output_file_name_here.txt

Best Outputs

The best output of all the models compared in the paper are present in the outputs folder