The code is a full implementation of SPNet in my paper Abstractive Dialog Summarization with Semantic Scaffolds(https://arxiv.org/abs/1910.00825), which is my first trial in independent research.

Our model is built in AllenNLP, which is an open-source NLP research library built on PyTorch. You can install it following its GitHub repo.

We adapt MultiWOZ-2.0 to our abstractive dialog summarization task. We release the delexicalized data with domain annotation in ./data folder. And./reqt_dict.json is the requests in test set, used for evaluation.

We use default setting of SPNet in ./models/mtModel and ./train.py, run python3 train.py to start training. Note that you may need to specify the GPU in the code. Logs and models will be saved in ./models_saved/SPNet. We also provide pretrained weights in ./models_saved/model_weights.th.

Run python3 inference.py and the predictor will generate ./predictions.txt with 1,000 lines of summaries. They are the prediction of 1,000 test samples and the reference summaries are saved in ./references.txt. Note that it will use our pre-trained model as a default choice.

We have two automatic evaluation metrics: ROUGE and CIC. To separate the inference and evaluation, we use file2rouge tools to measure the ROUGE scores. You can follow the instructions in files2rouge GitHub page.

CIC evaluation method is shown in ./evaluate.py.