e2e seq2seq NLG

This is the code for doing e2e seq2seq NLG using Fastai.

It is targeted at the e2e NLG competition.

Medium article explaining this work is available here.

Requirements

• python==3.7.4
• fastai==1.0.59
• fasttext==0.9.1
• torch==1.3.1

External resources

• e2e NLG dataset
• e2e NLG metrics script.
• Fasttext English word vectors.
• Classifier model for reranking (vocab.pkl and classifier_model.pth)

Code content

• notebooks for training and testing the models.
• Fastai code implementing the seq2seq model, and taken from here.
• code for preprocessing texts and meaning representations before feeding it to seq2seq: cleaning, delexicalization.
• code for postprocessing texts and meaning representations before feeding them to evaluation scripts: relexicalization, displaying outputs in same order as input and references files.
• code for data augmentation, inferencing using MR classifier, and reranking.