We currently have the following training scripts for the models:

• GPT2 Baseline Text + Fact
• Knowledge Dependent Policy Driven Neural Response Generator using Mezza Tags

For any clarification related to the above code, please reach out to Rishi Rajasekaran (rrajasek@ucsc.edu)

Scripts to train Seq2Seq and Transformer models on the Amazon Topical-Chat Corpus. This code serves as the baseline for DSTC9 Track 3.

To train: python3 train.py --use_knowledge --transformer --save_path transformer/

To test: python3 test.py --use_knowledge --transformer --save_path transformer/

To serve interactive model with TF-IDF based fact selection: python3 dynamic.py --use_knowledge --transformer --save_path transformer/

The pre-processed data can be found in data.zip. If you would like to use a different pre-processing strategy, please download the original data from here.

The dataset preparation code is split between the utils.py file and the tc_dataset.py. The data loading and tokenization is done in utils.py while the data preparation to feed into the model is done in tc_dataset.py.

If you experience any issues with this code, please contact me at mehrishikib@gmail.com

• spacy
• python -m spacy download en_core_web_lg
• nltk.download('punkt')