This repo contains the implementation of two regularization techniques for Deep Matching Networks, built on top of the code from https://github.com/yangliuy/NeuralResponseRanking. We added a parameter to employ either Domain Adversarial Learning (DAL) to induce domain-agnostic representations, or to apply multi-task learning for domain classification (MTL) inducing domain-aware representations. We modified the .config files to receive extra inputs such as the out-of-domain prediction set.
To enable MTL or DAL use the following parameter with either 'DMN-ADL' or 'DMN-MTL'as input :
python main_conversation_qa.py --domain_training_type '$REGULARIZATION'
To see some examples and run the code you can also use this google colab notebook that clones the repo, downloads the dataset and run experiments.