Neural Response Ranker for Alana, Heriot-Watt University's Alexa Prize Socialbot

Code for the paper "Neural Response Ranking for Social Conversation: A Data-Efficient Approach" by Igor Shalyminov, Ondřej Dušek, and Oliver Lemon (Search-Oriented Conversational AI, an EMNLP 2018 Workshop) [Paper] [Slides].

1. Train a new model

python alana_learning_to_rank/learning_to_rank.py <trainset>.json <devset>.json <testset>json <model_folder>

2. Evaluate accuracy of a training model

python alana_learning_to_rank/evaluate.py <model_folder> <eval_dataset>.json

We provide 2 models trained from Alexa Prize dialogue transcripts, length_model_1M and rating_model, with length and rating targets respectively.

For a greater applicability, the models were trained without features involving bot names (which are specific to our setup), therefore the actual accuracy numbers might differ from reported ones.

Due to the copetition rules, we can't share any data containing real users' interactions with the system. Therefore, we only provide examples of datasets for training and evaluation stages (pandas json dataframe format) which you can check with while collecting your own data.