We aim to improve the bAbI benchmark as a means of developing intelligent dialogue agents. To this end, we propose concatenated-bAbI (catbAbI): an infinite sequence of bAbI stories. catbAbI is generated from the bAbI dataset and during training, a random sample/story from any task is drawn without replacement and concatenated to the ongoing story. The preprocessig for catbAbI addresses several issues: it removes the supporting facts, leaves the questions embedded in the story, inserts the correct answer after the question mark, and tokenises the full sample into a single sequence of words. As such, catbAbI is designed to be trained in an autoregressive way and analogous to closed-book question answering.

catbAbI models can be trained in two different ways: language modelling mode (LM-mode) or question-answering mode (QA-mode). In LM-mode, the catbAbI models are trained like autoregressive word-level language models. In QA-mode, the catbAbI models are only trained to predict the tokens that are answers to questions—making it more similar to regular bAbI. QA-mode is simply implemented by masking out losses on non-answer predictions. In both training modes, the model performance is solely measured by its accuracy and perplexity when answering the questions.