Bootstrapping incremental dialogue systems from minimal data: the generalisation power of dialogue grammars

摘要

We investigate an end-to-end method for automatically inducing task-baseddialogue systems from small amounts of unannotated dialogue data. It combinesan incremental semantic grammar - Dynamic Syntax and Type Theory with Records(DS-TTR) - with Reinforcement Learning (RL), where language generation anddialogue management are a joint decision problem. The systems thus produced areincremental: dialogues are processed word-by-word, shown previously to beessential in supporting natural, spontaneous dialogue. We hypothesised that therich linguistic knowledge within the grammar should enable a combinatoriallylarge number of dialogue variations to be processed, even when trained on veryfew dialogues. Our experiments show that our model can process 74% of theFacebook AI bAbI dataset even when trained on only 0.13% of the data (5dialogues). It can in addition process 65% of bAbI+, a corpus we created bysystematically adding incremental dialogue phenomena such as restarts andself-corrections to bAbI. We compare our model with a state-of-the-artretrieval model, MemN2N. We find that, in terms of semantic accuracy, MemN2Nshows very poor robustness to the bAbI+ transformations even when trained onthe full bAbI dataset.