Key points are not available for this paper at this time.
We investigate the task of modeling open-domain, multi-turn, unstructured, -participant, conversational dialogue. We specifically study the effect of different elements of the conversation. Unlike previous efforts, focused on modeling messages and responses, we extend the modeling to context and participant's history. Our system does not rely on handwritten or engineered features; instead, we train deep neural networks on a large dataset. In particular, we exploit the structure of Reddit and posts to extract 2. 1 billion messages and 133 million. We evaluate our models on the task of predicting the next in a conversation, and we find that modeling both context and improves prediction accuracy.
Al‐Rfou et al. (Wed,) studied this question.