Conversational Contextual Cues: The Case of Personalization and History for Response Ranking
Venue
preprint (2016)
Publication Year
2016
Authors
Rami Al-Rfou, Marc Pickett, Javier Snaider, Yun-hsuan Sung, Brian Strope
BibTeX
Abstract
We investigate the task of modeling open-domain, multi-turn, unstructured, multi-
participant, conversational dialogue. We specifically study the effect of
incorporating different elements of the conversation. Unlike previous efforts,
which focused on modeling messages and responses, we extend the modeling to long
context and participant’s history. Our system does not rely on handwritten rules or
engineered features; instead, we train deep neural networks on a large
conversational dataset. In particular, we exploit the structure of Reddit comments
and posts to extract 2.1 billion messages and 133 million conversations. We
evaluate our models on the task of predicting the next response in a conversation,
and we find that modeling both context and participants improves prediction
accuracy.
