Conversational Contextual Cues: The Case of Personalization and History for Response Ranking

TL;DR

This paper explores enhancing response ranking in open-domain multi-turn conversations by incorporating conversational context and participant history using deep neural networks trained on a large Reddit dataset.

Contribution

It extends previous models by integrating long conversation context and participant history without relying on handcrafted features, using a large-scale neural network approach.

Findings

Modeling context and participants improves response prediction accuracy.

Deep neural networks trained on Reddit data effectively capture conversational cues.

Long conversation history contributes significantly to response relevance.

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.