Towards a Human-like Open-Domain Chatbot

TL;DR

This paper introduces Meena, a large neural network chatbot trained on social media data, and proposes a new human evaluation metric, SSA, demonstrating its effectiveness in measuring human-like conversation quality.

Contribution

The paper presents Meena, a 2.6B parameter open-domain chatbot trained end-to-end, and introduces SSA, a novel human evaluation metric for multi-turn conversations.

Findings

Strong correlation between perplexity and SSA scores.

Meena achieves 72% SSA on multi-turn evaluation.

Full Meena scores 79% SSA, outperforming existing chatbots.

Abstract

We present Meena, a multi-turn open-domain chatbot trained end-to-end on data mined and filtered from public domain social media conversations. This 2.6B parameter neural network is simply trained to minimize perplexity of the next token. We also propose a human evaluation metric called Sensibleness and Specificity Average (SSA), which captures key elements of a human-like multi-turn conversation. Our experiments show strong correlation between perplexity and SSA. The fact that the best perplexity end-to-end trained Meena scores high on SSA (72% on multi-turn evaluation) suggests that a human-level SSA of 86% is potentially within reach if we can better optimize perplexity. Additionally, the full version of Meena (with a filtering mechanism and tuned decoding) scores 79% SSA, 23% higher in absolute SSA than the existing chatbots we evaluated.