Neural Generation Meets Real People: Towards Emotionally Engaging Mixed-Initiative Conversations
Ashwin Paranjape, Abigail See, Kathleen Kenealy, Haojun Li, Amelia Hardy, Peng Qi, Kaushik Ram Sadagopan, Nguyet Minh Phu, Dilara Soylu, Christopher D. Manning Stanford NLP {ashwinpp,abisee,kkenealy,haojun,ahardy,pengqi, equal contribution
Abstract
We present Chirpy Cardinal, an open-domain dialogue agent, as a research platform for the 2019 Alexa Prize competition. Building an open-domain socialbot that talks to real people is challenging – such a system must meet multiple user expectations such as broad world knowledge, conversational style, and emotional connection. Our socialbot engages users on their terms – prioritizing their interests, feelings and autonomy. As a result, our socialbot provides a responsive, personalized user experience, capable of talking knowledgeably about a wide variety of topics, as well as chatting empathetically about ordinary life. Neural generation plays a key role in achieving these goals, providing the backbone for our conversational and emotional tone. At the end of the competition, Chirpy Cardinal progressed to the finals with an average rating of 3.6/5.0, a median conversation duration of 2 minutes 16 seconds, and a 90 ${}^{\text{th}}$ percentile duration of over 12 minutes.
中文速览
开放域闲聊机器人（socialbot）需要同时应对海量话题知识、自然对话风格和情感共鸣三大挑战，现有系统往往只能在其中一两项上取得突破。Chirpy Cardinal 的研究团队通过将实体追踪、导航意图分类器、多个模块化响应生成器（Response Generator）与 GPT-2 神经生成模型有机组合，让机器人既能沿预设对话路径深入讨论维基百科、Reddit 等来源的世界知识，又能以神经模型生成流畅、感同身受的日常情感回应。在 2019 年 Alexa Prize 决赛中，该系统获得 3.6/5.0 的用户评分，中位对话时长达 2 分 16 秒，前 10% 用户的对话甚至超过 12 分钟。这项工作表明，将神经生成与符号化对话管理相结合，是实现"用户主导、内容丰富、情感真实"三者兼顾的有效路径，为开放域社交对话系统的后续研究提供了可复现的工程范本。
原文 arXiv:2008.12348;中英对照 + 大白话阅读 https://aha.fim.ai/paper/2008.12348v2