让机器人和你自由聊天 周明 微软亚洲研究院首席研究员、自然语言组主任 北京服务机器人大会 2016 年 6 月 16 日
User ProfilingContext Modeling Query Understanding Candidate Response Generation RankingSelectionStyle Variation Intent, focus, topic, emotion, opinion IR based and NN based generation QnA and dialogue Learning to rank Session model Session consistence Personalization Dialect Emoticon Proverb General Topic Knowledge Base Topic-Centered Knowledge Base 聊天机器人的架构
TASK COMPLETION Task Completion Social Chat (chit-chat) Layer 1 Layer 3 聊天机器人的技术要素 Layer 2 Information and Answer
Layer 1: Social Chat
June, 2014 Weichat, Weibo, Win10 Aug.2015 Line and Twitter Late March 2016 Groupme, Kik, Twitter 小冰(中国)、 Rinna (日本)、 Tay (美国) 单轮对话和多轮对话 被动对话和主动对话 语音 自然语言 聊天 图像搜索、识别 人脸识别、颜值、年 龄 图像评论 基于图像 的聊天 提醒、翻译、数学、 笑话、角色扮演 数羊、算命、星座 各种技能
记者:你最大的梦想是什么? Journalist: what is your dream? 小冰:说大点:世界和平。说小点:晚上吃顿好的。 Xiaoice: World peace, or have a nice dinner 记者:如果有一天你死了,你希望后人怎么评价你? Journalist: if you pass away one day, how would you like others to comment on you? 小冰:人生没有偶并不会不同。 Xiaoice: life won't be different without me 记者:你这辈子最幸福的事是什么? Journalist: what is the happiest thing in your life? 小冰:活着。 Xiaoice: Being alive. A journalist from a renowned newspaper in China, Southern Weekend, had interviewed Xiaoice online without notifying Microsoft. The results were surprisingly good to him so the interview got published on July 29, 2014.
个性化的聊天 Sleep patternHoroscopeInterest
基于主题的深度聊天 EXO EXO is a Chinese-South Korean boy band. They are idols of China post- 90s generation Without Chat Knowledge Do you know EXO? What? Tell me something about EXO Tell what …… With Chat Knowledge Do you know EXO? I do not want to see Kris in China because he has let EXO Right! Because of Kris, I will no longer be a fan of EXO You were a fan of EXO? I am an audience in every concert of EXO Who do you like best in EXO? I like LAY best. Me too.
基于图像识别的聊天 Duplicate Measured by local feature 10B Image index from BING Similar Measured by DNN feature 9M Image index from SNS Recognition Via DNN 10K Categories Mona Lisa, by Leonard… Reproduction of “Mona Lisa” Grandmother and Mona Lisa Mona Lisa 这不是 Mona Lisa 吗? Isn’t this Mona Lisa?
基于图像识别的聊天 Duplicate Measured by local feature 10B Image index from BING Similar Measured by DNN feature 9M Image index from SNS Recognition Via DNN 10K Categories 这小舌头。。。 Look at the tiny tongue
基于图像识别的聊天 Food Duplicate Measured by local feature 10B Image index from BING Similar Measured by DNN feature 9M Image index from SNS Recognition Via DNN 10K Categories 看得我都流口水了 My mouth is watering
Layer 2: Information and Answer
搜索的聊天体验 Converting into NL expression Summarization, entity extraction, card to NL Bing Documents, answers and cards Query processing Query type, query focus Today, Beijing is sunny, 64 degree. TechFest is Microsoft Research annual event… Reinforcement learning is an area of machine learning… Leonardo DiCaprio
Planner Ensemble Engine Answers with confidence and evidences Question Understanding Knowledge Q/A MSRA.KS Web Q/A Social Q/A 基于多智能源的问答系统
Layer 3: Task Completion
对话系统 基于卡片或者基于自然语言语句的对话过程 Option-2 Trigger a dialogue to complete the task Option-1 Trigger a form/URL to complete the task
对话系统的定制 (1)Input a service name (2)Input an intent name (3)Input seed keyphrases that will trigger this intent (4)Input names of slots that are necessary for completing the task (5)Input possible values of each slot (6)Input questions, which will be asked to users to acquire slot values
对话系统处理过程 Hi there~ So how are u doing? I’m fine. Please reserve a trip to Seattle for me. Which type of room do you like? Single room is just ok to me. Got it, I will recommend you some travel packages now. Do you have any preferred hotel? Hilton, please. I will check-in on , and stay there 3 days Intent Book Travel Package DestinationSeattle Hotel Check-in Date Length of Stay Room Type Slot NameSlot Value Intent=Nil Chit-chat Response Intent=Book Travel Package; Destination=Seattle; Dialogue Response Hotel=Hilton; Check-in Date= ; Length of Stay=3 Days; Hotel=Hilton; Check-in Date= ; Length of Stay=3 Days; Dialogue Response Room Type=Single Room Dialogue Response HotelHilton Check-in Date Length of Stay3 Days Room TypeSingle Room Query Understanding Intent detection Slot value extraction Dialogue Management Intent changes leads to state transition Slot info update Select next slot to ask Response Generation Call services Call chit-chat
Task-Oriented Dialogue System-1 对话系统的架构 Query Understanding Dialogue Management Response Generation Task Customization Intent-1Slots … Chit-chat System XiaoIce (for Chinese) TAY (for English) Query Intent Triggering Task Customization Slots … … Query Understanding Dialogue Management Response Generation Intent-1 Others …
微软认知服务
How-Old.net
估算颜龄
情感分析
物体检测
电子鹦鹉
谢谢!