MSRA周明：对话机器人的关键技术_9616811

2020-02-27 58浏览

  • 1.Three Key Engines of A Conversational Bot Dr. Ming Zhou Principal Researcher Microsoft Research Asia Aug. 26, 2016 mingzhou@microsoft.com
  • 2.Xiaoice, Rinna and Tay NL chat •Single-turn and multi-turn •Passive and proactive •Speech enabled Imagebased chat •Image search and recognition •Face detection, beauty score and age detection •Image commenting Skills •Reminder, translation, mathematics •Joking, role play •Count sheep, fortune-telling, horoscope • Weichat, Weibo, Win10 June, 2014 Aug. • Line and Twitter 2015 Late March 2016 • Groupme, Kik, Twitter
  • 3.General Architecture of Chatbot User Profiling 1 Context Modeling 2 3 4 Query Understanding Candidate Response Generation Ranking Intent, focus, topic, emotion, opinion Chit-chat, QnA and dialogue Session consistency General Topic Knowledge Base Data & Index 5 Selection Personalization Style Variation Language style Dialect Emoticon Topic-Centered Knowledge Base
  • 4.Three Key Engines TASK TaskCOMPLETION Completion Layer 3 Information and Answer Layer 2 Social Chat (chit-chat) Layer 1
  • 5.Engine 1: Social Chat
  • 6.Personalized Chat Sleep pattern Horoscope Interest
  • 7.Deep-Chat With Chat Knowledge Without Chat Knowledge Do you know EXO? Do you know EXO? I do not want to see Kris in China because he has let EXO What? Right! Tell me something about EXO Because of Kris, I will no longer be a fan of EXO You were a fan of EXO? Tell what …… I am an audience in every concert of EXO Who do you like best in EXO? I like LAY best. Me too.
  • 8.Image-Based Chat Duplicate Measured by local feature 10B Image index from BING Mona Lisa, Reproduction Grandmother by Leonard… of “Mona Lisa” and Mona Lisa 这不是Mona Lisa吗？ Similar Measured by DNN feature 9M Image index from SNS Isn’t this Mona Lisa? Recognition Via DNN 10K Categories Mona Lisa
  • 9.Image-Based Chat Duplicate Measured by local feature 10B Image index from BING 这小舌头。。。 Similar Measured by DNN feature 9M Image index from SNS 这小舌头。。。 Look at the tiny tongue Recognition Via DNN 10K Categories
  • 10.Image-Based Chat Duplicate Measured by local feature 10B Image index from BING 看得我都流口水了 Similar Measured by DNN feature 9M Image index from SNS My mouth is watering Recognition Via DNN 10K Categories Food
  • 11.
  • 12.Engine 2: Information and Answer
  • 13.Botification of Search Results Query processing Today, Beijing is sunny, 64 degree. Query type, query focus Bing Documents, answers and cards Converting into NL expression Summarization, entity extraction, card to NL TechFest is Microsoft Research annual event… Reinforcement learning is an area of machine learning… Leonardo DiCaprio
  • 14.QnA Based on Multi-Intelligences Question Understanding Planner Knowledge Q/A Web Q/A MSRA .KS Ensemble Engine Answers with confidence and evidences Social Q/A
  • 15.Engine 3: Task Completion
  • 16.Dialogue System • Enabling both card-based or chat-based Option-1 Option-2 Trigger a form/URL to complete the task Trigger a dialogue to complete the task
  • 17.Dialogue Process Hi there~ Chit-chat Response Intent=Nil So how are u doing? I’m fine. Please reserve a trip to Seattle for me. Do you have any preferred hotel? Intent=Book Travel Package; Destination=Seattle; Hilton, please. I will check-in on 2015-10-01, and stay there 3 days Hotel=Hilton; Check-in Date=2015-1001; Length of Stay=3 Days; Dialogue Response Room Type=Single Room Dialogue Response Which type of room do you like? Single room is just ok to me. Got it, I will recommend you some travel packages now. Dialogue Response Slot Name Slot Value Intent Book Travel Package Destination Seattle Hotel Hilton Check-in Date Length of Stay Room Type 2015-10-01 3 Days Single Room Dialogue Management Query Understanding • • Intent detection Slot value extraction Response Generation • • • • • Call services Call chit-chat Intent changes leads to state transition Slot info update Select next slot to ask
  • 18.Applications Shopping Guider Robot Brain Customer Service
  • 19.
  • 20.
  • 21.
  • 22.Robot Brain
  • 23.Customer Service
  • 24.Customer Service (HI) Unsatisfied  Satisfied
  • 25.Research Opportunities • Productivity of knowledge engineering • Learning from existing data to build hierarchical knowledge base • Powerful tool for human editing and updating • Context-sensitive understanding and responding • Better context modelling about current topic and task • Allow various styles of user’s input to map into intent space • Semantic computing (distance, entailment) • Embedding of word, entity, sentence • Lexical semantic relation (synonym, hyponym, hypernym, distance)
  • 26.Thanks!