conversational ai from cat i to iii

14
convospot Conversational AI from CAT I to III @homerquan [email protected]

Upload: huangmaohomer-quan

Post on 12-Apr-2017

133 views

Category:

Technology


0 download

TRANSCRIPT

Page 1: Conversational AI  from CAT I to III

convospot

Conversational AI from CAT I to III

@[email protected]

Page 2: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

Conversational AI

Cognition

Knowledge

LanguageUnderstanding

LanguageSynthetic

• How to model knowledge and cognition • How to get and update those models

Page 3: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

Rule-based chatbot

Page 4: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

How voice assistant works

Source: http://www.macworld.com/article/3067407/ios/meet-viv-the-new-voice-assistant-from-the-creators-of-siri.html

Entity TaggingNLP + Query plan

Pattern match

Page 5: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

Knowledge from data sources

Source: Google Knowledge Graph Search

Dynamic info

Page 6: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

More or less UX tricks

Source: Cheat Sheet: All Facebook Chatbot Interactions - chatbot magazine

Combined with ML improvement in NLPThey are already very interesting and powerful! e.s.p Alexa’s voice understanding

Page 7: Conversational AI  from CAT I to III

Challenges

• How to easily acquire cognition and knowledge

• Knowledge with reasoning

• Limited knowledge, new concept

Page 8: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

CAT I,II,III

• CAT I Understand language commands e.g., Siri, Viv, Alexa

• CAT II Natural interaction with portable knowledge and memory (knowledge)

• CAT III Reasoning and acquiring new knowledge proactively (knowledge & cognition)

Page 9: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

Universal Language Model

• A Neural Knowledge Language Model

• Merge of graph model and probabilistic model to represent knowledges in our language. e.g., Composing graphical models with neural networks for structured representations and fast inference

Page 10: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

Method of learning structured behavior

• Learning alone timeline: Long short-term memory (LSTM)

• Dynamic Memory: Memory Network

• Demo: http://yerevann.com/dmn-ui/#/

Page 11: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

Practices of deep learning• Trained by raw data (end-to-end)

(*But NLP is still necessary)

• Using fast inference to update the model

• Invert Reinforce Learning (IRL) for apprenticeship learning (convospot.io)

Page 12: Conversational AI  from CAT I to III

https://goo.gl/7K05nf

Reference papers

Page 13: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

• A channel to deliver conversational AI • A platform to digitalize knowledges

Page 14: Conversational AI  from CAT I to III

©Reflen Inc 2017 - confidential and property

Team

Homer Quan Founder and CEO

www.homerquan.me

Qiang LiuMentor on AI

www.cs.dartmouth.edu/~qliu/

Techstars Alumni (2015)

Yahoo Search Entity GraphYahoo NewsYahoo E-commerceYahoo Local

Assistant Professor

Postdoc