A.I. and Translation iflytek Research : Gao Jianqing 11-2017
1. Introduction of iflytek and A.I. 2. Application of A.I. in Translation
Company Overview Founded in 1999 A leading IT Enterprise in China The largest public AI & Speech Technology company in Asianpacific Region Over 9000 employee, over 5.5 Billion US dollars market value
Introduction to A.I. Artificial Intelligence A system or machine that can percept, make decisions and think like human beings On Mar 9-16,AlphaGo defeated Lee Sedol Century 5 By 4-1 MIT: 45% activities can be automated by current technology CEO, 20% of the work may be substituted with a robot The McKinsey Global Institute Report: How AI redefine work AI comes faster than you could imagine
iflytek Key Technologies Three levels of Artificial Intelligence Computational Intelligence Compute Perceptive Intelligence Listen Speak Recognize Cognitive Intelligence Understand Think
Intelligent Speech Technology Intelligent Speech Tech:Enabling machines to listen & speak, including speech synthesis, speech recognition, etc. Speech synthesis Converting text to speech Installing artificial mouth for computers Speech recognition Converting speech to text and recognize speech content, speaker, language and other information Installing artificial ear for computers Cognitive intelligence Understanding and Thinking Speech evaluation, Machine translation, Smart Customer Service
Perceptive Intelligence Research Achievement Machines pronounce better than ordinary human speakers 5 4 自 3 然度 2 4.7 Natural Voice 4.2 IFLYTEK:The only system which exceeds 4 in naturalness 3.6 3.5 3.1 2.5 2.3 2.3 2.3 1.7 1.5 Blizzard Challenge championship for consecutive 11 years English language speech synthesis exceeded ordinary people in naturalness for the first time in 2012 Blizzard challenge 1 0 A C I F B G D H J K E Other participants: Microsoft, IBM, MIT, CMU in the United States Edinburgh in England, NITech, ATR, etc. in Japan IFLYTEK won the Blizzard Challenge again in 2016
Progress on Automatic Speech Recognition (ASR) 2011-8 Large Scale Language LM 2011-1 Updating based on real data 2013-7 RNN LSTM Model 81.2% 92% 86.2% 95%+ 2012-5 DNN Model 2016-10 DFCNN Model 75.3% 60.2% 2010-10-28 Speech Cloud
Recent progress on ASR 1 st Place of 2016 CHiME Challenge WER(Six Microphones) WER(Double Microphones) 14 12 10 8 6 4 2.24 2.91 2.98 3.48 4.31 4.68 5 5.69 6.41 6.55 6.75 8.98 10.1 11.52 16 14 12 10 8 6 4 3.91 5.32 5.96 6 7.28 8.32 8.81 9.14 9.64 10.77 12.82 13.69 13.9 2 2 0 0 IFLYTEK B C D E F G H I J K L M N 1 科大讯飞机构 2 机构 3 机构 15 机构 5 机构 7 机构 6 机构 9 机构 8 机构 11 机构 16 机构 12 机构 13
From listen and speak "to understand and think Obama administration promotes "brain activity patterns Program" EU "Human Brain Project" "China Brain Project started at the end of 2015
Cognitive intelligence hotspot - Exam Robot US US Exam Robot Turing Center at University of Washington Objectives: Pass American high school biology exam Japan Japan Todai college entrance exam robot Japan National Institute of Informatics Objective: Admitted to University of Tokyo China China humanoid answering robot IFLYTEK, Tsinghua University, Chinese Academy of Sciences, Beijing University... Objective: Admitted to China Tier-1 universities
IFlyTek Super Brain Project - robots pass college-entrance examination Ministry of Science and Technology launched China first AI major projects " humanoid answering robot" v v v v v IFLYTEK has officially released Super Brain Project on August 20, 2014 Language Understanding Knowledge Presentation Association Reasoning Self-learning Objective: the robot can go to university!
IFlyTek super brain milestones :Machines outperform humans in scoring essays and test papers of all disciplines Scan Examination Paper Handwriting Recognition Automated Essay Scoring Automatic Machine Review OCR Scan Automatic Review
A.I. Open Platform 150% 212% 333% 1.5Billion+ 130k+ 1Billon+ Super Brain FACE 2014 Nov 2016 March ASR Gesture TTS Evaluation Translation User number of speech cloud Daily number of request The third party partner
1. Introduction of iflytek and A.I 2. The Application of A.I. in Translation
Speech Translation ASR Speech Input Speech Output TTS MT
Machine Translation Converting the source language to the target language using a computer Statistical based MT Sematic represe ntation Nueral network based MT Who am I? Decoder Source Syntax SMT Phrase SMT Target Attention Encoder + + + + 我是谁?
BLE Recent Progress on MT Champion of IWLST2014 MT performance (manually scored) 20 15 10 5 iflytek NICT of Japan MIT of America 0 CH-EN HKUST IWSLT has the most academic influence in the machine translation of spoken language Translation materials came from TED recordings and text Oral Edu Pol Meeting Lecture
itrans Application of itrans 123 4.97 Million 1100 Translating Chinese into English Japanese Korea French Spanish Countries Number of services Daily using Running on local device
How will MT help human translators?
Man-Machine Cooperation Definition of man-machine cooperation Examples WordSmith IBM Waston Machine assists mankind:machine help mankind to solve computing-intensive and repetitive problem via AI Mankind helps machine:mankind perform some difficult problems based on the initial result of machine News reports written by machines based on statistics Reporters can write a further analysis report Machine analyzes genetic data and the latest research results to help doctors choose more effective treatment
Trans based on Human-Machine Cooperation iflyrec for conference Cooperation Modes Optimizing the performance of MT based on manual parallel data Assist interpreters in preparation for conferences Research on the UI design of the system
Trans based on Man-Machine Cooperation iflyrec for conference Advantages Improving audience experience Relaxing equipment requirements Assisting interpreting by showing ASR result of source language Evaluating the level of interpreters Transcribe the interpreter s speech into text
Thank you very much!