AI, MT and Language Processing Symposium2

Transcription

1 1

2 Speaker Overview Dr. Anthony Scriffignano Senior Vice President & Chief Data Scientist Dun & Bradstreet Dr. Anthony Scriffignano, Senior Vice President and Chief Data Scientist for Dun & Bradstreet, is an internationally recognized thought leader in the data science space. He leads a team of data scientists focused on advancing Dun & Bradstreet's core capabilities and IP globally. With extensive background in advanced algorithms and linguistics, he holds multiple patents and presents globally on data and technology trends, multilingual challenges in business identity, and artificial intelligence. Warwick Matthews Senior Director of Identity Data Engineering Dun & Bradstreet Warwick is Senior Director of Identity Data Engineering at Dun & Bradstreet. Based in Melbourne Australia, his work focuses largely on creating complex cross-border multilingual data flows. AI, MT and Language Processing Symposium2

3 Keynote: Confounding Characteristics and Resolution of Complex Business Identity in a Multilingual Context The presentation has three major themes: More Better Faster Technology and Decision making Risk and Response Disruptive Evolution, Malfeasance and how we will Respond The Future is Here Quantum Computing, Machine Intelligence, New Mindsets, Recommendations for the future Dr. Anthony Scriffignano Senior Vice President & Chief Data Scientist Dun & Bradstreet Warwick Matthews Senior Director of Identity Data Engineering Dun & Bradstreet AI, MT and Language Processing Symposium3

4 Confounding Characteristics and Resolution of Complex Business Identity in a Multilingual Context AI, MT AND LANGUAGE PROCESSING SYMPOSIUM 28 MARCH 2018 Anthony J. Scriffignano, Ph.D. SVP / Chief Data Scientist Warwick Matthews Senior Director, Identity Data Engineering

5 OUR CURIOUS WORLD M O R E T E C H N O L O G Y, F A S T E R D E C I S I O N S T O D AY COMPUTATIONAL LINGUISTIC CHALLENGES R O M A N I Z A T I O N O F B U S I N E S S I D E N T I T Y D A T A THE RISKS AND OUR RESPONSE D I S R U P T I V E E V O L U T I O N A N D H O W W E W I L L R E S P O N D THE FUTURE IS HERE T R E N D S, D E V E L O P M E N T S, C A S E S T I U D I E S

6 OUR CURIOUS WORLD M O R E T E C H N O L O G Y, F A S T E R D E C I S I O N S

7 We live in an age of promise. Advanced linguistic methods are making things possible that were science fiction only a few short years ago.

8 Changes in our environment have continuously influenced business decisionmaking. What is changing is the speed and degree of globalization. 1 1 Hypergeometric Data Growth 2 Unstructured Data 3 Globalization Challenges 4 Virtual Businesses Hypergeometric digital data growth can make it more difficult to determine what is valuable vs. noise It is estimated percent of all business information exists as unstructured data The globalization of business relationships can overwhelm many businesses with multilingual data A business geographic location, structure, and physical customer interaction are becoming irrelevant The importance of language should not be underestimated. Language contains nuance, changes constantly, and informs our thinking on levels that we often rely upon in subtle and powerful ways. A. Scriffignano Note: Unstructured data includes data which lacks (an exposed ontology) and which appears to belie attempts to understand any implied categorization. The term is often used where the content is not actually unstructured, but rather only poorly understood at the time of ingestion or inspection 8

9 In business today, we will hone our skills for using dynamic, unstructured data, or we will begin to drown in it. There is no guarantee that things get better. Unstructured Data Situational awareness More than 85% of data creation is unstructured Language and use of language are constantly evolving Commonly available tools and solutions only address a small part of this space 9

10 Machine Translation is good enough Many unmet challenges remain in linguistics Myths / Inconvenient Truths: More Data is better Lots of data in 1 place is sufficient to learn AI can find answers Machine learning will find hidden truth Natural language processing removes all language barriers Data vs. noise Data at rest vs. data in motion AI methods have preconditions Regression vs. unprecedented change Language is constantly changing

11 COMPUTATIONAL LINGUISTIC CHALLENGES R O M A N I Z A T I O N O F B U S I N E S S I D E N T I T Y D A T A

12 Today s problem is not a lack of data.. Dun & Bradstreet is continually acquiring large amounts of non-latin data, particularly in Asia, which needs to be Romanized in order to enter our Global Data Supply Chain. Translation is traditionally largely manual, time consuming and very expensive when you have millions of records to process. When it comes to Romanizing pure Identity Data we have a special problem. Name and Address data has no context, and this is particularly challenging when we are talking about new business entities who have not existing footprint to work from. And we need to solve this problem millions of times per day, automatically. TITLE OF PRESENTATION (EDIT USING: INSERT MENU > HEADER AND FOOTER) 12

13 A D&B Challenge: Romanization of Business Identity Data Business Names are a unique mix of Translated and Transliterated (phonetic) data. And this data has no context (there is nothing around a name to guide the system) Addresses are easy to translate for big geos like Cities. But address detail is often vague, idiosyncratic and defies systematic classification (especially in China!). 13

14 Why don t we just A.I. our way out of the problem?

15 AI has limitations.. AI is a grey box at best difficult to get actionable qualitative feedback to aid in automated decision-making It does not actually understand what it is doing, so it cannot tell when its output is nonsense. inspirobot.com Todai Robot Microsoft Tay 15

16 ..none of the modern AIs, including Watson, Siri and Todai Robot, is able to read.....it doesn't understand any meaning. Dr Noriko Arai, Tokyo University

17 So we must leverage multiple simultaneous approaches.. AI SHEN/X Lexicon/Stats-based routines Decisioning system Machine Learning System UI Human Adjudication

18 THE RISKS AND OUR RESPONSE D I S R U P T I V E E V O L U T I O N A N D H O W W E W I L L R E S P O N D

19 There are three use cases which represent True North for our innovation Organic growth/decay Discovering new businesses or changes in business status New business that would otherwise go undetected Additional input for Intelligence Engine to transform a Single Source Record (SSR) Full-file maintenance (e.g. Out of Business) Fraud / Malfeasance Ingesting information that can be used to detect bad behavior Common social footprint shared by more than one persona (e.g. identity theft) Patterns of observation that suggest clusters of bad behavior (e.g. fraud rings) Discovery of new types of malfeasance (e.g. Phishing) Personal Identity Discovering data elements that can help resolve people in business context New Social Media handles or sources of social data (e.g. opinion blogs) Data which can be aggregated to pre-existing clusters of identity (e.g. photos) for additional resolution (e.g. hyperclusters) Data about groups of individuals associated with a business context (e.g. user groups, discussion boards) 19

20 Areas of focus Entity Extraction Existing capabilities include extraction of entities, tokenization, part of speech tagging, usage vectors, language detection Understanding Person and Perspective Multiple patents including identity resolution, people in the context of business, geospatial inference, flexible alternative indicia Sentiment Attribution/ Clustering The current state of art is highly dependent on training and has challenges with precision and recall, reproducible results Multiple interpretations. Becomes Proper Noun due to inference about verb. The political party? The fruit? Apple destroys competition Яблоко пережило несколько стадий развития Transitive verb, requires actor. The company? Changing regulatory environment globally 20

21 FLOCCINAUCINIHILIPILIFICATION

22 Confounding characteristics USE CASES CONFOUNDING CHARACTERISTICS DERIVING EMPIRICAL MEASURES THAT INFORM USE CASES Entity Extraction Sentiment Attribution Context / Behavior Sarcasm ABC corporation is a wonderful company, if you don t do business with them. Neologism Be sure to like us on FaceBook and use #shallow when you Tweet. Passive metric Grammar variations FBI is Hunting Terrorists With Explosives. Punctuation Hi mom! vs. Hi, mom? Intentional mis-spelling RU There? 22

23 Positioning Confounding Characteristics in the Curation Process Recursive Discovery Entity Extraction Vetting/ Adjudication Sarcasm Words or predicates juxtaposed in such a way as to convey hidden meaning that is opposed to that which comes from cursory interpretation Example: BP is an excellent company to do business with, if you like destroying nature. Neologism Words or phrases which are newly constructed and taken collectively to have some shared meaning. Example: - Hashtags in Twitter Grammar variations Word usage which is intentionally or unintentionally incorrect, leading to ambiguous or non-dispositive interpretation Example: FBI is Hunting Terrorists With Explosives Punctuation Usage of punctuation in a non-standard or inconsistent way or lack of punctuation, leading to ambiguous or contradictory interpretation Example: Eats shoots and leaves vs. Eats, shoots, and leaves Intentional Mis-Spelling Invented, incorrect, or adopted spelling that results in inconsistent, incorrect, or non-dispositive interpretation Example: RU There? Mixing of Languages or Scripts Including foreign words/phrases or characters, especially in non-standard ways. Examples: He has a certain je ne sais quoi or Please have some Synthesis Who is speaking? About whom? How do they feel? In what context? 23

24 THE FUTURE IS HERE T R E N D S, D E V E L O P M E N T S, C A S E S T I U D I E S

25 Depth Assessing capabilities in language synthesis for identified use cases Understand dependencies across sources Analyze relative usefulness of new sources Analyzing impact on specific use cases Using source metadata for insights Semantic vector space models Semantic disambiuation Understanding Person & Perspective Entity Extraction Sentiment Analysis Translating scores into degree of text confoundedness Detecting & Measuring scores of Confounding Language Characteristics Create dedicated & scalable infrastructure for unstructured data Breadth Detecting additional confounding factors (e.g. foreign language) Improving robustness of existing detection algorithms 25

26 Reality check 26

27 Watch this space Inter- and intra-language correlation : deciding when things mean the same thing Inter- and intra-language transformation : transforming inference among languages Changing behavior to attract/obviate grapheme analysis : reacting to changing language Emerging metalanguage (e.g. textspeak ) : reacting to language about language A language of things : reacting to new languages used by automation Using language to hide language : reacting to attempts to obscure via language Unicode is not universal : understanding the limitations of automation 27

28 The journey of discovery involves asking new questions? What is the community at large saying about this business? How do I understand and measure customer sentiment? Can I trust the social data on my partners? How are opinions changing over time? Are they authentic? How can I see birth and death of a business more quickly? How can I detect inconsistent behavior? What does it mean? What about modes: Is there a measurable difference between leaders and their organizations' opinions? 28

29 Warwick Matthews Thank You! 謝謝 Dankjewel merci ありがとう धन यव द Anthony Scriffignano 29

30 30