2. A little bit about LexisNexis Legal & Professional
• LexisNexis Legal & Professional is a leading global provider of legal regulatory
and business information and analytics that helps customers increase productivity,
improve decision-making and outcomes
• 10,700 employees in global and customers in >130 countries
Our Mission:
• Providing opportunity for easy-to-access and
accurate law & reporting of injustices anytime
anywhere
• Innovating to support each customer's success
• Combining information & analytics to help
customers make more informed decisions & achieve
better outcomes
3. A little bit about myself
• Min Chen is responsible for driving and delivering customer-centric legal
innovations across 7 markets in Asia Pacific through artificial intelligent
technologies as well as data analytics and visualization.
• Leading 230 engineers and product people in Asia Pacific, main teams are in
Shanghai (124 HCs) and India (81 HCs)
My Passion:
AI-powered
SEARCH
USER
EXPERIENCE
Machine Reading
Comprehension
4. A little bit about Shanghai Tech Center
Raleigh
~ 385, 2019
London
~ 103, 2019
Shanghai
~ 100, 2019
LexisNexis 3 Global Technology Centers FTE number as of Jan, 2019
Legal Research & Data Analytics
LexisNexis Microsoft Office
LexisNexis Information Offline e-Reader
AsOne Manager – Compliance Workflow
Augmented Reality & Machine Learning Powered Solutions
Main expertise in Shanghai:
• Search & AI (NLP, DNN, ML)
• Data analytics, mining, visualization
• Mobility, drafting, and workflow solution
5. A few examples of projects by leveraging
Artificial Intelligence
• Legal information is very often complex, professional and lengthy text-based
documents. Given this specialty, our expertise and focus are particularly on
NLU (Natural Language Understanding) for machine reading
comprehension and NLG (Natural Language Generation) for
computational “text (written or spoken)” producing.
• The typical use case we are working on are: semantic search,
recommendation engine, document auto-summary, key information
extraction, context connection and etc.
6. Case Study 1: when AI meets Search
Semantic search, which focuses on understanding
user intent, is the solution to improve customer
satisfaction.
Semantic Search employs AI technologies such as
natural language processing and machine
learning to understand and organize data, predict
the intent of the search query, improve relevancy
of results, and automatically tune the relevancy of
results over time.
Problem: Customers are searching a “thing”, not a “string”. Besides, our users’ search expectation
continues to increase due to the rapid evolution of online search world (like Google) – return the most
desirable result on the first page in the first shot
Let’s see a sample of one
small component of semantic
search – semantic parsing !
7. 7LexisNexis Confidential
Before “Semantic Parsing” – Sample
Before this release: search
engine & algorithm either
parses query by words, or
by incorrect segments
which are not in semantic
units in legal domain. For
this example, the query is
parsed by words
search results related to
individual strings vs. user
intent, leading user to have
a negative experience. For
this example, top result
contains a lot “directors” or
“liability” which is not
relevant at all
User searches for
“limitation of liability for directors”
Traditional query parsing breaks this query into
“limitation”, “of”, “liability”, “for”, “directors”
Lexis Advance
AU example
8. 8LexisNexis Confidential
After boosting “Semantic Parsing” – Sample
User searches for
“limitation of liability for directors”
Lexis Advance
AU example
Semantic query parsing breaks this query by semantic units of:
“limitation of liability” for “directors”
After this release
The most relevant AM (Analytics
Material) document rendered at top of
the result due to semantic parsing that
limitation of liability is considered as
one phase and directors is treated as
another segment
Solution:
• Parse the query by semantic unit
in legal domain
• Build a legal domain phrase
dictionary by extracting qualified
phrases out of LexisNexis
documents & user queries via
machine learning and data
mining techniques
9. 9LexisNexis Confidential
Learning to Rank – Lexis AsOne
Training Data
(Usage log)
Learning algorithm
Metrics (A/B testing)
LTR Model
Solr Index
Learning to rank is a machine learning approach to build relevance model based on user data
with different features in order to return sophisticated ranking result.
10. LexisNexis Confidential
Identify “Adjacent Cases” which
shared similar legal principles/issues
by leveraging deep learning
Case Study 2: Concept embedding for recommendation
Problem:
1. You can’t search something you
haven’t thought about !
• Understand the breadth of tactical
options by going beyond legal issues
outside a linear chain of citing cases
2. Recommendation based on
similar legal principles is hard
even by human !
• No existing legal issue/principle
category assigned to case documents
• No rule-based pattern to define “legal
principles” or “legal issues” of cases as
legal principle is a concept
Customers:
Lawyers (any segment), Knowledge Managers & Researchers (including clerks
and paralegals), Students, or any persona conducting legal research
11. 11LexisNexis Confidential
Solution – Concept embedding deep learning
Deep Neural
Network
Online Prediction
Learn
Predict
Offline Training Process
Legal Issue
Embedding
Similarity
comparison
Near Real Time Indexing
Case Doc
Share Similar
Legal Principle?
Case Doc
Release
Trained Network
• Train up a deep neural network model to get legal principle embedding (Feeding 191,291 cases with 1,095,047
citations into this model). Every case is converted into a legal issue vector space to represent legal principle.
Hence, cases with similar legal principle should be clustered together
• Such technique can be used in many scenarios such as recommendation, search & analytics solutions
12. 12LexisNexis Confidential
Sample – Recommended cases by AI
Source case
Intellectual property — Trade marks — Opposition — Appeal
Winton Shire Council v Lomas (2002) 119 FCR 416|(2002) 56 IPR 72|(2002)
AIPC 91-794|[2002] FCA 288|BC200200989
Appeal from unsuccessful opposition to registration of trade mark 'Waltzing
Matilda’ in respect of various goods and services.
Opponents are operators of tourist attraction in Winton known as 'The
Waltzing Matilda Centre'.
Evidence establishes that 'Waltzing Matilda' was composed near Winton.
Applicants claim to be first to use trade mark.
Whether trade mark is likely to deceive or cause confusion.
Whether was a use prior to priority date of 'Waltzing Matilda' in respect of
goods and services identified in opposed application.
Whether applicant is owner of trade mark.
No Name Restaurants (Cesare) Pty Ltd v No Name Restaurants Pty Ltd (1996) 36
IPR 488|(1997) AIPC 91-301
Intellectual property — Trade marks — Registration of trademark — Honest
concurrent use
Whether should permit registration of trade marks NO NAMES notwithstanding
substantially identical or deceptively similar.
Where applicant and opponent originally partners carrying on restaurant business under
name NO NAMES. Where applicant and opponent subsequently decided to operate two
restaurants independently, but did not decide on use of name.
Consideration of quantum of concurrent user: size of reputation of mark in respect of
restaurant services. Where parties conducted restaurants under same name independently
for some years.
Consideration of inconvenience to applicant and loss of goodwill if registration refused and
applicant required to adopt another trade mark.
Highly relevant case found by AI
Rated by legal experts in Australia: We are now able to get AI to recommend complex cases which are difficult for human!!
1. The source case of ‘Winton Shire Council v Lomas’ (on the left side) has 4 different aspects in sense of legal principles: Legal
Issue (trade mark, prior use); Legal Argument (First to use trade mark and Whether like to deceive); Legal Principles
(Substantially identical or similar to, and weight to be given to the opinion of the delegate use, prior to priority date of
mark); Motion (Opposition to registration of trademark)
2. Those 4 types of principles are identified by human and NOT tagged in the document
3. AI recommended case (on the right side) covered these 4 types of principles. According to the rater “this case shows a
really different way to tackle/reason through the same legal issue, and therefore highly relevant ”
13. 13LexisNexis Confidential
Case Study 3: Cognitive automation – auto summary
US case auto summary is a project aiming to massively improve editorial efficiency and content coverage with
millions of saving via deep learning to understand natural language(NLU) and generate summary
language(NLG) based on court case opinion. Underneath this project, achievement of 4 key NLP tasks can be
leveraged for other projects
Key information extraction
Find & extract most relevant information
which can represent the document for
different tasks.
Natural language generation (NLG)
Generate natural language such as summary,
overview, outcome and other tasks by machine
Auto labeling
Automated training data creation which can
save human efforts and reduce
dependencies with expensive SME resource.
Factorize joint inference
Large scale taxonomy classification for scenarios
with tens of thousands of categories in data.
Re-usable NLP tasks for other projects State-of-act technology underpinning these tasks
BERT/MT-DNN
Generate baseline and benchmark for our solutions
and improve the performance of existing solutions
Transfer learning
Fine-tuned word contextual embedding to enhance
the performance in legal domain
Graph Neural Network (Future)
Use structured priori knowledge in knowledge graphs
and to improve performance on different tasks.
Reinforcement Learning/GAN (Future)
Use a generative network to generate natural language
and use a discriminative network to evaluate the quality
14. Asia Legal AI will become predictive and prescriptive
15. Decision-making tool for penalties/damages (1 of 2)
Lawyers would
want to know the
Penalties History
issued by
regulator, with
relevant case list
showed on the
bottom of the
page in an
interactive way
Lawyers wants to know what sort of penalties have been handed down in the past; whether the penalties have
historically been higher or lower if the person has contested them in court instead of just paying up front etc.
Descriptive analytics
16. Lawyers would love to know estimated penalties with max and min range, which
are generated based on previous judgments that are relevant to given information.
Lawyers could refer to it, to challenge regulatory decisions they want to take to court.
Lawyers would like to best estimate the end result for their clients’ cases before they go to trial, particularly in
relation to penalty or damages. This would help lawyers make the right strategic call in advance: whether it’s
worth proceeding with the case, or better to negotiate a settlement or plea deal.
It would also be
helpful to lawyers to
see top 5 relevant
categories and how
they influence
estimated penalties,
on the left. This will
help lawyers to
advise their clients
how to proceed –
whether they should
pay the fine, or
contest it, and take
other actions.
Decision-making tool for penalties/damages (2 of 2)
Predictive analytics