Closed Set Extraction (CSE) Enav Weinreb Text Metadata Services, Thomson Reuters November 18 th, 2014.

Closed Set Extraction (CSE)Enav WeinrebText Metadata Services, Thomson ReutersNovember 18th, 2014

AGENDA• Thomson Reuters, Text Metadata Services

• Closed Set Extraction– Problem description

– Solution

• Aspects of the solution

• Conclusions and questions

THOMSON RETUERS

NewsBroker ResearchBondsFundamentalsPress Releases

Case LawAdmin DecisionsPublic RecordsDocketsArbitration

Editorial Analysis Scholarly Articles PatentsTrademarksDomain NamesClinical TrialsDrugs

TEXT METADATA SERVICES (TMS)• In charge of extracting metadata from unstructured

contents– Entity extraction and resolution

– Relations, events, and facts

– Topics

• Formerly called ClearForest

• Owner of OpenCalais

• I lead a small applied research team within TMS

COMPANY EXTRACTION

• Task: Identify companies within input text

• Motivation to concentrate on known companies:– Public companies have a special page in TR Eikon

– Better quality for an easier problem

– Faster development

known

CLOSED SET EXTRACTION• Starting point: lexicon with company aliases

• Given an input document:– Find company aliases as candidates (TRIE)

– Filter out noisy candidates using a machine learning layer

EXAMPLES• “China, Apple's third biggest market in the world

behind the US and Europe, has been a hefty cash cow for Apple.” 

• “Apple partisans (a real faction) often claim that the typical American grocer stocks no more than 12 varieties: Red Delicious, Gala, Fuji, et al.”

TRAINING PROCESS• Collect input texts

• Identify company candidates (TRIE)

• Send candidates to manual tagging

• Extract features

• Train a classifier (logistic regression)

• Results for Reuters news documents are of very high quality– F measure of ~96%

HISTORY• Started with a rule based algorithm with

– Good quality, but

– Bad latency

• Requirements: – Under 50ms per document

– Only public companies

We used the rule based algorithm as a data labeler

It turns out that learning from an existing rule based algorithm can have surprising positive results

EXTRACTING CANDIDATES• Lexicon search with a basic TRIE implementation

• Coverage: extending the lexicons using the rule based algorithm– Co mentions with known aliases in a large corpus

• Fuzzy lexicon search?– Vowel free lexicon

– Fuzzy TRIE???

– How to treat fuzzy candidates?

FEATURE ENGINEERING• Positional unigrams and bigrams

– “CEO of” before the instance is a positive feature

– “Mr.” before the instance is a negative feature

– 2-3 tokens before and after the instance give best results

• Company fingerprint – positive and negative– “tennis” is in the negative signature for “Williams Energy”

– “Gas” is in the positive signature for “Williams Energy”

– Here we look in a wider window around the instance

• Company suffix (Inc., Ltd., etc.)

• Part of speech

• More…

CSE

DATA LABELING• Discovery versus CSE

– Searching for companies in text versus answering multiple choice questions

• Ideal for crowd sourcing

• 10 times cheaper than labeling for company discovery algorithms

CSE

CROWD SOURCING EXAMPLEPlease read the following text:

“A high-achiever from a young age, Kim Williams is known as one of our top business leaders. Now, the workaholic's sudden departure as News Corp's CEO has given him time for reflection.”

In the above text, “Williams” is best described as:

o A company name

o Part of a company name

o Neither a company name nor a part of one

o There is not enough information in the text to determine

LOGISTIC REGRESSION VS CRF• Faster improvement iterations

• Large training sets:100Ks of examples labeled by– Our rule based algorithm

– The crowd

• More data or smarter machine learning?

CSE

DOMAIN ADAPTATION MADE EASIER• Domain adaptation is simpler for logistic regression

• Example– Trained a model on Reuters news

– Ran it on research reports to get 77% f-measure

– Adapted a model for research reports

– F-measure improved to 87%

• Main idea: cluster news and research instances together and boost confidence for research instances within positive news clusters

boost!

CSE

SYNTACTIC VS. WORLD-KNOWLEDGE• Syntactic features

– Words before/after candidate instance

– Part of speech

– Regular expressions on the candidate

• World knowledge features– Company signature

– Candidate aliases

• We trained each model separately and then took a logical OR of the results

• Result model introduces significant improvement over baseline

SOLVING PRODUCTION ISSUES• Suppose we have a bug in production

“ISIS is turning us all into its recruiting sergeants”

• We do not want to wait for the next drop to fix it

• “Negative signature” – get a list of related bugs and create a patch to solve it on-the-fly– Content specialist needs to answer a few yes/no

questions and the system builds a classifier to identify further instances of the bug and clean it

DOCUMENT LEVEL TRIAGE• Let’s stay with our friends in ISIS example

• Sometimes the sentence itself really looks as if it has a company mention

“ISIS is turning us all into its recruiting sergeants”

• As a human, we immediately identify the document is not about a company

• We built a classifier at the document level answering the question: How likely is this document to contain any company?

UNSUPERVISED VERSION OF CSE• Find candidate instances in the corpus

• Using clustering and basic statistics infer clearly positive and clearly negative instances

• Identify unambiguous instances to be used as positive examples– Avoid introducing a feature that looks at these aliases

• Use sibling lexicons as negative examples

• Use negative and positive set to train a classifier to decide on more instances

• Iterate…

CONCLUSIONS AND FUTURE WORK• When only known entities matter, entity extraction

becomes easier

• Labeling data is cheaper

• Faster improvement iterations

• Easier to use world knowledge

• Easier to apply domain adaptation

WE ARE HIRING!