ie.pdf

Transcript of ie.pdf

Information Extraction - Part I

SNLP 2014

CSE, IIT Kharagpur

October 27, 2014

SNLP 2014 (IIT Kharagpur) Information Extraction October 27, 2014 1 / 33
Goal: machine reading

GoalAcquire structured information knowledge from unstructured text

Information Extraction

Information Extraction (IE) SystemsFind and understand limited relevant parts of texts

Gather information from many pieces of textProduce a structured representation of relevant information:

I Relations (in the database sense)I A knowledge base

GoalsOrganize information so that it is useful to people

Put information in a semantically precise form that allows furtherinferences to be made by computer algorithms



Gather information from many pieces of text

Produce a structured representation of relevant information:I Relations (in the database sense)I A knowledge base



Information Extraction (IE)

DefinitionInformation extraction is the task of finding structured information fromunstructured or semi-structured text.

What sort of information?IE Systems extract clear, factual information

Roughly: Who did what to whom when? etc.

E.g., Gathering earnings, profits, headquarters etc. from company reportsThe headquarters of BHP Billiton Limited, and the global headquarters ofthe combined BHP Billiton Group, are located in Melbourne, Australia.

headquarters(BHP Biliton Limited, Melbourne, Australia)



What sort of information?

IE Systems extract clear, factual information





ExampleIn 1998, Larry Page and Sergey Brin founded Google Inc.We can extract the following information,

FounderOf(Larry Page, Google Inc.),

FounderOf(SergeyBrin, Google Inc.),

FoundedIn(Google Inc., 1998)

Such information can be used by search engines and database managementsystems to provide better services to end users.


ExampleIn 1998, Larry Page and Sergey Brin founded Google Inc.

We can extract the following information,





Applications in Biomedical domain

Biomedical domainA large amount of scientific publications

Need to look for discoveries related to particular genes, proteins or otherbiomedical entities

Biomedical entities often have synonyms and ambiguous names

Critical task: automatically identify mentions of biomedical entities in textand link them to their corresponding entries in existing knowledge bases.

Biomedical domain

More applications of IE

Building and extending knowledge bases and ontologies

Scholarly literature databases: Google Scholar, CiteSeerX

People directories: Rapleaf, Spoke, Naymz

Bioinformatics: clinical outcomes, gene interactions, ...

Stock analysis: deals, acquisitions, earnings, hirings and firings

Intelligence analysis for business and government

A Sample document containing a terrorism event

Santiago, 10 Jan 90 - Police are carrying out intensive operations in the townof Molina in the seventh region in search of a gang of alleged extremists whocould be linked to a recently discovered arsenal. It has been reported thatCarabineros in Molina raided the house of of 25-year-old worker Mario MunozPardo, where they found a fal rifle, ammunition clips for various weapons,detonators, and material for making explosives. It should be recalled that agroup of armed individuals wearing ski masks robbed a businessman on arural road near Molina on 7 January. The businessman, Enrique OrmazabalOrmazabal, tried to resist; The men shot him and left him seriously wounded.He was later hospitalized in Curico. Carabineros carried out severaloperations, including the raid on Munoz home. The police are continuing topatrol the area in search of the alleged terrorist command.

Information Extraction as per the template

Relation Extraction

Relation types

For generic news text ...

Relation types from ACE 2003

ROLErelates a person to an organization or a geopolitical entitysubtypes: member, owner, affiliate, client, citizen

PARTgeneralized containmentsubtypes: subsidiary, physical part-of, set membership

ATpermanent and transient locationssubtypes: located, based-in, residence

SOCIALsocial relations among personssubtypes: parent, sibling, spouse, grandparent, associate

Relation types: Freebase

Relation types: geographical

More relations: disease outbreaks

More relations: protein interactions

Relation extraction: 5 easy methods

Hand-built patterns

Bootstrapping methods

Supervised methods

Distant supervision

Unsupervised methods

Hand-written Information Extraction: use regex

You can also use Tregex

Identifying Patterns in TreesSimple example: NP < NN

Tregex: Other relations

Rule-based Extraction Examples

Determining which person holds what position in what organization

[person], [position] of [org]Vuk Draskovic, leader of the Serbian Renewal Movement

[org] (named, appointed, etc.) [person] Prep [office]NATO appointed Wesley Clark as Commander in Chief


Determining where an organization is located

[org] in [loc]NATO headquarters in Brussels

[org] [loc] (division, branch, headquarters, etc.)KFOR Kosovo headquarters

Patterns for learning hyponyms

Intuition from Hearst (1992)Agar is a substance prepared from a mixture of red algae, such as Gelidium,for laboratory or industrial use.

What is Gelidium?

How do you know?



What is Gelidium?

How do you know?



What is Gelidium?

How do you know?

Hearsts lexico-syntactic patterns

Automatic Acquisition of Hyponyms

Y such as X((,X) (, and/or) X)such Y as X

X or other Y

X and other Y

Y including X

Y , especially X

Examples of Hearst patterns

Patterns for learning meronyms

Berland and Charniaks patternsSelected initial patterns by finding all sentences in a corpus containingbasement and building

Patterns for learning meronyms

Berland and Charniaks patternsSelected initial patterns by finding all sentences in a corpus containingbasement and building

Problems with hand-built patterns

Requires hand-building patterns for each relation!I hard to write; hard to maintainI there are zillions of themI domain-dependent

Dont want to do this for all possible relations!Plus, wed like better accuracy

I Hearst: 66% accuracy on hyponym extractionI Berland and Charniak: 55% accuracy on meronyms

Bootstrapping approaches

If you dont have enough annotated text to train on . . .But you do have:

I some seed instances of the relationI (or some patterns that work pretty well)I and lots and lots of unannotated text (e.g., the web)

can you use those seeds to do something useful?

Bootstrapping can be considered semi-supervised

Bootstrapping example

Target relation: burial place

Seed tuple : [ Mark Twain, Elmira ]

Google for Mark Twain and Elmira

Use those patterns to search for new tuples

Bootstrapping relations

Bootstrapping problems

Requires that we have seeds for each relationI Sensitive to original set of seeds

Generally have lots of parameters to be tunedNo probabilistic interpretation

I Hard to know how confident to be in each result

Supervised relation extraction

The supervised approach requiresDefining an inventory of output labels

I Relation classification: located-in, employee-of, inventor-of,

Collecting labeled training data: MUC, ACE . . .

Defining a feature representation: words, entity types, . . .

Choosing a classifier: Nave Bayes, MaxEnt, SVM, . . .

Evaluating the results

ie.pdf

Documents

Transcript of ie.pdf