This document provides an overview of information extraction (IE). It describes IE as the process of scanning text to extract relevant entities, relations, and events. The document outlines common IE tasks like named entity recognition and discusses approaches to IE like using cascaded finite-state transducers and learning-based methods. It also addresses challenges in IE like measuring performance and how systems are progressing towards overcoming the 60% accuracy barrier.

1. Information Extraction
Joan Hurtado
Ignacio Delgado

2. Contents
• INTRODUCTION
• IE TASKS
• IE WITH CASCADED FINITE-SATATE TRANSDUCERS
• LEARNING-BASED APPROACHES TO IE
• HOW GOOD IS IE

3. INTRODUCTION

4. What is IE?
 Information Extraction is the
process of scanning text for
relevant information to
some interest
 Extract:
 Entities
 Relations
 Events
 Who did what to whom
when and where

5. Why IE?
 Need for eficient processing of texts in
specialized domains
 Focus on relevant parts, ignore the rest
 Typical applications:
 Gleaning business
 Government
 Military intelligence
 WWW searches (more specific than keywords)
 Scientific literature searches
 …

6. Most common uses
 Named Entity Recognition
 Identify names, special
entities (dates, times)
 Uses textual patterns
 Important at biomedical
applications
 IE is more than NER
 Recognition of events
and their participants

7. How to measure
performance
 Recall
 What percentage of the correct answers did the
system get
 Precision
 What percentage of the system’s answers were
correct
 F-score
 Weighted harmonic mean between recall and
precision

8. c
IE TASKS

9. Unstructured vs. Semi-structured
text
Unstructured
 Natural language
sentences
 Semantics depends on
linguistic analysis
 Examples:
 News stories
 Magazines articles
 Books
 …
Semi-structured
 Structured data
 Semantics defined by its
organization
 Physical layout plays role in
interpretation
 Examples:
 Job postings
 Rental ads
 …

10. Single-document vs. Multi-
document
 Originally IE systems designed for individual
documents
 Nowadays new systems to extract facts from WWW
 Both use similar techniques
 Distinguishing issue: redundancy
 Multi-document can exploit redundancy
 However need to challenge cross-document
coreference resolution
 Multi-document IE systems also are referred as
open-domain

11. Assumptions about
Incoming Documents
 Relevant only documents
 Single event documents

12. IE WITH CASCADE FINITE-STATE TRANSDUCERS

13. Complex Words
Basic Phrases
Complex Phrases
Domain Events
Template Generation

14. Complex Words
 Identify multiwords, company names, people
names, locations, dates, times and basic entities
 Recognition strategies:
 Patterns
 Dictionaries
 Context

15. Basic Phrases
 Some syntactic constructs can be
identified with reasonable
reliability:
 Noun group
 Verb group
 Strategies:
 Simple finite-state grammars
 Ambiguities
 Noun-verb ambiguity
 Verbs locally ambiguous
 Problems
 Not al languages have high
distinction between noun and
verb groups

16. Complex Phrases
 Recognize complex noun and verb groups
 Complex noun groups
 Appositives
 Measure phrases
 Prepositional attachments (of, for)
 Noun group conjunction
 Complex verb groups
 Verb conjunction
 Verb groups with same significance
 Domain-relevant entities can be recognized

17. Domain Events
 Ignore anything not identified in previous phases
 Domain events require domain-specific patterns
for identification
 Strategy:
 Finite-state machines
 Certain kind of “pseudo-syntax” can be done
 Nowadays IE systems begin to rely in full-
sentence parsing

18. Template Generation:
Merging Structures
 Previous stages operate within bounds of single
sentences
 Operate over whole text to combine previous
collected information into a unified whole
 If recognizing multiple events:
 Determine how many distinct events
 Assign each entity to appropriate event

19. LEARNING-BASED APPROACHES TO IE

20. Supervised Learning of
Extraction patterns & rules
 Reduce knowledge engineering bottleneck
required to create an IE system for a new domain
 Examples:
 AutoSlog  create lexico-syntactic patterns
 PALKA  patterns generalized based on words
semantics
 LIEP  identify syntactic paths related to roles
 CRYSTAL  “concept nodes” with lexical, syntactic
and semantic constrains
 WHISK  learn regular expressions
 Many others: SRV, RAPIER, …

21. Supervised Learning of
sequential classifier models
 View IE as a classification problem that can be
tackled using sequential learning models
 Read sequentially and label each word as an
extraction or a non-extraction
 Typical labeling scheme IOB
 Inside
 Outside
 Beginning of desired extraction
 Strategies:
 Hidden Markov Models
 Maximum Entropy Classifiers
 Support Vector Machines

22. Weakly supervised and
unsupervised approaches
 Annotating training text still requires time and
complexity
 Further techniques to learn extraction using weakly
supervised and unsupervised systems
 Examples
 AutoSlog-TS (preclassifed corpus which texts identified
as relevant or irrelevant)
 Ex-Disco (manually defined seed, patterns ranked, best
patterns selected added to seed)
 Meta-bootstraping (seed nouns that belong to
semantic class)
 On-Demand Information Extraction (dynamically learns
from queries)

23. Discourse-oriented
approaches to IE
 Most IE systems patterns focus only on local
context surrounding
 Extend systems to have more global view
 Strategy:
 Add constrains to connect entities in diferent
clauses
 Decision trees (WRAP-UP)
 Set of classifiers to identify new templates (ALICE)

24. HOW GOOD IS IE

25. How IE systems are
progressing?
 The 60% barrier in performance
 Biggest mistakes in entity and event coreference
 The implicit knowledge on NL not translated to texts
 Problems on training data not found on test data
 Good IE systems typically recognize 90% of entities
 An event requires about 4 entities
 0.9*0.9*0.9*0.9 = 65.61%

26. THANKS