A Survey of WEB Information Extraction Systems

1
A Survey of WEB Information Extraction Systems

• Chia-Hui Chang
• National Central University
• Sep. 22, 2005

2
Introduction

• Abundant information on the Web
• Static Web pages
• Searchable databases Deep Web
• Information Integration
• Information for life
• e.g. shopping agents, travel agents
• Data for research purpose
• e.g. bioinformatics, auction economy

3
Introduction (Cont.)

• Information Extraction (IE)
• is to identify relevant information from
  documents, pulling information from a variety of
  sources and aggregates it into a homogeneous form
• An IE task is defined by its input and output

4
An IE Task
5
Web Data Extraction
Data Record
Data Record
6
IE Systems

• Wrappers
• Programs that perform the task of IE are referred
  to as extractors or wrappers.
• Wrapper Induction
• IE systems are software tools that are designed
  to generate wrappers.

7
Various IE Survey

• Muslea
• Hsu and Dung
• Chang
• Kushmerick
• Laender
• Sarawagi
• Kuhlins and Tredwell

8
Related Work Time

• MUC Approaches
• AutoSolg Riloff, 1993, LIEP Huffman, 1996,
  PALKA Kim, 1995, HASTEN Krupka, 1995, and
  CRYSTAL Soderland, 1995
• Post-MUC Approaches
• WHISK Soderland, 1999, RAPIER califf, 1998,
  SRV Freitag, 1998, WIEN Kushmerick, 1997,
  SoftMealy Hsu, 1998 and STALKER Muslea, 1999

9
Related Work Automation Degree

• Hsu and Dung 1998
• hand-crafted wrappers using general programming
  languages
• specially designed programming languages or tools
• heuristic-based wrappers, and
• WI approaches

10
Related Work Automation Degree

• Chang and Kuo 2003
• systems that need programmers,
• systems that need annotation examples,
• annotation-free systems and
• semi-supervised systems

11
Related Work Input and Extraction Rules

• Muslea 1999
• IE from free text using extraction patterns that
  are mainly based on syntactic/semantic
  constraints.
• The second class is Wrapper induction systems
  which rely on the use of delimiter-based rules.
• The third class also processes IE from online
  documents however the patterns of these tools
  are based on both delimiters and
  syntactic/semantic constraints.

12
Related Work Extraction Rules

• Kushmerick 2003
• Finite-state tools (regular expressions)
• Relational learning tools (logic rules)

13
Related Work Techniques

• Laender 2002
• languages for wrapper development
• HTML-aware tools
• NLP-based tools
• Wrapper induction tools (e.g., WIEN, SoftMealy
  and STALKER),
• Modeling-based tools
• Ontology-based tools
• New Criteria
• degree of automation, support for complex
  objects, page contents, availability of a GUI,
  XML output, support for non-HTML sources,
  resilience and adaptiveness.

14
Related Work Output Targets

• Sarawagi VLDB 2002
• Record-level
• Page-level
• Site-level

15
Related Work Usability

• Kuhlins and Tredwell 2002
• Commercial
• Noncommercial

16
Three Dimensions

• Task Domain
• Input (Unstructured, semi-structured)
• Output Targets (record-level, page-level,
  site-level)
• Automation Degree
• Programmer-involved, learning-based or
  annotation-free approaches
• Techniques
• Regular expression rules vs Prolog-like logic
  rules
• Deterministic finite-state transducer vs
  probabilistic hidden Markov models

17
Task Domain Input
18
Task Domain Output

• Missing Attributes
• Multi-valued Attributes
• Multiple Permutations
• Nested Data Objects
• Various Templates for an attribute
• Common Templates for various attributes
• Untokenized Attributes

19
Classification by Automation Degree

• Manually
• TSIMMIS, Minerva, WebOQL, W4F, XWrap
• Supervised
• WIEN, Stalker, Softmealy
• Semi-supervised
• IEPAD, OLERA
• Unsupervised
• DeLa, RoadRunner, EXALG

20
Automation Degree

• Page-fetching Support
• Annotation Requirement
• Output Support
• API Support

21
Technologies

• Scan passes
• Extraction rule types
• Learning algorithms
• Tokenization schemes
• Feature used

22
A Survey of Contemporary IE Systems

• Manually-constructed IE tools
• Programmer-aided
• Supervised IE systems
• Labeled based
• Semi-supervised IE systems
• Unsupervised IE systems
• Annotation-free

23
(No Transcript)
24
Manually-constructed IE Systems

• TSIMMIS Hammer, et al, 1997
• Minerva Crescenzi, 1998
• WebOQL Arocena and Mendelzon, 1998
• W4F Saiiuguet and Azavant, 2001
• XWrap Liu, et al. 2000

25
A Running Example
26
TSIMMIS

• Each command is of the form
• variables, source, pattern where
• source specifies the input text to be considered
• pattern specifies how to find the text of
  interest within the source, and
• variables are a list of variables that hold the
  extracted results.
• Note
• means save in the variable
• means discard

27
Minerva

• The grammar used by Minerva is defined in an EBNF
  style

28
WebOQL

• Select Z!.Text
• From x in browse (pe2.html), y in x, Z in y
• Where x.Tag ol and Z.TextReviewer Name

29
W4F

• Wysiwyg support
• Java toolkit
• Extraction rule
• HTML parse tree (DOM object)
• e.g. html.body.ol0.li.pcdata0.txt
• Regular expression to address finer pieces of
  information

30
Supervised IE systems

• SRV Freitag, 1998
• Rapier Califf and Mooney, 1998
• WIEN Kushmerick, 1997
• WHISK Soderland, 1999
• NoDoSE Adelberg, 1998
• Softmealy Hsu and Dung, 1998
• Stalker Muslea, 1999
• DEByE Laender, 2002b

31
SRV

• Single-slot information extraction
• Top-down (general to specific) relational
  learning algorithm
• Positive examples
• Negative examples
• Learning algorithm work like FOIL
• Token-oriented features
• Logic rule

Rating extraction rule- Length(1),
Every(numeric true), Every(in_list true).
32
Rapier

• Field-level (Single-slot) data extraction
• Bottom-up (specific to general)
• The extraction rules consist of 3 parts
• Pre-filler
• Slot-filler
• Post-filler

Book Title extraction rule- Pre-filler slot-fille
r post-filler word Book Length2 wordltbgt word
Name Tag nn, nns word lt/bgt
33
WIEN

• LR Wrapper
• (Reviewer name lt/bgt, ltbgt, Rating lt/bgt,
  ltbgt, Text lt/bgt, lt/ligt)
• HLRT Wrapper (Head LR Tail)
• OCLR Wrapper (Open-Close LR)
• HOCLRT Wrapper
• N-LR Wrapper (Nested LR)
• N-HLRT Wrapper (Nested HLRT)

34
WHISK

• Top-down (general to specific) learning
• Example
• To generate 3-slot book reviews, it start with
  empty rule ()()()
• Each parenthesis indicates a phrase to be
  extracted
• The phrase in the first set of parenthesis is
  bound to variable 1, and 2nd to 2, etc.
• The extraction logic is similar to the LR wrapper
  for WIEN.

Pattern Reviewer Name lt/bgt (Person) ltbgt
(Digit) ltbgtTextlt/bgt() lt/ligt Output
BookReview Name 1 Rating 2 Comment 3
35
NoDoSE

• Assume the order of attributes within a record to
  be fixed
• The user interacts with the system to decompose
  the input.
• For the running example
• a book title (an attribute of type string) and
• a list of Reviewer
• RName (string), Rate (integer), and Text
  (string).

36
Softmealy

• Finite transducer
• Contextual rules

slt,RgtL HTML(ltbgt) C1Alph(Rating)
HTML(lt/bgt) slt,RgtR Spc(-) Num(-) sltR,gtL
Num(-) sltR,gtR NL(-) HTML(ltbgt)
37
Stalker

• Embedded Category Tree
• Multipass Softmealy

38
DEByE

• Bottom-up extraction strategy
• Comparison
• DEByE the user marks only atomic (attribute)
  values to assemble nested tables
• NoDoSE the user decomposes the whole document in
  a top-down fashion

39
Semi-supervised Approaches

• IEPAD Chang and Lui, 2001
• OLERA Chang and Kuo, 2003
• Thresher Hogue, 2005

40
IEPAD

• Encoding of the input page
• Multiple-record pages
• Pattern Mining by PAT Tree
• Multiple string alignment
• For the running example
• ltligtltbgtTlt/bgtTltbgtTlt/bgtTltbgtTlt/bgtTlt/ligt

41
OLERA

• Online extraction rule analysis
• Enclosing
• Drill-down / Roll-up
• Attribute Assignment

42
Thresher

• Work similar to OLERA
• Apply tree alignment instead of string alignment

43
Unsupervised Approaches

• Roadrunner Crescenzi, 2001
• DeLa Wang, 2002 2003
• EXALG Arasu and Garcia-Molina, 2003
• DEPTA Zhai, et al., 2005

44
Roadrunner

• Input multiple pages with the same template
• Match two input pages at one time

Sample page 01 lthtmlgtltbodygt 02 ltbgt 03
Book Name 04 lt/bgt 05 Data mining 06
ltbgt 07 Reviews 08 lt/bgt 09
ltOLgt 10 ltLIgt 11 ltbgt Reviewer Name
lt/bgt 12 Jeff 13 ltbgt Rating
lt/bgt 14 2 15 ltbgtText lt/bgt 16
17 lt/LIgt 18 ltLIgt 19 ltbgt
Reviewer Name lt/bgt 20 Jane 21 ltbgt
Rating lt/bgt 22 6 23 ltbgtText
lt/bgt 24 25 lt/LIgt 26
lt/OLgt 27lt/bodygtlt/htmlgt
Wrapper (initially) 01 lthtmlgtltbodygt 02
ltbgt 03 Book Name 04 lt/bgt 05
Databases 06 ltbgt 07 Reviews 08
lt/bgt 09 ltOLgt 10 ltLIgt 11 ltbgt
Reviewer Name lt/bgt 12 John 13 ltbgt
Rating lt/bgt 14 7 15 ltbgtText
lt/bgt 16 17 lt/LIgt 10
lt/OLgt 11lt/bodygtlt/htmlgt
parsing
String mismatch
String mismatch
String mismatch
String mismatch
tag mismatch
45
DeLa

• Similar to IEPAD
• Works for one input page
• Handle nested data structure
• Example
• ltPgtltAgtTlt/AgtltAgtTlt/Agt Tlt/PgtltPgtltAgtTlt/AgtTlt/Pgt
• ltPgtltAgtTlt/AgtTlt/PgtltPgtltAgtTlt/AgtTlt/Pgt
• (ltPgt(ltAgtTlt/Agt)TltPgt)

46
EXALG

• Input multiple pages with the same template
• Techniques
• Differentiating token roles
• Equivalence class (EC) form a template
• Tokens with the same occurrence vector

47
DEPTA

• Identify data region
• Allow mismatch between data records
• Identify data record
• Data records may not be continuous
• Identify data items
• By partial tree alignment

48
Comparison

• How do we differentiate template token from data
  token?
• DeLa and DEPTA assume HTML tags are template
  while others are data tokens
• IEPAD and OLERA leaves the problems to users
• How to apply the information from multiple pages?
• DeLa and DEPTA conduct the mining from single
  page
• Roadrunner and EXALG do the analysis from
  multiple pages

49
Comparison (Cont.)

• Techniques improvement
• From string alignment (IEPAD, RoadRunner) to tree
  alignment (DEPTA, Thresher)
• From full alignment (IEPAD) to partial alignment
  (DEPTA)

50
Task domain comparison

• Page type
• structured, semi-structured or free-text Web
  pages
• Non-HTML support
• Extraction level
• Field level, record-level, page-level

51
Task domain comparison (Cont.)

• Extraction target variation
• Missing attributes, multiple-value attributes,
  multi-order attribute permutation
• Template variation
• Untokernized Attributes

52
(No Transcript)
53
Technique-based comparison

• Scan pass
• Single pass vs mutiple pass
• Extraction rule type
• Regular expression vs. logic rules
• Feature used
• DOM tree information, POS tags, etc.
• Learning algorithm
• Machine learning vs pattern mining
• Tokernization schemes

54
(No Transcript)
55
(No Transcript)
56
Conclusion

• Criteria for evaluating IE systems from the task
  domain
• Comparison of IE systems from various automation
  degree
• The use of various techniques in IE systems

57
Future Work

• Page Fetching
• XWrap, W4F, WNDL
• Schema Mapping
• Full information
• Partial information
• Query Interface Integration

58
References

• C.-H. Chang, M. Kayed, M. R. Girgis, K. Shaalan,
  A survey of Web Information Extraction Systems.