Learning to construct knowledge bases form the world wide web

1
Learning to construct knowledge bases form the
world wide web

• Yang Lu
• luyang_at_cse.msu.edu
• Department of Computer Science and Engineering
• Michigan State University
• Advisor Dr Sakti Pramanik

2
Learning to construct knowledge bases form the
world wide web

• Goal of the project
• Automatically create a computer understandable
  knowledge base whose content mirrors that of the
  world wide web.

3
The problem

• Huge amount of web pages.
• Classification of web pages.
• Relations among the web pages.
• Machine understandable.

4
An example

• Dr. Pramaniks web page

5
How to construct and maintain the knowledge base

• Using machine learning to create information
  extraction methods for each of the desired types
  of knowledge.
• Applying the learned information extraction
  methods to extract symbolic, probabilistic
  statements direct from the web hypertext.

6
The WebKB system overview

• The system is first trained to extract
  information of the desired types, then it is
  allowed to browse new web sites to extract a new
  knowledge base.

7
Training the WebKB system

• The training data
• A specification of the classes and the relations
  of interest.
• Training examples that describe instances of the
  ontology classes and relations.

8
Example of training data

• Classes person, student, course,
  research_project
• Relations
• advisor_of( student, faculty)
• The samples are hand labeled.

9
Assumption due to the variation nature of the web

• How class instance are described on the web
• Each instance of an ontology class is
  represented by one or more segments of hypertext
  on the web.( a single webpage, continuous text
  within a page or a root graph of several pages
  linked by hyperlinks.)
• How relation instances are described on the web
• An undirected path of hyperlinks
• A segment of text representing A that contains
  the segment that represents B
• Hypertext segment for A satisfies some learned
  model for relatedness to B.

10
An example from the paper
11
Recognize class instance

• Statistical bag-of-words method
• Ignore the sequence in which the words occur.
• Assume word occurrence is isolated.

12
Recognize class instance

• Estimate the word probability
• Need to be robust for infrequent words.
• Estimate the vocabulary
• Limit the vocabulary size to 2000 words.

13
Recognize class instance

• Results

14
Recognize class instance

• Results
• Top vocabulary list for each class

15
Recognize class instance

• Results
• Top vocabulary list for each class

16
Recognize class instance

• Results
• Accuracy/coverage of full text

17
Recognize class instance

• Results
• Accuracy/coverage of hyperlink

18
Recognize class instance

• Results
• Accuracy/coverage of title/heading

19
Recognize class instance

• First-order text classification
• Learning to classify web pages by using a learner
  which is able to induce first-order rules
• Using FOIL algorithm.
• Example A page is a course homepage if it
  contains the words textbook and TA and is linked
  to a page that contains the word assignment.

20
Recognize class instance

• First-order text classification
• Apply the FOIL 6.4 with default setting to learn
  a set of clauses consisting the classes mentioned
  above except other.
• other is the default class.

21
Recognize class instance

• First-order text classification
• Learned rule classifier example

• Explanation of the rule
• The page has the word instructor, but does not
  have the word good.
• The page contains a hyperlink to a page which
  does not contain any hyperlink to other pages.
• This linked page contains the word assign.

22
Recognize class instance

• Combine the statistical classifier and the
  first-order rule based classifier
• Using a simple voting algorithm

23
Recognize relation instance

• Recognize relations of interest that exist among
  extracted class instances.
• Relation among class instances are represented by
  hyperlink paths in the web.

24
Recognize relation instance

• Using FOIL algorithm to derive the relations.
• Two step process
• Path finding
• Rule finding

25
Recognize relation instance

• Path finding process an example
• Rule finding process FOIL algorithm

26
Extract text fields

• Using SRV algorithm
  ( sequence rules with validation )
• SRV is a first-order rule learner based on FOIL.
• Input is a set of web pages.
• Output is a set of information extraction rules.
• Work on the HTML codes and tried to match a
  sequence of tokens.

27
Conclusion

• Advantages
• Automatically extraction.
• Good performance.
• Disadvantages
• Need training sample.
• Algorithm needs improvement.

28
Unsupervised information extraction

• Soft matching patterns
• Represent patterns by
• Combining lexical tokens alongside
  part-of-speech classes and punctuations.
• Adopting a probabilistic framework that combines
  slot content and sequential fidelity in computing
  the degree of pattern match.

29
Unsupervised information extraction

• Tokens for the soft matching patterns

30
Unsupervised information extraction

• Example of generalizing soft patterns

31
Unsupervised information extraction

• Pattern matching algorithm
• Cosine similarity.
• Express the soft pattern as following
• ltSlot-w, , Slot-2, Slot-1, SCH_TERM , Slot1,
  Slot2, Slotw Pagt
• lt(tokeni1, weighti1), (tokeni2, weighti2)
  (tokenim, weightim) Slotigt