Information Retrieval

1
Information Retrieval

• A Eugene M. Taranta II Presentation

2
Topics Of Discussion

• IR
• Definitions
• Process
• Research Areas
• Models
• Vector-space

3
What is IR ?

• Information Retrieval (IR) is concerned with
  techniques that can provide effective access to
  large collections of objects containing primarily
  text from 1

4
IR Definitions

• Object(Document) Source of information in text
  form
• Scientific Journals
• Emails
• News Articles
• Manuals
• Etc.

5
IR Definitions

• Object Descriptions contain a set of attributes
  that describe the content of a document
• Object Descriptions may be the actual object
  itself in the case of smaller documents

6
IR Definitions

• Information Need A set of attributes that define
  a researchers interest
• Query A Description of the Information Need
• Often expressed in natural language

7
IR

• Is not database query processing
• Database Query Processing
• No use of natural language
• Uses well defined semantics
• Ex Inventory System

8
IR Process

• Generate a representation for each object
  description
• Generate a representation of the information need
• Compare 1 and 2 to select objects that most
  likely satisfy research need

9
Areas of IR Research

• Document Representation
• Information Need Representation
• Comparison of 1 and 2
• Evaluation

10
IR Models

• How IR Process is achieved is an IR Model
• Four Common Models
• Boolean
• Cluster-Based Retrieval
• Vector Space Retrieval
• Probabilistic Retrieval

11
IR Model Boolean

• Basis for most commercial IR Systems
• Object Description is defined by a set of Boolean
  attributes
• Information Need described by a Boolean expression

12
IR Model Boolean

• No Ranking System
• Difficult to use
• Low precision
• These are due to lack of attribute weights

13
IR Model Cluster-Based

• Assumes similar documents satisfy same
  information need
• Create clusters of like documents
• Each cluster gets an Average representation
• Retrievals returns clusters rather than
  independent documents

14
IR Model Vector-Space

• Individual docs are described in vectors
• Elements of Vectors represent weighted values
• Weights are determined from frequency models
• These models are generally considered Ad Hoc

15
IR Model Probabilistic

• Based on Probability Ranking Principle
• Overall effectiveness will be achieved when docs
  are ranked in decreasing order
• Involves estimations of probability in relevance

16
IR Model Notes

• Vector-space and Probabilistic share similarities
  with Boolean
• All support partial matching and may use Boolean
  logic in queries.
• Probabilistic contributed to to our understanding
  of term weighting, ranking and relevance feedback

17
IR Architecture
18
Vector-space The Pieces

• Inverse Document Frequency (idf) A weighing
  factor of term frequency among the collection of
  documents
• Term Frequency (tf) Number of times a term
  appears in a document

19
Vector-space The Pieces

• idf can be calculated as
• log( d/df )
• d, number of documents
• df, document frequency
• idf gives us a sense of term importance

20
Vector-space
21
Vector-space
22
Vector-space

• Cell value tfDoci idf

23
Vector-space

• Three Representation Vectors
• Non-zero vector
• Column-index vector
• Row vector

• Vector-space algorithms
• Coordinate Storage (COO)
• Compressed Sparse Column (CSC)
• Compressed Sparse Row (CSR)
• Block Sparse Row (BSR)

24
CSR non-zero vector
25
CSR column vector
26
CSR row vector
27
Vector-space Query

• Query Vector
• Entry for each term
• Value is based on idf

Information Need Researcher is looking for
information on Apple Computers
Query Apple Computer
28
CSR Algorithm

• M ? Number of Documents
• for count ? 0 to M
• temp ? 0
• row_idx ? row_vector count
• while row_idx lt ( row_vectorcount 1 1 )
• col_idx ? col_vector row_idx
• temp ? temp(non_zero_vectorrow_idx
• Qcol_idx)
• CSR_outputcount ? temp

Examine each doc
Grab pointer to where row starts in non-zero
vector
Examine each term in doc
If term is in query, add relevance
Save doc result
29
CSR Example Query

• Examine Doc 0

(.44.22)
(.8.0)
.097
Examine Doc 1
(.8.0)
(1.40)
0
Examine Doc 2
(.22.22)
(.70)
(.4.4)
.21
Examine Doc 3
(.22.22)
(.70)
(.70)
(.70)
.05
Examine Doc 4
(.4.4)
(.70)
.16
30
CSR Example Query Result

• Examine Doc 0

(.44.22)
(.80)
.097
Examine Doc 1
(.8.0)
(1.40)
0
Examine Doc 2
(.22.22)
(.70)
(.4.4)
.21
Examine Doc 3
(.22.22)
(.70)
(.70)
(.70)
.05
Examine Doc 4
(.4.4)
(.70)
.16
31
Summary

• Information Retrieval uses natural language to
  query into a collection of documents
• Four basic methods include Boolean, Cluster,
  Vector-space, and Probabilistic
• Vector-Space use three representation vectors
  (non-zero, column and row) and one query vector
• Vector multiplication yields a relevance source
  by which we can rank results

32
References

• Croft, Bruce W. and Turtle, Howard R. Text
  Retrieval and Inference Text-based Intelligent
  Systems Current research and practice in
  information extraction and retrieval 1992 pp
  127-155
• Goharian, Nazli Jain,Ankit and Sun, Qian,
  "Comparative Analysis of Sparse Matrix Algorithms
  for Information Retrieval", Journal of Systemics,
  Cybernetics and Informatics, 2003.
• Salton Gerard, Automatic Text Processing The
  Transformation, Analysis, and Retrieval by
  Computer. Addison-Wesley Pub Co 1998.