A Probabilistic Classifier for Table Visual Analysis

1
A Probabilistic Classifier for Table Visual
Analysis

• William Silversmith
• TANGO Research Project
• NSF Grant 0414644 and 04414854
• Greetings Prof. Embley!

2
Outline

• Motivation
• High level view
• Algorithm
• Results of a small experiment
• Discussion
• Appendix

3
Motivation (1)

• Tables use visual cues to present information.
• Known approaches (to us) exclusively use
  structure or pay little attention to visual
  information.
• Potential use Segment category regions from
  delta cells, allowing further automation of
  TANGO.

4
Source http//www40.statcan.ca/l01/cst01/legal13-
eng.htm
5
Motivation (2)

• However,visual features are not used consistently
• Ideally, automatically analyze them anyway
• Globally
• Within an internet domain (ie Canada Statistics)
• Within a subject domain
• Train a program to understand a given domain!

6
High Level View

• Find a group of tables that are accompanied by a
  listing of their categories and delta cells
• Train the program on them to recognize how visual
  features predict the splits
• Use trained program on new tables to allow TANGO
  to mine data

7
Algorithm (1)

• Find a training set (tables with verifying data)
• Select visual features to analyze
• Examples
• Font, Indentation, Empty Cells, Background Color,
  Font Color, Many More!

8
Algorithm (2)

• Training the classifier
• for each attribute
• Form the difference table // See appendix
• Sum along the rows and columns
• The highest row and column indicates the
  horizontal and vertical split indicated by the
  attribute
• Compare this coordinate with the one indicated by
  the verification data
• Track the number of hits and misses persistently

9
Algorithm (3)

• The probability that each attribute predicts a
  correct answer is its weight.
• Weights can be tabulated for horizontal features,
  vertical features, and combined.
• Certain features may be more sensitive along a
  given direction

10
Algorithm (4)

• Use classifier
• for each table in the data set
• Form the difference table by summing all weighted
  attributes together for each difference
• Sum the resulting distinctions along rows and
  columns
• The highest sum for the top or leftmost split for
  horizontal and vertical splits respectively is
  the predicted category/delta segmentation

11
Experiment (1)

• A small experiment was conducted by hand to test
  the efficacy of the algorithm.
• Procedure
• Select six tables from Canada Statistics
• Use four to train the classifier
• Assess it on the last two
• By chance, one of the training set tables was a
  concatenated table.

12
Experiment (2)

• Analyzed characteristics
• Font style (normal/italic/bold)
• Indentation (left/center/right/offset)
• Data type (string or reasonably recognizable
  number)
• Adjacent empty cell (whitespace)
• Empty/Nonempty transitions do not count except
  for the whitespace measure

13
Experiment Training Results (3)
14
Experiment Training Results (4)
15
Experiment Classifier Results (5)

• Both data set targets were properly segmented
  using both combined and distinct weights
  approaches for horizontal and vertical cuts
• Neither of the targets were concatenated tables

16
Experiment Classifier Results (6)

• Uniformly random classifier
• P( T1 and T2 ) (1/49)(1/120) 1 in 5880
  trials
• P( T1 or T2 ) (1/49)(119/120)(1/5880
  )(1/120)(48/49)
• 
  1 in 35 trials
• Sample space points in the table

17
Discussion Results (1)

• Font style less important that previously
  thought
• Font style still seemed to indicate the presence
  of aggregates
• Some decisions made by slim margins
• Vertical cuts predicted by indentation

18
Discussion Results (2)

• Whitespace is useful, but sometimes misleading
• A row of whitespace below the column categories
  confuses the classifier.
• Some choices of validating splits are debatable

19
Discussion Future Issues (3)

• Current formulation ignorant of number of cuts?
• Potentially useful in describing large numbers of
  similar tables
• Where is the training data?

20
Discussion Future Issues (4)

• Sources of training data are hard (or possibly
  expensive) to come by
• Canada Statistics provides access to internal
  databases for 5000
• Luckily, Raghav produced about 200 samples by
  accident!
• Format is DoclabXML!
• Hooray for Raghav and TAT!

21
Discussion Enhancements (5)

• Use a rule based learning system?
• Detect visual patterns?
• Combine other approaches?

22
Appendix 1 Difference Table

• Form a difference table by taking all the cells
  adjacent on the top, bottom, left, and right and
  checking to see if there is a difference in the
  attribute(s) you are looking at.
• A difference returns 1, no difference returns 0.
• Do this for all cells in the table.
• Edges of the table do not count as cells.

23
Experimental Data Sources

• Canada Statistics
• Training Set
• http//www40.statcan.ca/l01/cst01/legal13-eng.htm
• http//www40.statcan.ca/l01/cst01/serv11a-eng.htm
• http//www40.statcan.ca/l01/cst01/famil36a-eng.htm
  
• http//www40.statcan.ca/l01/cst01/labor50a-eng.htm
  
• Data Set
• http//www40.statcan.ca/l01/cst01/serv22-eng.htm
• http//www40.statcan.ca/l01/cst01/agrc18a-eng.htm