1

ListReader:Wrapper Induction for Lists

in OCRed Documents

Thomas PackerBYU CS DEG2012.03.17

2

We Love Data (In Digital Form)

3

Lots of Paper Documents in the World

4

Lots of Text in Paper Documents

5

Lots of Lists in TextLots of Data in Lists

6

Manual Data Entry

7

Wrappers: Individualized Extraction Rules

8

Semi-Automatic Wrapper Induction

9

Weakly-supervised

Wrapper Induction

Semi-supervised Wrapper Induction

10

Data: Image

11

Data: OCR

12

Data: Hand-Labeled OCR

13

Induced Regex Wrappers

Single-Specific

Single-General

Multi-General

(i)(\. )(Lydia)( )(Lewis)(\*, )(b\. ) …

([a-z]{1,5})([ \t\r\n\.]{1,6})([a-zA-Z]{3,9})([ \t\r\n]{1,5}) ([a-zA-Z]{3,9})([ \t\r\n\*,]{1,7})([ \t\r\n\.a-z]{1,7}) …

([iv]{1,3})([ \.]{2,2})([ACHJLacdehilmnrstuvy]{4,6})([ ]{1,1}) ([Leisw]{5,6})([ \*,]{2,3})([ \.abp]{3,5}) …

Character Classes[a-z][A-Z][0-9][ \t\r\n]<each punct.>

15

Preliminary Results(Field Label F-measure, Small Dataset)

Single-Specific Single-General Multi-General0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

31%36%

54%

45% 46%

60%

TrainingTest

16

Conclusions and Future Work

Single-Specific Single-General Multi-General Transduction0%

10%

20%

30%

40%

50%

60%

70%

80%

90%

100%

31%36%

54%

76%

45% 46%

60%

79%

TrainingTest

17

All done. Suggestions?