Towards Domain-Independent Information Extraction from Web Tables
Wolfgang Gatterbauer, Paul Bohunsky, Marcus Herzog,Bernhard Krupl, and Bernhard Pollak
Presented by Aaron StewartBYU CS 652
Table Extraction Using Spatial Reasoning in the CSS2 Visual Box Model
Database and Artificial Inteligence GroupVienna University of Technology, Austria
Wolfgang Gatterbauer and Paul Bohunsky
Contributions
1. Classify visually structured data2. Non-tree IE formalism3. Argue to defer semantic interpretation of
output4. Ground truthing method5. Web table test set6. Visual results
Introduction
Source: Gatterbauer et al. 2007
Visually Structured Data on the Web
• Tables• Lists• Aligned Graphs
Visually Structured Data on the Web
Source: Gatterbauer et al. 2007
Formal Setup
• DOM Tree Representation• Visual Box Representation– Visualized Element Nodes (VENs)• DOM nodes with bounding boxes
– Visualized Words• Text words with bounding boxes
Formal Setup
Source: Gatterbauer et al. 2007
Information Extraction
• Visualized Element Nodes Table extraction (VENTex)
• Steps:– Table location– Table recognition– Table interpretation
Information Extraction
Source: Gatterbauer et al. 2007
Table Extraction
Source: Gatterbauer et al. 2007
Table Extraction
1. Gather 8 HTML node attributes2. For text, add link3. Only accept TH, TD, DIV html nodes4. Tables must form frames5. Remove duplicate bounding boxes
Table Extraction
6. Adjacency: 3 pixels7. LOCATEFRAMES algorithm8. No overlapping cells9. Minimum 3 rows, 2 columns10. Remove empty rows/columns (spacers)
LOCATE FRAMES Algorithm (earlier paper)
• Visual table model• Expansion algorithm
Visual Table Model
Source: Gatterbauer et al. 2007
Double Topographical Grid???
• Two origins– Upper left corner– Lower right corner
• Sorted lists of pixel positions– The numbers are indices– But pixels remain in regular coordinates
Neighbor Relations
Source: Gatterbauer et al. 2007
Neighbor Relations
• Expand to include neighbors 1,2,3,4– within or equal – Not bigger– Not outside– Not stepped
Expansion Algorithm
Source: Gatterbauer et al. 2007
Basic Algorithm
• http://www.dbai.tuwien.ac.at/staff/gatter/work/AAAI_2006_Presentation_Table_Extraction_Spatial_Reasoning.pdf
Table Interpretation
• Argument– Few details about the method actually used– Take data as it comes– Pass it on to a later semantic processing stage
Table Interpretation
Source: Gatterbauer et al. 2007
Performance
• Load + render: O(n)• Double topographical grid: O(n sqrt(n))• About 5 seconds per page
Web Table Ground Truthing
• Tool to copy web pages– (not easy!)– http://
www.dbai.tuwien.ac.at/user/pollak/webpagedump
• Students selected and submitted pages– 493 web tables– 269 web pages– 63 students– http://www.dbai.tuwien.ac.at/staff/gatter/ventex/
Experimental Results
Source: Gatterbauer et al. 2007
Future Work• Table extraction• Table interpretation• Nested substructures• Other visually structured data• Information integration
Source: Gatterbauer et al. 2007
My Conclusions
• Useful table-building algorithm– For electronic data only– Requires strict alignment
• Could be expanded– Other electronic formats (PDF, even ASCII text)– Probabilistic model for jitter