TableVision uses printed sums to check scanned table structure
In plain words
A scanned table can have readable numbers but still be wrong if the numbers are put under the wrong county or candidate. TableVision is a workflow for rebuilding and checking the layout of scanned printed tables, especially historical election returns.
It groups rows into counties, assigns figures to cells, and checks whether each column adds to its printed total before deciding to release, flag, or withhold it.
On seven clean panels, using the same docTR readings with TableVision’s layout put more figures in the right cells than docTR’s own tables did.
The research
When a scanned table’s digits are mostly readable, how can a reader tell whether each number still belongs to the right county and candidate? In historical election returns, a misplaced row or shifted column can leave figures legible while making the resulting county totals unreliable.
TableVision reconstructs county columns in scanned election tables and uses printed totals and rereads to decide which columns can be released.
On seven clean panels, filing the same docTR words with TableVision’s column comb put 97.9% of figures in the right cell, versus 83.7% in docTR’s own tables, and raised fully correct county columns from 11 of 30 to 24. Because the reader’s words stayed fixed, this comparison points to structure recovery as the source of better placement.
On three hard pages, one stored TableVision pass filed 99.2% correctly, and its gate released 51 columns, all exact.
These results come from specific historical election tables, not a random sample of historical documents. A matching sum can still hide misfiled or compensating errors, and the workflow’s reader was evaluated on only three hard pages.
Analysis
Deep
- Original paper authors
- Todd Y. Zhou, James M. Snyder
- Submitted to arXiv
- License
- CC BY 4.0