A recent paper I wrote, now under
submission, examines the impact of
OCR errors on a large collection of scanned pages I constructed
specifically for this task and which I am making available to the
international research community to help foster similar studies. This
data set is derived from the well-known
Reuters-21578
news corpus.
"Split-000" of the Reuters corpus was first filtered to remove articles
that consist primarily of tabular data that would be inappropriate to
parse using NLP techniques (see “Medium-Independent Table Detection” (
PDF 301 kbytes)). Then I
formatted each of the remaining articles as a single page typeset in
Times-Roman 12-point font. In doing so, I discarded articles that were
either too long to fit on a single page or too short to provide a good
test case (fewer than 50 words). Of the 925 articles in the original
set, 661 remained after these various criteria were applied.
These pages were then printed on a Ricoh Aficio digital photocopier and
scanned back in using the same machine at a resolution of 300 dpi
bitonal using the copier's automatic sheet feeder. One set of pages was
scanned as-is ("orig"), another two sets were first photocopied through
one and two generations with the contrast set to the darkest possible
setting ("dark1" and "dark2"), and two more sets were similarly
photocopied through one and two generations at the lightest possible
setting ("light1" and "light2") before scanning. This resulted in a
test set totaling 3,305 pages. We then ran the resulting bitmap images
through the
Tesseract
open source OCR package.
The NLP stages I employed consist of:
- Sentence boundary detection using the MXTERMINATOR
package by Reynar and Ratnaparkhi.
- Tokenization using the Penn
Treebank tokenizer.
- Part-of-speech tagging using Ratnaparkhi's MXPOST.
Here's an example of a page from the collection (derived from Page 1 of
Split-000 of Reuters-21578):
The complete set of 3,305 pages is broken into "batches" of 100-pages
each (except for the last, which contains the final 61 pages), and
further subdivided in terms of the five types of source
document (original page, first- and second-generation dark photocopies,
and first- and second-generation light photocopies). These are then
collected into TAR archives and compressed using gzip. Click on the
links below to download the indicated archives. Note that each of these
sets is between 3 and 5 megabytes in size.
The corresponding ground-truth files for each batch are listed below.
Note that the approach I use in my work is relativistic. That is, there
is no universal ground-truth, but rather I compare the performance of
the various text analysis stages on clean and noisy inputs. An "error"
is considered to have occurred when the two sets of results differ.
There may already in fact be errors present, even for clean inputs.
COLLEGE HOME
| LEHIGH HOME | SEARCH
