Challenge
The client's document workflow depended on reading Korean reliably at production volume, and the accuracy of that reading set the ceiling for everything downstream: search, extraction, and compliance. A missed character is a missed hit, and a wrong digit is wrong data. Korean made this hard in ways an English-tuned pipeline cannot absorb. Hangul is a composed syllabic script rather than a flat alphabet, and real scanned documents mix Hangul, Hanja, Latin, and digits on a single line. An off-the-shelf OCR engine tuned for Latin text demonstrated well in an afternoon and then failed on the real archive, where scan quality, layout, and mixed scripts all vary. Labeled training data for the target documents was scarce, which ruled out simply throwing more data at the problem.
The client's document workflow depended on reading Korean reliably at production volume, and the accuracy of that reading set the ceiling for everything downstream: search, extraction, and compliance. A missed character is a missed hit, and a wrong digit is wrong data. Korean made this hard in ways an English-tuned pipeline cannot absorb. Hangul is a composed syllabic script rather than a flat alphabet, and real scanned documents mix Hangul, Hanja, Latin, and digits on a single line. An off-the-shelf OCR engine tuned for Latin text demonstrated well in an afternoon and then failed on the real archive, where scan quality, layout, and mixed scripts all vary. Labeled training data for the target documents was scarce, which ruled out simply throwing more data at the problem.
Solution
We built a Korean-first pipeline rather than tuning an English-centric one. The decisive modeling choice was to recognize the roughly 51 jamo components and compose syllables from them, instead of trying to learn more than 11,000 syllable classes directly, which is the difference between a recognizer that needs an enormous labeled corpus and one that does not. Five stages run on the line: preprocessing to deskew, binarize, and normalize resolution; layout detection; text-line detection; jamo-aware recognition with mixed-script routing so Hangul, Hanja, Latin, and digits are each handled correctly; and post-processing that composes and corrects structured text with confidence scores. Underneath sit two disciplines that made it hold: a synthetic-data and augmentation strategy to cover the cases real labels did not, and evaluation on real held-out documents measured by character and word error rate per script and per quality band.
We built a Korean-first pipeline rather than tuning an English-centric one. The decisive modeling choice was to recognize the roughly 51 jamo components and compose syllables from them, instead of trying to learn more than 11,000 syllable classes directly, which is the difference between a recognizer that needs an enormous labeled corpus and one that does not. Five stages run on the line: preprocessing to deskew, binarize, and normalize resolution; layout detection; text-line detection; jamo-aware recognition with mixed-script routing so Hangul, Hanja, Latin, and digits are each handled correctly; and post-processing that composes and corrects structured text with confidence scores. Underneath sit two disciplines that made it hold: a synthetic-data and augmentation strategy to cover the cases real labels did not, and evaluation on real held-out documents measured by character and word error rate per script and per quality band.
Results
The pipeline reached 96.8 percent accuracy on the documents actually received, sustained at a throughput of more than 50,000 documents a month. Critically, that number was measured on the real, messy archive rather than a clean benchmark, so the downstream steps that depend on the text could be built on a figure the team could trust. The client got a production document pipeline that holds on the documents it really sees, at the volume it really has.
The pipeline reached 96.8 percent accuracy on the documents actually received, sustained at a throughput of more than 50,000 documents a month. Critically, that number was measured on the real, messy archive rather than a clean benchmark, so the downstream steps that depend on the text could be built on a figure the team could trust. The client got a production document pipeline that holds on the documents it really sees, at the volume it really has.
