A Divergence-Driven Framework for Chinese Literary Text Digitization: Multi-Engine OCR Accuracy Evaluation and Multi-LLM Post-Correction

The digitization of Chinese literary texts for corpus construction faces three challenges: vertical typesetting, a set covering more than 80,000 Chinese characters codified in China’s national standard GB 18030, and frequent use of rare characters. To address gaps in existing research on multi-engine comparison and accuracy quantification, this study adopts a three-phase framework: engine evaluation, workflow design, and end-to-end validation. Engine evaluation comprises two experiments. Experiment 1, on 13 pages of vertical Traditional Chinese, compares three engines: Kandian Guji v2.0.7 (Kandian) achieves 0.25% Character Error Rate (CER), far surpassing PaddleOCR-VL-1.5 (Paddle1.5) at 4.77% and ABBYY FineReader 15 (ABBYY) at 5.96%. Experiment 2 extends to four engines on 21 pages of horizontal Simplified Chinese by adding PP-OCRv5 (PP5); Paddle1.5 achieves the lowest CER at 0.74%, and its error distribution diverges across all three error classes from Kandian’s, supporting a multi-engine scheme. ABBYY and PP5 are excluded from the final workflow due to error patterns. The four-stage workflow is adopted for the Simplified Chinese scenario: multi-engine Optical Character Recognition (OCR), character-level diff detection, multi-Large Language Model (LLM) arbitration, and final human proofreading. Experiment 3 validates the end-to-end workflow on 20 pages of Simplified Chinese text, where Kandian and Paddle1.5 produce initial CERs of 0.54% and 0.28%; diff detection identifies 62 divergences, of which 49 reach consensus, reducing CER to 0.26%; human proofreading reduces it to 0.00%. For the Traditional Chinese scenario, a rule-based script grounded in character co-occurrence legitimacy is adopted, reducing CER from 0.286% to 0.074%. The framework offers a reproducible quality-control method for Chinese literary digitization.