Design and optimization of NLP algorithms for optically-aware intelligent diagnosis of English text errors

Automated diagnosis of English text errors is essential for computer-assisted language learning (CALL), especially given the diverse error patterns in second-language (L2) writing. In real-world settings, such as processing scanned or handwritten essays, optical acquisition quality directly affects textual accuracy. To address this, we integrate optical character recognition (OCR) enhancement and multi-spectral imaging into the diagnostic pipeline. Existing approaches often conflate detection with correction: rule-based systems lack generalizability, while neural models prioritize fluency over transparency and struggle across languages. A modular natural language processing (NLP) architecture is proposed that decouples token-level error classification from constrained correction. Input text is first refined using illumination-invariant optical features to suppress OCR errors, and optical uncertainty, derived from metrics such as edge sharpness or spectral contrast, is fused with linguistic confidence to support robust diagnosis. The model combines a shared Bidirectional Encoder Representations from Transformers (BERT) encoder, a Conditional Random Field (CRF)-enhanced classifier for 12 Corpus -First Certificate in English (CLC-FCE) error types, and an attention entropy-based calibration module. Evaluated on CoNLL-2014 and a new DiagLang-6 test set spanning six First Language (L1) backgrounds, including optically acquired learner texts, it achieves a Detection F1 of 78.4±1.3 and competitive correction quality (F0.5 = 61.2±0.8), with narrow cross-L1 variance (3.8 points) and high calibration accuracy (79.6±2.0). This work establishes a physically grounded, optically aware framework for equitable and deployable educational NLP.