Beyond OCR: Quantifying The Field-Association Gap in Automated Electrical Nameplate Extraction

Abstract

This paper evaluates, on a public 286-image real-world dataset, the same extraction architecture implemented in the authors' nameplate web application: local optical character recognition followed by deterministic rule-based field association and engineering validation, with no call to a hosted API. The application originally used a hosted multimodal model (Gemini) for extraction; that call has since been replaced, in the application's own source, by an in-browser OCR-plus-rules pipeline, so the design evaluated here is the design the application now runs. Two OCR engines are scored with an engineering-aware protocol, field exact match, all-critical-fields-correct rate, hallucination and order-of-magnitude error rates, and validation-rule flag recall and precision, adapted from a companion review of deep-learning nameplate extraction. Reference values for the dataset were produced by a large vision-language model (Qwen3.6-35B) rather than by human annotators, so every number below is a measurement of agreement with that machine-generated reference, not of ground-truth accuracy, and is worded that way throughout rather than only in a caveat. RapidOCR-plus-rules achieves 20.5% field-level agreement with the reference labels and 67.0% precision (bootstrap 95% CI 17.5-23.9%), against 8.8% and 50.3% for Tesseract-plus-rules; both are far below the all-critical-fields-correct rate a deployment would require (1.0% and 0.3% respectively). A diagnostic breakdown splits the loss between reading and association: 52.5% of reference critical-field values can be found in the raw RapidOCR transcript, but only 20.5% are correctly assigned to their field by the deterministic parser, so 61% of the values the engine does read are lost at the association step. The 51 reference records that fail the engineering-validation rules (17.8% of the dataset) were adjudicated against their OCR transcripts: 76% are consistent with a genuine reference-label error, 10% are false positives on machines the validator does not model (synchronous, servo, geared and DC motors), and 14% could not be resolved from text alone. Every point at which this evaluation fell short of its original protocol is disclosed explicitly, most importantly that GPU-dependent recognizers and a human-verified gold-label subset could not be produced in the environment used for this analysis, and that no public transformer-nameplate dataset with field labels exists, so a motor-nameplate dataset was used to validate the pipeline design rather than the application's exact field list.