Character Error Rate (CER) is a standard metric for OCR and vision-language models, but it treats every mistake as equal. A single typo and a misspelled proper name incur the same penalty, even though the latter is usually far more damaging in real-world use.
To address this, the authors propose semCER, a severity-weighted variant. semCER keeps CER's alignment and denominator, so it remains directly comparable to the original metric, but it charges each grapheme edit according to its severity rather than a flat cost.
The paper argues that this better reflects the practical harm of OCR errors. While the abstract does not specify the exact weighting scheme, the intent is clear: evaluation should account for how much an error matters, not just how many errors occur.