Ikibazo gihuza ibibazo · Hagati Urwego
OCR Ukuri: Ikiranga n'Ijambo Ikosa Ikigereranyo
Ikizamini CER na WER ibisobanuro, ibisanzwe hamwe no guhitamo ibimenyetso, nimpamvu guteranya ibiciro bya OCR bikenera kugenzura inyandiko yihariye.
Ikibazo 1 Bya 8
Nigute igipimo cyamakosa yimiterere gikunze kubarwa?
Reba buri kibazo8 ibibazo
Ibisubizo nibisobanuro bigaragara mugihe ufata ikibazo.
- Nigute igipimo cyamakosa yimiterere gikunze kubarwa?
- (Substitutions + deletions + insertions) divided by the number of reference characters
- Correct characters divided by the number of recognized characters
- The number of reference words divided by total edits
- Substitutions only divided by the number of reference characters
- WER itandukaniye he na CER?
- It uses the recognized words as the denominator in all cases
- It measures semantic similarity rather than edit distance
- It counts edit operations at the word level and divides by reference words
- It measures only punctuation differences
- Kuki CER cyangwa WER ishobora kurenga 1.0?
- Substitutions and deletions alone can exceed the reference length even without insertions
- Character error rate divides edits by the number of recognized characters
- Whitespace normalization can make a score exceed one even when the edit count is smaller than the reference
- A large number of insertions can make the edit count exceed reference length
- Ni ubuhe buryo busanzwe butuma CER igereranya yororoka muri sisitemu?
- Lowercase strings but leave punctuation, whitespace, and Unicode handling unspecified
- The exact Unicode, punctuation, and whitespace transformations applied consistently to the reference and output
- Remove punctuation from the reference but preserve it in the recognized output
- Normalize only the recognized output while leaving the reference transcript unchanged
- Why might WER comparisons be difficult across languages?
- WER does not use text tokens
- All languages define whitespace-delimited words identically
- WER ignores reference text length
- Word segmentation and tokenization rules can differ across scripts and languages
- Niki muri rusange CER ishobora kwihisha mubikorwa byakazi?
- A critical error in a short identifier, total, or table field
- That every character was substituted
- That reference normalization was not used
- That OCR has no impact on downstream extraction
- How should two OCR systems be compared fairly?
- Compare training loss from one model with test CER from another
- Choose different reference documents so each system sees its best case
- Use the same reference set, normalization, tokenization, and scoring implementation
- Use each system’s preferred normalization and compare raw scores
- For invoices, which measure checks whether dates or totals are correct as complete fields beyond CER or WER?
- Table-cell boundary alignment for the page layout
- Field-level correctness for extracted dates, totals, and identifiers
- Per-document mean character error rate
- Confidence calibration for individual character predictions