OCR Swaps Decimal Points and Commas in Invoice Amounts: Causes and Fixes

Issue: OCR misreads punctuation — decimals become commas, currency symbols drop, dashes get interchanged — silently corrupting parsed amounts

Frequency: Common

Symptoms

• Decimal points become commas (or vice versa), breaking numeric parsing
• Currency symbols misread or dropped from extracted amounts
• Hyphens, dashes, and minus signs interchanged, flipping signs or breaking ID formats

Root Cause Small punctuation marks are often damaged in scans or rendered differently across fonts. Regional formatting differences (. vs , for decimals) add ambiguity.

Example

Input Image: Total: $1,234.56
Expected: 1234.56
Actual: 1.23456 (comma interpreted as decimal)

Result: Invoice processed for wrong amount

How to Fix Punctuation and Decimal Misreads in OCR

Mitigation Strategies

Prevention

1. Locale-aware numeric parsing: Detect document locale/currency (from sender metadata, currency symbol, or address) before parsing and apply the corresponding decimal/thousands-separator convention, since . vs , ambiguity is fundamentally a locale-detection problem rather than an OCR problem. Trade-off: requires reliable locale signals; documents with no locale context still default to a guess.
2. Dual-pass raw-and-parsed extraction with cross-check: Extract both the raw character string (1,234.56) and a structurally-parsed numeric value, then verify the parsed value’s magnitude and decimal-position are consistent with the raw string’s punctuation pattern before accepting it, catching cases like a comma being misread as a decimal point. Trade-off: adds a verification step and requires a canonical raw-string extraction path in addition to numeric parsing.
3. Currency-symbol-first parsing order: Detect and extract the currency symbol before parsing the numeric value, since knowing the currency (and therefore its conventional format) resolves ambiguity that pure digit/punctuation recognition cannot. Trade-off: symbol misreads (dropped or misrecognized currency marks) can still propagate the wrong format assumption.

Detection & Response

1. Sanity-bound range validation: Reject or flag parsed values outside an expected range for the field/document type (e.g., an invoice line item over $1M, or a total under $1 when line items sum higher), routing the flagged value for correction rather than passing a misparsed amount downstream.
2. Value-distribution monitoring: Track the distribution of parsed values for a field across a document source; a sudden cluster of unusually small values (e.g., many totals under $1 when historically larger) signals systematic comma-as-decimal misparsing rather than isolated errors.
3. Downstream parsing exception correlation: Monitor parsing/type exceptions in downstream financial systems and correlate them back to the originating document source and extraction confidence, since a spike there is often the first visible symptom of a punctuation misread upstream.

Architecture Patterns

1. Confidence-gated human-in-the-loop review queue: Route any parsed value that fails sanity bounds or raw/parsed cross-check to human review before it reaches payment or accounting systems.
2. Locale-inference-then-parse pipeline: Architect extraction as locale/currency detection first, then format-specific numeric parsing second, rather than a single locale-agnostic parser applied uniformly to all documents.
3. Reconciliation checkpoint: For invoices, verify that extracted line items sum to the extracted total within tolerance; a mismatch is strong evidence of a punctuation-driven misparse in one of the values and should trigger reprocessing before acceptance.

Metrics

1. sanity_bound_rejection_rate: Target: < 1% of parsed numeric fields; Alert threshold: > 4%
2. line_item_total_reconciliation_mismatch_rate: Target: < 2%; Alert threshold: > 6%
3. downstream_parsing_exception_rate: Target: < 0.5%; Alert threshold: > 2%
4. locale_detection_confidence: Target: > 95% of documents with a confidently detected locale; Alert threshold: < 85%

Alerts

1. Value Distribution Anomaly (P2): Condition - a document source shows a spike in parsed values below $1 or an unexplained order-of-magnitude shift. Action: Sample documents, verify locale detection and decimal parsing for that source.
2. Reconciliation Mismatch Spike (P1): Condition - line-item-to-total mismatch rate exceeds 6% for a source. Action: Halt auto-posting for that source, route to review, investigate punctuation misparse.
3. Downstream Exception Spike (P2): Condition - payment/accounting system parsing exceptions traced to extraction exceed 2% in a rolling window. Action: Correlate to source/template, escalate to extraction pipeline owner.

References

• Why OCR Is the Weakest Part of Document AI - Punctuation issues
• IDP Accuracy Reckoning 2026 - OCR accuracy challenges