ambrosfitz/webster1865-headwords-cleaned

Webster 1865 cleaned headwords

This dataset contains normalized and deduplicated headwords extracted from the 1865 Webster dictionary. It is derived from the raw extraction in ambrosfitz/webster1865-headwords and is intended for cross-edition joins with the cleaned Webster 1828 headwords and a 1913 Webster reference index.

Main output

headwords_cleaned.txt contains 51,584 sorted, lowercase, unique join keys. The compatibility alias headwords_deduped.txt has identical contents.

Normalization removes apostrophe-like stress marks and dictionary syllabification hyphens while preserving likely lexical compound hyphens. The pipeline used Gemma 4 E4B with a closed action vocabulary: the model could keep hyphens, remove hyphens, or assign a quarantine category, but could not invent corrected spellings.

Audit summary

• Raw extraction occurrences: 87,157
• Accepted occurrences: 75,834
• Quarantined occurrences: 11,323
• Unique cleaned headwords: 51,584
• Duplicate occurrences removed: 24,250
• Headwords matched to the 1913 reference: 32,518 (63.0%)
• Hyphenated headwords retained: 4,644

All structural output checks passed: the main list is sorted, lowercase, duplicate-free, blank-free, restricted to letters and internal hyphens, and has A-Z coverage.

Files

• headwords_cleaned.txt: primary cleaned headword list
• headwords_deduped.txt: compatibility alias
• audit_report.json: counts, policies, model metadata, and validation results
• quarantined.jsonl: rejected occurrences and reason codes
• merge_audit.json: canonical-to-raw merge provenance
• headword_groups.jsonl: one record per canonical headword
• headword_occurrences.jsonl: accepted occurrence-level provenance
• headwords_1913_matched.txt: cleaned forms found in the 1913 reference
• headwords_1913_unmatched.txt: cleaned forms absent from that reference
• headwords_hyphenated_review.txt: all retained hyphenated forms
• headwords_hyphenated_unmatched_review.txt: retained hyphenated forms absent from the 1913 reference and therefore the highest-priority manual review set
• clean_webster_1865_headwords.py: reproducible cleaning pipeline

The raw scan-derived extraction is intentionally not duplicated here. It remains in ambrosfitz/webster1865-headwords.