Skip to content

ADR-v2-079 — Self-Learning Correction Dictionary

Status: Accepted (2026-07-02) · Wave J Context links: [[adr-v2-026-phonetic-corrector]] (pronunciation vs edit history), [[adr-012-self-improvement-loop]], [[adr-v2-050-redaction-ink]], [[adr-011]]

Context

Wave J research (#6) — reuse the opt-in EditWatcher read-back: when you fix the same ASR-output→ edit pair repeatedly ("yaz says" → "YazSes"), promote it to a boundary-guarded, longest-match find→replace applied automatically to future output — a high-precision patch table mined from your own accepted edits. Distinct from the Phonetic Corrector (pronunciation-based), the vocabulary (biases initial_prompt), and tune (config diffs) — this rewrites output from edit history. Anchor: arXiv 2406.07589 ("Tag and correct: high-precision post-editing of ASR errors").

Decision

Add an opt-in Self-Learning Correction Dictionary: [corrdict] enabled=false, min_support=3. Pure cores: mine_substitutions(events, min_support) counts (wrong→right) edit pairs and keeps only those meeting min_support with an unambiguous dominant correction, and apply_corrections(text, table) applies them longest-match, word-boundary-guarded. Dependency-free and deterministic (it can only apply corrections it has actually seen — it can't hallucinate new errors). A seq-tagger for generalization is deferred. OFF by default.

Consequences

  • Learns your recurring ASR errors and fixes them automatically; high precision by construction.
  • Distinct from Phonetic Corrector / vocabulary / tune.
  • Privacy (ADR-011/012): derived only from local edits, stored in the encrypted corpus, honours redact_patterns; no keystroke logging.
  • Caveat: only exact seen substitutions (min-support, dominant) → generalization is the deferred tagger tier; off by default.