Correction rules for the initial-sound rule, jamo errors, and punctuation

Bareun corrects conjugation errors, word spelling errors, and punctuation errors according to Korean grammar rules. These corrections are classified under GRAMMER (grammar and spelling rules) and WORD (word rules) of RevisionCategory, and each correction carries a rule source (rule_article) that shows "which rule it was corrected by."

GRAMMER: conjugation, particle combination, and punctuation

GRAMMER deals with grammar errors that arise when words change form or combine within a sentence.

WORD: sai-siot, the initial-sound rule, and Sino-Korean combinations

WORD deals with spelling rules for words themselves. A typical one is the initial-sound rule (dueum beopchik).

Building an NFD dictionary: searching rules by phoneme

Bareun finds correction targets by searching Hangul spelling and standard language rules based on actual dictionary data, not human intuition. To do this, it builds the dictionary in NFD (decomposed into initial, medial, and final jamo) phoneme units.

In Hangul, the visible syllable form (NFC) and its decomposed phoneme form (NFD) differ in length.

When the dictionary is decomposed into phonemes, you can search for words a rule applies to by jamo pattern, such as "stems whose medial vowel is ㅗ/ㅜ." Bareun applies a pattern search (regexp_like(nfd, ...)) to the dictionary column decomposed in NFD and extracts the set of words covered by a given spelling rule.

Based on the extracted word set, Bareun generates error data and trains the correction model with those error rules. Because the rule is applied to the entire dictionary rather than to examples in someone's head, every word the same rule applies to can be covered without omission.

14 morphological error types caught at the segmentation stage

Morphological errors such as conjugation and contraction are identified in advance at the segmentation stage of morphological analysis, before entering the correction pipeline. Of Bareun's 115 segmentation labels, 14 are labels for error detection (spell checking), and eojeols with segmentation errors are restored to the correct form.

Jamo error correction

Jamo-level errors, where an initial, medial, or final jamo is mistyped or missing while typing, are also correction targets. Bareun uses jamo-level edit distance to find a correct spelling close to the original text, and rejects changes that move too far from the original to prevent over-correction.

graph TD
  A[Eojeol to correct] --> B{Which rule is violated?};
  B -- Conjugation, particle, punctuation --> C[GRAMMER];
  B -- Sai-siot, initial-sound rule, Sino-Korean --> D[WORD];
  C --> E[Attach ReviseHelp<br>comment + rule_article];
  D --> E;
  E --> F[Correction result + rule source];

The rule source is provided in ReviseHelp

The key to grammar and word rule corrections is the basis. ReviseHelp in the response's helps provides the following.

For example, a word spacing correction carries "Hangul Spelling Rules Article 2: each word of a sentence is, in principle, written separated by spaces" in rule_article, and a standard word correction carries "Standard Language Rules principle of determination, Article 22."

Frequently asked questions

Q. Does it also fix initial-sound rule errors?

Yes. The initial-sound rule is corrected under the WORD (word rule) category. For example, 로인 is corrected to 노인, and the rule source is provided in rule_article.

Q. How are conjugation errors and word spelling errors told apart?

Grammar errors such as conjugation, particle combination, and punctuation are classified as GRAMMER, and word spelling rules such as sai-siot, the initial-sound rule, and Sino-Korean combination are classified as WORD.

Q. Where can I check the rule basis of a correction?

You can check the rule source in rule_article of the ReviseHelp in the response's helps. The reason for the correction (comment) and additional examples (examples) are provided too.

Q. What is the NFD dictionary, and why search by phoneme?

The NFD dictionary is a dictionary indexed in phoneme units by decomposing Hangul into initial, medial, and final jamo. For example, 아름다운 is 4 characters as syllables (NFC) but 10 characters as phonemes (NFD). When decomposed this way, you can search the dictionary directly by jamo pattern, such as "stems whose medial vowel is ㅗ/ㅜ," for words covered by a spelling rule, and extract them as correction and training targets.

Q. How many error types are caught at the segmentation stage?

Of Bareun's 115 segmentation labels, 14 are labels for error detection. Morphological errors such as ㅂ, ㅅ, ㄹ, 러, and 르 irregular conjugation errors, 외 contraction errors, copula ending errors, and vowel progressive assimilation errors are identified at the segmentation stage, and eojeols with errors are restored to the correct form.