Why Bareun morphological analysis reaches 99.6% accuracy
Bareun's Korean morphological analysis accuracy is 99.6% for POS tagging and 99.7% for eojeol (space-delimited word unit) split/restore.
These numbers did not come by chance. They are the result of many decisions stacked up, from data cleaning to model architecture design.
This article lays out where that high accuracy comes from.
Two accuracy metrics
| Metric | Accuracy | Meaning |
|---|---|---|
| POS tagging | 99.6% | Share of morphemes assigned the correct one of 47 POS |
| Eojeol split/restore | 99.7% | Share of eojeol split correctly into morphemes with base forms restored |
POS tagging accuracy shows "whether the labels were attached correctly", and eojeol split/restore accuracy shows "whether the eojeol was cut correctly and returned to its original form".
How the 99.6% was measured
This figure comes from a strict, large-scale evaluation, not a small sample.
- Whole-sentence tagging criterion: On 150,000 sentences of the Modu Corpus (Modu-ui Malmungchi), if even one morpheme in a sentence is wrong, the whole sentence is counted as wrong. It is a demanding criterion that requires getting all 47 POS right to count as correct.
- Ambiguity evaluation set: We collected cases where the same eojeol is analyzed differently depending on context, at a scale of 35,396 sentences / 8,285 surface forms, and measure whether Bareun picks the analysis that fits the context.
- VCP (copula '이') restoration evaluation: We separately evaluate whether the often-omitted '이' is correctly restored in the analysis, using 12,129 sentences extracted from the Modu Corpus.
The results measured from these several angles come together as the POS tagging figure of 99.6%.
Three pillars that build accuracy
graph TD A[Corpus cleaning] --> ACC[High analysis accuracy]; B[Separate segmentation and tagging stages] --> ACC; C[In-house tagging guidelines] --> ACC;
1. Corpus cleaning
We fix tagging errors in the Sejong, Modu, and news corpora, align them for consistency, and train on them. When a model trains on clean data, it does not learn confusing signals.
2. Separating the segmentation stage from the tagging stage
Bareun splits "where to cut" (segmentation) and "what POS it is" (tagging) into separate models. Each model solves only a simpler, clearer problem, so accuracy is higher than when one model handles everything.
3. In-house tagging guidelines
Criteria for segmentation and POS handling are set consistently word by word, as in 누가 → 누 + 가.
The same word is always analyzed the same way, so results are stable.
Added robustness from spacing augmentation and out-of-vocabulary handling
Not shaken by informal input
Training with word spacing augmentation lets Bareun find morpheme boundaries even when users get word spacing wrong, and it guesses new words that are not in the dictionary from context. This robustness supports accuracy in real-world use.
All these elements together produce accuracy in the high 99% range, a level that a dictionary-based analyzer would find hard to reach.
Frequently asked questions
Q. How accurate is Bareun's morphological analysis?
POS tagging accuracy is 99.6%, and eojeol split/restore accuracy is 99.7%.
Q. How is this accuracy possible?
It rests on three pillars: corpus cleaning, separation of the segmentation stage from the tagging stage, and consistent in-house tagging guidelines. Word spacing augmentation and context-based guessing of out-of-vocabulary words add robustness.
Q. How do POS tagging accuracy and eojeol split/restore accuracy differ?
POS tagging accuracy is the share of morphemes given the correct POS, and eojeol split/restore accuracy is the share of eojeol split correctly into morphemes with base forms restored.
Q. How was the 99.6% measured?
Whole-sentence tagging is evaluated on 150,000 sentences of the Modu Corpus, with a strict criterion: if even one morpheme is wrong, the whole sentence is counted as wrong. The result also adds an ambiguity evaluation set of 35,396 sentences and a VCP ('이') restoration evaluation of 12,129 sentences.