Korean ambiguity dataset: 12,175 cases of semantic ambiguity

In Korean, the same characters often mean something entirely different depending on context. 가는 alone can come from '가늘다' (to be thin), '가다' (to go), or '갈다' (to grind). Bareun built 12,175 such ambiguity cases itself, and has released part of them to contribute to Korean natural language processing research.

What is ambiguity

Ambiguity is the state in which one surface form can have several morphological analysis results. A typical example is 가는.

All three look like 가는, but the analysis is completely different depending on which verb or adjective it comes from. The only clue that decides the right answer is the context.

graph TD
  A[가는] --> C1["가는 실 (a thin thread) → 가늘/VA"];
  A --> C2["학교에 가는 길 (the way to school) → 가/VV"];
  A --> C3["칼을 가는 사람 (a person sharpening a knife) → 갈/VV"];

Why we built an ambiguity dataset

For this purpose, the Bareun team built 12,175 ambiguity cases itself. Each case captures how the same surface form should be analyzed differently depending on context, which becomes the fuel for the model to learn context-based judgment.

Ambiguity resolution evaluation set: 35,396 sentences / 8,285 surface forms

Beyond training, Bareun built a separate evaluation dataset to measure how well it resolves ambiguity. It covers 35,396 sentences / 8,285 surface forms, and surface forms with too little data were reinforced with additional data.

A typical example is "트는". The eojeol (space-delimited word unit) looks the same, but the analysis differs.

VCP restoration evaluation: 12,129 sentences

In Korean, the copula '이' (VCP) is often omitted. Bareun extracted 12,129 sentences from the Modu Corpus (the National Institute of Korean Language's "Corpus for Everyone") in which '이' is included or omitted, and evaluated whether it correctly restores the omitted '이'.

Full 47-POS tagging evaluation: 150,000 sentences

The strictest evaluation is whole-sentence tagging. On 150,000 sentences from the Modu Corpus, if even one morpheme in a sentence is wrong, the whole sentence counts as incorrect. It is a demanding standard: all 47 parts of speech have to be right for a sentence to count as correct.

Public dataset

Frequently asked questions

Q. What is Korean ambiguity?

It is the phenomenon in which the same surface form can have several morphological analysis results depending on context. For example, the analysis of 가는 differs depending on whether it comes from '가늘다', '가다', or '갈다'.

Q. How much ambiguity data has Bareun built?

It built 12,175 ambiguity cases itself. They capture how the same surface form should be analyzed differently depending on context, so the model learns context-based judgment.

Q. How is ambiguity resolution performance evaluated?

It is measured on a separately built evaluation set of 35,396 sentences / 8,285 surface forms. Using cases such as "트는" (틀다/트다) where the eojeol looks the same but the analysis differs, it checks whether Bareun picks the analysis that fits the context.

Q. What is the VCP (copula '이') restoration evaluation?

It checks whether the often-omitted copula '이' is correctly restored. It extracts 12,129 sentences from the Modu Corpus in which '이' is included or omitted, and measures whether the omitted '이' is brought back in the analysis, as in 평갑니다 (평가+이+ㅂ니다).

Q. How is full 47-POS tagging evaluated?

It is evaluated on 150,000 sentences from the Modu Corpus with a strict standard: if even one morpheme in a sentence is wrong, the whole sentence counts as incorrect.

Q. Is the ambiguity dataset public?

Yes. Part of it is public at github.com/bareun-nlp/korean-ambiguity-data, and you can use it in Korean natural language processing research.