Why word embeddings distinguish 가#V, 가#J, and 가#N
Korean has many morphemes that share the same characters but have different parts of speech (POS).
Even 가 alone can be a verb (가다, "to go"), a particle (the subject particle, josa), or a noun (a unit, 가).
Bareun's word embeddings distinguish such homographic (同形異義) morphemes in a form that includes a POS hint.
That is why 가#V, 가#J, and 가#N each have a different embedding.
Same characters, different morphemes
Let's look at how 가 differs depending on context.
| Notation | POS hint | Meaning | Example sentence |
|---|---|---|---|
가#V | Predicate (V) | Stem of the verb '가다' (to go) | 학교에 가다 (go to school) |
가#J | Particle (J) | Subject particle | 비가 온다 (it rains) |
가#N | Nominal (N) | A noun such as '가' (edge) | 강가를 걷다 (walk along the riverside) |
All three are written 가, but their grammatical roles and meanings are completely different.
Bareun embeds them as separate tokens, such as 가#J (particle), 가-#V (가다), and 가#N (noun).
If you merged them into one in the embedding, the meanings of three unrelated words would be mixed into a single vector.
During training, Bareun found as many as 12,175 words like this that have the same surface form but different forms. Training on this many homographic morphemes separated by form is what leads to Bareun's improvement in disambiguation.
Putting the POS hint in the embedding key
Bareun sets the unit of its embeddings not as the written characters alone but as a combination of form + POS hint. The POS hint is written with the eight major morphological units.
| Notation | Morphological unit |
|---|---|
#N | Nominal |
#V | Predicate |
#E | Ending (eomi) |
#J | Particle (josa) |
#M | Determiner |
#A | Adverb |
#I | Interjection |
#S | Symbol |
graph TD GA[가] --> V[가#V<br/>verb stem vector]; GA --> J[가#J<br/>subject particle vector]; GA --> N[가#N<br/>noun vector];
Why form-based embeddings?
If morphemes with different meanings share the same vector, the model loses the basis for telling apart the 가 in "비가 온다" from the 가 in "학교에 가다".
When the POS hint is put into the embedding, the model learns different vectors by role even for the same characters,
which makes analysis that fits the context possible.
This distinction is directly tied to the accuracy of morphological analysis. The ability to correctly identify the 가 in 비가 온다 as a subject particle (가#J)
and the 가 in 학교에 가다 as a verb stem (가#V) comes from exactly this.
The foundation of homograph handling
Underlying Bareun's ability to handle ambiguity (the same notation with different meanings) is this form-based embedding that carries a POS hint. Homographic morphemes that could not be resolved by embeddings that looked only at notation are separated and trained by POS.
Frequently asked questions
Q. What do 가#V, 가#J, and 가#N mean?
They are three morphemes that are all written 가 but have different POS hints. 가#V is a verb stem,
가#J is a subject particle, and 가#N is a noun. Each is trained as a different vector in the embedding.
Q. Why include the POS hint in the embedding?
Even with the same characters, meaning and role differ when the POS differs. Without the POS hint, unrelated morphemes are mixed into one vector and cannot be told apart. With the hint, different vectors are trained for each role.
Q. How does this approach help with homograph handling?
When embeddings are separated by POS, there is a basis for picking the POS that fits the context even for the same notation.
It can correctly tell apart the particle 가 in 비가 온다 and the verb 가 in 학교에 가다.
Q. How many words have the same surface form but different forms?
During training, Bareun found 12,175 homographic morphemes that are written the same but have different forms. It improves disambiguation by embedding them as different tokens by form.
Q. What are POS hint notations like #N and #V?
They are notations for the eight major morphological units. They mark the morphological unit of a morpheme: nominal #N, predicate #V, ending #E, particle #J,
determiner #M, adverb #A, interjection #I, and symbol #S.