Negation, intensifiers, and tuning

A plain lexicon lookup gets 안 좋다 (“not good”) wrong, because it never sees the negation. kosac offers opt-in composition to fix this.

The problem

from kosac import SentimentAnalyzer

plain = SentimentAnalyzer("polarity")
plain.analyze("이 영화는 안 좋다")["features"]["polarity"]["label"]   # 'POS'  (wrong!)

The positive word 좋 dominates, so the sentence scores POS despite the 안.

Turn on negation

negating = SentimentAnalyzer("polarity", negation=True)
negating.analyze("이 영화는 안 좋다")["features"]["polarity"]["label"]   # 'NEG'

When a negation marker (안 / 못 / 않 / 없 …) falls within a window of a matched expression, its POS↔NEG mass is swapped:

sentence	plain	negation=True
이 영화는 안 좋다	POS (0.66)	NEG (0.99)
별로 안 좋다	POS (0.59)	NEG (0.84)

Each affected match is flagged in the result:

for m in negating.analyze("이 영화는 안 좋다")["features"]["polarity"]["matches"]:
    print(m["entry"], m["negated"])
# 이/MM False
# 영화/NNG True
# 는/JX 안/MAG False
# 좋/VA True
# 다/EC True

(영화/NNG and 다/EC are also flagged here because they fall inside the window around 안 — see the caveat below.)

Intensifiers

Intensifier markers (정말 / 너무 / 매우 …) up-weight a nearby match, sharpening the dominant label:

plain   = SentimentAnalyzer("polarity")
boosted = SentimentAnalyzer("polarity", intensifier=True, intensifier_factor=2.0)

plain.analyze("정말 좋다")["features"]["polarity"]["probs"]["POS"]     # 0.927
boosted.analyze("정말 좋다")["features"]["polarity"]["probs"]["POS"]   # 0.992

You can enable both at once: SentimentAnalyzer("polarity", negation=True, intensifier=True).

Aligning Kiwi to the lexicon

align=True seeds Kiwi’s user dictionary with the lexicon’s entries, so its segmentation matches the lexicon’s surface/POS units more often:

analyzer = SentimentAnalyzer("polarity", align=True)
analyzer.analyze("이 영화 정말 좋다")["features"]["polarity"]["label"]   # 'POS'

Customising the markers

The window size, boost factor, and marker sets are all configurable, and the defaults are exported for reuse. Korean negation is a small closed class (안 / 못 / 않 / 말 / 아니) already covered by DEFAULT_NEGATIONS; intensifiers are an open class, so adding colloquial ones is the common case:

from kosac.analyzer import DEFAULT_INTENSIFIERS

custom = SentimentAnalyzer(
    "polarity",
    intensifier=True,
    intensifiers=DEFAULT_INTENSIFIERS | {"완전/MAG"},   # add 완전 ("totally")
    window=3,                                            # tokens to look around a match
)
SentimentAnalyzer("polarity").analyze("이 영화 완전 좋다")["features"]["polarity"]["probs"]["POS"]   # 0.888
custom.analyze("이 영화 완전 좋다")["features"]["polarity"]["probs"]["POS"]                          # 0.987

⚠️ Caveat

Negation/intensifier handling is a window-based heuristic. With a wide window an adjacent morpheme can occasionally be flagged by mistake; the final label is usually unaffected, but tune window (or post-process the per-match negated flags) if you need precise negation scope.