Getting started

Your first sentiment analysis with kosac, end to end.

Install

pip install "kosac-lexicon[kiwi]"

The [kiwi] extra adds the Kiwi morpheme tokenizer — it installs from PyPI with no Java required.

Analyze a sentence

from kosac import SentimentAnalyzer

analyzer = SentimentAnalyzer("polarity")
result = analyzer.analyze("이 영화 정말 좋다")

result["features"]["polarity"]["label"]   # 'POS'
round(result["features"]["polarity"]["prob"], 3)   # 0.968

Read the result

analyze() returns a plain, JSON-serialisable dict. Each feature carries the top label, its prob, the full probs distribution, and the matched entries — each with a character span back into the original text:

for m in result["features"]["polarity"]["matches"]:
    print(f'{m["entry"]:>8}  {m["span"]}  {m["text"]!r}  {m["max_value"]}')

    이/MM  [0, 1]  '이'   NEG
  영화/NNG  [2, 4]  '영화'  POS
  정말/MAG  [5, 7]  '정말'  POS
    좋/VA  [8, 9]  '좋'   POS
    다/EF  [9, 10]  '다'  NEG

Notice that function morphemes such as 이/MM and 다/EF carry noisy polarity (their corpus distribution leans NEG). The probability aggregation weighs all the evidence — dominated here by 좋/VA (POS) — and still yields POS overall.

span lets you map a match back to the source text:

m = result["features"]["polarity"]["matches"][3]
result["text"][m["span"][0]:m["span"][1]]   # '좋'

All six features at once

KOSAC annotates six independent features. Ask for them together with features="all":

SentimentAnalyzer("all").analyze("이 영화는 정말 좋았고 너무 행복했다")

feature

label

prob

polarity

POS

0.98

intensity

Medium

1.00

expressive-type

dir-speech

1.00

nested-order

1

1.00

subjectivity-polarity

POS

1.00

subjectivity-type

Argument

0.98

Next steps