Counting sentiment words¶

Instead of a model’s probability, you can classify text by counting sentiment words — the number and proportion of positive vs. negative words. kosac provides this word-counting (frequency) method via count(). It is the approach commonly used in content analysis and social-science studies.

Per-document counts¶

count() tallies each matched morpheme by its dominant label (max.value):

from kosac import SentimentAnalyzer

analyzer = SentimentAnalyzer("polarity")
c = analyzer.count("서비스가 형편없고 너무 실망스러웠다")["features"]["polarity"]

c["counts"]        # {'COMP': 0, 'NEG': 3, 'NEUT': 0, 'None': 0, 'POS': 2}
c["proportions"]   # {'NEG': 0.6, 'POS': 0.4, ...}
c["total"]         # 5
c["label"]         # 'NEG'  (most frequent label)

Over a corpus¶

count_frame() returns one row per document, with a count column per label:

texts = [
    "이 정책은 정말 훌륭하고 만족스럽다",
    "서비스가 형편없고 너무 실망스러웠다",
    "그저 그런 평범한 하루였다",
]
analyzer.count_frame(texts)

                  text polarity.label  polarity.total  polarity.NEG  polarity.POS  ...
이 정책은 정말 훌륭하고 만족스럽다            NEG               8             5             3
서비스가 형편없고 너무 실망스러웠다            NEG               5             3             2
     그저 그런 평범한 하루였다           COMP               5             2             0

⚠️ Caveat: function morphemes add noise¶

Look at row 0: a clearly positive sentence is counted as NEG, because every matched morpheme is counted — including particles and endings (조사, 어미) whose lexicon polarity is noisy. This is a known limitation of the pure word-counting approach.

For per-sentence classification, prefer the probability method analyze(), which weighs the evidence instead of counting it:

analyzer.analyze_frame(texts)

                  text polarity.label  polarity.prob
이 정책은 정말 훌륭하고 만족스럽다            POS           0.59
서비스가 형편없고 너무 실망스러웠다            NEG           0.66
     그저 그런 평범한 하루였다            NEG           0.46

Now row 0 is correctly POS.

When the count method shines: corpus-level trends¶

Word counts are most useful as descriptive statistics over a whole corpus, not single sentences. Sum the per-document counts to gauge the overall tone of a collection:

frame = analyzer.count_frame(texts)
totals = frame[["polarity.POS", "polarity.NEG", "polarity.NEUT"]].sum()
totals.to_dict()
# {'polarity.POS': 5, 'polarity.NEG': 10, 'polarity.NEUT': 1}

positivity = totals["polarity.POS"] / (totals["polarity.POS"] + totals["polarity.NEG"])
round(positivity, 3)   # 0.333

With only three sentences here the noise still dominates (positivity is 0.33 despite a balanced set) — which is exactly why this method needs a large corpus. Across thousands of documents the per-morpheme noise averages out and the positive/negative ratio becomes a meaningful indicator, as used in the literature.

Reading texts from a file¶

count_frame / analyze_frame take any iterable of strings, so a pandas column works directly:

import pandas as pd

df = pd.DataFrame({"text": texts})         # e.g. pd.read_csv("reviews.csv")
analyzer.count_frame(df["text"]).shape     # (3, 8)