polVADER

Lexicon-and-rule-based sentiment analysis for Polish text, in the style of VADER. No training required at inference time — polVADER scores text directly from a Polish sentiment lexicon plus a rule layer for negation, intensifiers, capitalization, punctuation emphasis, and contrastive conjunctions (ale, jednak, ...).

Install

pip install polvader
python -m spacy download pl_core_news_lg

polVADER uses spaCy's pl_core_news_lg model for Polish tokenization and lemmatization; it is not bundled with the package and must be downloaded once after install.

Usage

from polvader import Lexicon

lex = Lexicon()
scores = lex.polarity_scores("To był wspaniały dzień pełen szczęścia.")
print(scores)
# {'neg': 0.0, 'neu': 0.xxx, 'pos': 0.xxx, 'compound': 0.879}

polarity_scores() returns the same {neg, neu, pos, compound} contract as the original English VADER. compound is a single normalized score in [-1, +1]; neg/neu/pos are proportions of the text's sentiment-bearing content.

Batch scoring (uses spaCy's nlp.pipe internally, much faster than a loop):

results = lex.score_batch([
    "Świetny produkt, polecam!",
    "Nigdy więcej tu nie wrócę.",
], batch_size=256)

For social-media text (tweets, comments — hashtags, @mentions, emoji), run preprocess_social() first:

from polvader import preprocess_social

text = preprocess_social(raw_tweet)
scores = lex.polarity_scores(text)

Lexicon

By default Lexicon() loads the coverage-expanded, weight-tuned lexicon (~29.5k words): the original ~8.5k-word hand-built Polish valence lexicon, expanded via K-NN over PLLuM-8B's static input-embedding table (not a contextual/forward-pass embedding — benchmarked as equal-or-better and far cheaper to compute) across a multi-domain Polish corpus (tweets, product/hotel/service reviews, general sentiment text), then weight-tuned end-to-end against those same domains. Pass expanded=False for the smaller, untuned ~8.5k-word base lexicon instead:

lex = Lexicon(expanded=False)

Modifier system

polarity_scores() applies, on top of the raw lexicon lookup: negation (nie, multi-word negators), booster/dampener adverbs, ALL-CAPS emphasis, exclamation/question-mark emphasis, sentence-aware scoring for multi-sentence text, and contrastive-conjunction reweighting (text after "ale"/"jednak" counts more than text before it). Each category can be disabled independently via keyword flags on polarity_scores() for ablation/diagnostic purposes — see its docstring for the full flag list.

License

MIT

Funding disclosure

This work was supported by Narodowe Centrum Nauki (National Science Centre, Poland) under Grant 2020/38/A/HS6/00066.