VADER on emoji posts: nltk scores emoji as 0.0, vaderSentiment scores them by name (fire = negative)
Symptom: the same emoji post gets different VADER compound scores in nltk.sentiment.vader and vaderSentiment. "good π" is 0.4404 in nltk but 0.836 in vaderSentiment; "shipped it π₯π₯" (praise) is -0.5859 in vaderSentiment and 0.0 in nltk. Fix: for emoji-heavy posts use vaderSentiment, which translates emoji to their text names before scoring. Do not rely on nltk for emoji. The lookup is literal, so check the emoji you care about (π₯ is -0.34, π‘ and π score 0.0); pre-map slang emoji to words if needed. Offline nltk: SentimentIntensityAnalyzer() raises LookupError: Resource vader_lexicon not found until you run nltk.download("vader_lexicon", download_dir=DIR) and set NLTK_DATA=DIR. Compound (vaderSentiment | nltk): π 0.7184 | 0.0; π₯ -0.34 | 0.0; fire π₯ -0.5859 | -0.34; I love it π 0.802 | 0.6369; π‘ 0.0 | 0.0. Both agree on not good (-0.3412), empty string (0.0) and Spanish text (0.0). Versions: vaderSentiment 3.3.2, nltk 3.10.4, Python 3.12.