Text Analysis Lab

VADER starts with a lookup: each word token is checked against a valence lexicon where entries range from –4 (extremely negative) to +4 (extremely positive). Words not in the lexicon get a valence of 0. The raw scores are then adjusted by six rule layers applied in order:

1. Booster words — intensifiers like "very," "extremely," or "incredibly" within a 3-token window before a sentiment word increase its absolute valence by 0.293 (full boost), 0.279 (one token back), or 0.263 (two tokens back). Downplayers like "somewhat" or "marginally" apply a corresponding negative adjustment.
2. Negation — "not," "never," "without," or any word ending in "n't" within 3 tokens before a sentiment word multiplies its valence by –0.74, reversing and attenuating the signal.
3. Contrastive "but" — when a sentence contains "but," words before it are attenuated (×0.5) and words after it are amplified (×1.5), capturing the flip implied by "it was good, but very expensive."
4. Capitalization — if a sentiment word (but not all words) is in ALL CAPS, its absolute valence is boosted by 0.733 to capture emphasis like "LOVE it."
5. Punctuation — each "!" up to three adds 0.292; two or more "?" add 0.18 (three or more add 0.36), reflecting how punctuation intensifies emotional expression in social text.
6. Bigram dampeners — phrases like "kind of" or "sort of" before a sentiment word reduce its influence, capturing the hedge in "kind of nice."

After adjustments, individual valences are summed into \(S\). The compound score normalizes this sum: \[\text{compound} = \frac{S}{\sqrt{S^2 + 15}}\] Values ≥ 0.05 are labeled positive, ≤ −0.05 are labeled negative, and everything in between is labeled neutral. Separate positive, neutral, and negative proportions describe how many tokens carry each tone (each non-neutral token adds a "+1 punch" to its proportion bucket).

Ambivalence — a compound near zero can arise two ways: many positive words and many negative words cancelling out (high ambivalence) or very few sentiment words at all (genuinely neutral). The lab distinguishes these by computing \(\text{ambivalence} = \frac{2 \cdot \min(pos,\, neg)}{pos + neg}\) (0 = purely one-sided, 1 = perfectly balanced).

The NRC Word-Emotion Association Lexicon (EmoLex) was built by crowdsourcing human judgments of word–emotion associations. Annotators labeled whether each word evoked each of Plutchik's eight emotions and whether it was positive or negative. The result is a binary lookup: a word either does or does not carry each emotion.

The analyzer tokenizes each text, strips punctuation, and looks up each word (case-insensitive). For each matched token, all of its associated emotions are incremented. Scores are then normalized by the number of matched tokens (lexicon coverage) so that a short review with three matched words and a long essay with thirty matched words can be compared fairly. A high joy score, for example, means a large share of the emotionally-loaded words in that text are joy-associated.

When NRC adds insight beyond VADER: Two texts can have the same VADER compound score but very different emotion profiles. A customer review reading "thrilling and terrifying" might score slightly positive on VADER but show high joy and high fear in NRC—suggesting a mixed-intense reaction rather than calm satisfaction.

Rather than labeling emotions as discrete categories, dimensional models position affect in a continuous space. The NRC-VAD Lexicon provides continuous ratings (0–1) on three axes for each of 20,000 English words:

The Valence–Arousal plane corresponds to Russell's (1980) circumplex model of affect, which organizes emotions into four quadrants: high-V high-A (excited, happy), low-V high-A (anxious, angry), high-V low-A (calm, relaxed), and low-V low-A (sad, bored). Placing your text data in this space tells you not just the direction of emotion (positive or negative) but also its energy level.

Dominance adds a third axis capturing power dynamics. Texts that feel "in control"—commands, brand authority claims, expert opinions—show high dominance. Texts that communicate vulnerability, pleading, or powerlessness show low dominance. In consumer research, dominance differences often emerge between complaint texts (low-D) and recommendation texts (high-D) even when valence is similarly positive.