Text Analysis Lab

Text Analysis VADER · NRC · VAD

Qualitative data—customer reviews, survey open-ends, social posts, interview excerpts—carries meaning that goes far beyond a simple positive/negative label. This lab gives you three complementary methods for quantifying that meaning: sentiment (the overall positive–negative charge of a text), emotion profiling (which of eight discrete emotions are present), and dimensional affect (how pleasant, activated, or dominant the language feels). Run one method at a time to go deep, or run all three at once and compare what each lens reveals—and what it hides.

OVERVIEW & WHAT THIS TOOL SOLVES

Reducing thousands of customer reviews or social posts to a single "positive vs. negative" average hides critical marketing nuance: a 2-star review fueled by high-arousal anger and disgust demands an immediate PR response, whereas one driven by low-arousal sadness or disappointment calls for product or service recovery. The Sentiment & Affect Lab analyzes unstructured text through three complementaryValidated lexicons—VADER (rule-aware overall valence), NRC EmoLex (eight discrete Plutchik emotions), and NRC-VAD (Valence, Arousal, and Dominance dimensions).

🎯 Questions It Answers

Overall Polarity (VADER): How positive, negative, neutral, or ambivalent is each customer comment after accounting for negation, caps, and intensifiers?
Discrete Emotions (NRC): Which specific emotions—anger, fear, disgust, sadness, joy, trust, anticipation, or surprise—dominate our brand mentions?
Energy & Control (VAD): Does customer language feel calm vs. activated (Arousal) and empowered vs. powerless (Dominance)?

📂 What Data You Need

Text CSV or Preset Case Study: Customer reviews, social posts, survey verbatims, or ad copy.
Optional Grouping Column: Brand, competitor, product tier, or campaign wave to compare profiles across segments.
Analysis Mode Selection: Run VADER, NRC Emotions, VAD Circumplex individually, or all three side by side.

💡 What You'll Discover

Token-Level Rule Highlighting: Exact inspection of how boosters, negations, and "but" flips alter each review's score.
8-Emotion Radar & Bar Profiles: Normalized Plutchik emotion shares across documents and segments.
Russell's Circumplex Map: 2D Valence × Arousal scatter and Dominance comparisons across brands.

🔍 Key Nuances: VADER's 6 Rule Layers, NRC Emotions & Russell's Circumplex

📊

Sentiment Analysis (VADER)

How positive or negative is this text, overall?

Scores each token against 7,500+ lexical entries (−4 to +4), then applies six rule layers—boosters ("very"), negation ("not", ×−0.74), contrastive "but" (×0.5 before, ×1.5 after), ALL CAPS, punctuation ("!"), and bigram dampeners ("kind of")—to produce a normalized compound score.

VADER · Hutto & Gilbert, 2014

🎭

Emotion Profiling (NRC EmoLex)

Which discrete emotions does this text carry?

Maps 14,000+ English words onto Plutchik's eight primary emotions (anger, anticipation, disgust, fear, joy, sadness, surprise, trust) normalized by matched-token count so short and long reviews compare fairly.

NRC EmoLex · Mohammad & Turney, 2013

🧭

Dimensional Affect (NRC-VAD)

How pleasant, energized, and in-control does the language feel?

Rates 20,000 words from 0 to 1 on Valence (pleasure), Arousal (activation), and Dominance (control), placing every document in Russell's circumplex (e.g., high-V high-A = excitement; low-V high-A = rage/anxiety).

NRC-VAD · Mohammad, 2018

📐 Under the Hood: Model Equations & Variable Definitions

1. VADER Compound Normalization, Ambivalence Index & Lexicon Averages

\[ \text{compound} = \frac{S}{\sqrt{S^2 + \alpha}} \quad (\alpha = 15), \qquad \text{ambivalence} = \frac{2 \cdot \min(\text{pos},\, \text{neg})}{\text{pos} + \text{neg}}, \qquad E_m = \frac{1}{M} \sum_{t=1}^{M} \mathbb{1}[w_t \in L_m] \]

💬 Read it in plain English: VADER sums rule-adjusted word valences into S and normalizes it onto [−1, +1]; the Ambivalence score separates genuinely neutral text from mixed reviews where strong praise and strong complaints cancel out; and NRC/VAD normalize matched token scores across the M recognized words in the document.

S, α

Rule-Adjusted Valence Sum & Normalization Constant OUTPUT ⓘ

Total summed sentiment valence (S) after applying boosters, negations, caps, and punctuation rules, scaled by constant α = 15.

compound, pos, neg

Normalized Compound Score & Tone Proportions INFERENCE ⓘ

Overall polarity score in [−1, +1] and proportions of positive, negative, and neutral token weight.

ambivalence

Mixed-Sentiment Ambivalence Index OUTPUT ⓘ

Distinguishes a flat, emotionless review (0) from a conflicted review containing equal parts praise and criticism (1).

E_m, M, L_m

Normalized Emotion Share & Matched Tokens OUTPUT ⓘ

Proportion of the M matched lexicon tokens in a document that evoke emotion or affect dimension m.

DATA SOURCE

Load a marketing use case:

📊 Pre-loaded Text Datasets

Use these case studies to practice reading text analysis outputs before uploading your own data. Each scenario comes with a suggested analysis method—but you're always free to run any or all three engines on any dataset.

Upload CSV/TSV file

Drag & Drop CSV file (.csv, .tsv, .txt, .xls, .xlsx)

The first row should be column headers. One column should contain your text data.

CONFIGURE YOUR DATA

Text column

Choose which column contains the text you want to analyze.

Enable grouping by column

Optional: compare results across groups (e.g., brand, source, product rating).

Row identifier column (optional)

Selects a column to use as row labels in output tables and inspectors.

SELECT ANALYSIS METHOD

Choose how you want to analyze your text. You can switch methods at any time without re-uploading data.

📊

Sentiment Analysis

VADER · Overall positive / negative tone

Score each record on a compound scale from −1 (strongly negative) to +1 (strongly positive). Reveals overall valence, label distribution, ambivalence, and group differences.

Best for: customer satisfaction, brand perception, comparing positive/negative signal across groups.

🎭

Emotion Profiling

NRC EmoLex · Eight discrete emotions

Map each record onto Plutchik's eight emotions (anger, fear, joy, sadness, disgust, surprise, anticipation, trust). Reveals which emotional registers dominate your text data.

Best for: understanding which emotions drive engagement, distinguishing fear from anger, identifying trust-building language.

🧭

Dimensional Affect

NRC-VAD · Valence, Arousal, Dominance

Place records in continuous emotional space. Valence = pleasant vs. unpleasant. Arousal = activated vs. calm. Dominance = in-control vs. controlled. Visualized on Russell's circumplex.

Best for: segmenting arousal level (rage vs. disappointment), analyzing power framing in text, circumplex mapping.

🔬

Full Comparison

All Three Methods · Side-by-side

Run VADER, NRC, and VAD on the same data and compare what each method reveals. Convergence across methods strengthens conclusions; divergence points to complexity worth investigating.

Best for: publishable analyses, presentations where you need to defend your methodology, and rich qualitative-to-quantitative interpretation.

SENTIMENT RESULTS

Reading VADER output

VADER returns a compound score (−1 to +1) for each record: values ≥ 0.05 are labeled positive, values ≤ −0.05 are labeled negative, and everything in between is labeled neutral. The compound score is the primary output; the separate pos, neu, and neg proportions show how much of the text's vocabulary contributes each tone.

A compound score near zero deserves scrutiny. The ambivalence score below tells you whether neutral-looking records are genuinely low in sentiment ("the product arrived") or highly mixed ("outstanding service, but terrible product quality"). Mixed records may warrant qualitative follow-up even if their average looks benign.

Sentiment summary

Average compound score: –

Records labeled positive: –

Records labeled neutral: –

Records labeled negative: –

Run the analysis to see overall sentiment across your text records.

Labels distribution

Bar chart shows how records are classified by VADER's compound score thresholds. Mixed-neutral records (high ambivalence near zero) are distinguished from genuinely neutral ones.

DETAIL & WORKED EXAMPLE

Export Per-Record Results

Download all VADER scores (compound, pos, neu, neg, ambivalence, label) as a CSV—one row per record.

AI-Aided Semantic Coding & Benchmark

Claude Haiku 5.5 vs. VADER

Compare VADER's rule/lexicon heuristics with Claude Haiku 5.5's deep contextual understanding. Discover where algorithms agree, where they diverge on sarcasm or idioms, and how Aspect-Based Sentiment Analysis (ABSA) reveals nuanced marketing dimensions.

ℹ️ How this AI Comparative Benchmark Works (Methodology)

1. Full Dataset Screening: 100% of your uploaded dataset is analyzed locally in your browser using classical rule/lexicon heuristics (VADER).

2. Stratified Benchmarking: To provide rigorous validation without browser latency or token waste, an educational stratified sample (highest positive, lowest negative, and borderline ambivalent cases) is benchmarked with Claude Haiku 5.5.

3. Diagnostic Reliability: We compute Pearson correlation (r) and agreement rate across the sampled spectrum, isolate severe divergence (|Δ| ≥ 0.40) to diagnose sarcasm and pragmatic nuance, and extract Aspect-Based Sentiment (ABSA).

Note: Datasets between 5 and 5,000 records are supported for AI comparative benchmarking.

VADER Avg Compound

Claude Avg Valence

Label Agreement

Pos/Neu/Neg match %

Correlation (r)

Pearson correlation

Sarcasm Flagged

Implicit / Ironic

🤖 Claude Haiku Executive Takeaway:

⚖️ Head-to-Head Model Audit: Lexicon Heuristic (VADER) vs. Contextual LLM (Claude Haiku)

Both methods score the same sampled records. Claude receives up to 400 characters per record; VADER scores the full text. Select a matrix cell to inspect its comments. Counts and percentages describe this paired sample.

📊 Macro Sentiment Distributions Polarity Breakdown

VADER (Lexicon Rules) --

Claude Haiku (LLM) --

● Positive ● Neutral ● Negative

🔲 Cross-Classification Matrix Click cell to filter table

VADER ↓ \ AI →	Claude Haiku (Contextual AI)			VADER Total
VADER ↓ \ AI →	Positive	Neutral	Negative	VADER Total

Concordance (Agreement) Mild Divergence Polarity Reversal

⚠️ Disagreement Diagnostic: Where VADER & Claude Diverge (|Δ| ≥ 0.40)

These records represent critical teaching moments. They highlight where dictionary word-matching fails on sarcasm, double negatives, or polite complaints, whereas contextual AI captures true pragmatic intent.

Record	Text Excerpt	VADER	Claude AI	Δ Diff	AI Rationale & Nuance Diagnostic

🎯 Aspect-Based Sentiment Analysis (ABSA)

Instead of an uninformative single blanket score, customer feedback divides into distinct marketing touchpoints. Here is how Claude decomposed sentiment across key operational aspects:

View All Records Side-by-Side (VADER vs. Claude AI)

#	Text	VADER	Claude	Aspect	Sarcasm	Rationale

How VADER scored two examples

The worked example below shows the token-by-token valence coding for one relatively positive record and one relatively negative record. Color-coded chips show each token's base lexicon value; hover a chip to see the full modifier chain (negation, booster, capitalization) that produced the final contribution.

Run the analysis to see how each token in a relatively positive record contributes to its final sentiment score.

Run the analysis to see how each token in a relatively negative record contributes to its final sentiment score.

How this example is computed

Each token is looked up in the VADER lexicon. Tokens with a known valence receive a colored chip; words not in the lexicon appear as plain gray chips. The six rule layers (see Concepts above) then adjust those base scores before they are summed and normalized to the compound score shown.

Tip: Hover over any colored chip in the Record Inspector to see the full modifier chain—what the base lexicon value was, which rule modified it, and what the final contribution was.

🔗 From examples to patterns: These two examples show individual token logic. To see how sentiment is distributed across all your records, scroll to the Histogram and Box Plot below.

Record Inspector — Token-Level Valence Coding

Select any record to see how each token's valence contribution builds the final compound score. Colored chips indicate tokens with non-zero lexicon values; hover for the full modifier chain. Gray chips are lexicon misses.

Select a record:

Token-level breakdown will appear here after running the analysis.

ANALYSIS REPORT

APA-Style Statistical Reporting

Run the analysis to generate an APA-style report of your sentiment results.

Managerial Interpretation

Run the analysis to generate a managerial interpretation of your sentiment results.

Summary of Estimates

Measure	Estimate	Std. Dev.	Min	Max

Sentiment Distribution: Histogram

How to read bar colors: Green = positive territory (≥ 0.05), red = negative (≤ −0.05), gray = neutral. Bar height = record count in that range.

Run the analysis to see how compound sentiment scores are distributed across your text records.

EMOTION PROFILING RESULTS

Reading NRC output

Each emotion score represents the proportion of lexicon-matched tokens in a record that carry that emotion. A record scoring 0.40 on "joy" has 40% of its emotionally-loaded words tagged as joy-associated in the NRC lexicon. Scores do not sum to 1 because a single word can carry multiple emotions.

Lexicon coverage matters: a record with only two matched tokens may have an extreme emotion profile just by chance. The inspector below shows which tokens actually matched, so you can judge whether the scores are driven by meaningful signal or incidental vocabulary.

The radar chart makes cross-group comparison intuitive: a brand whose texts form a large "joy + trust" profile occupies a very different space from one whose texts form a "fear + anger" shape—even if their overall VADER sentiment scores are similar.

Run the NRC Emotion Lexicon analysis on your current text data.

Dominant Emotion

–

Avg positive (binary): –

Avg negative (binary): –

Avg lexicon coverage: –

Emotion Profile (Radar)

Each axis = proportion of matched tokens carrying that emotion. Overlapping traces = group comparison.

Emotion Scores (Bar Chart)

Bars sorted by average normalized score. When grouping is active, shows grouped bars per emotion.

Emotion Profile by Group

Per-group breakdown of dominant emotion, top emotions, and lexicon coverage. Radar and bar chart above also show full group traces.

Group	n	Dominant Emotion	Top 3 Emotions	Avg Coverage

Emotion Deviation from Overall Average

+Green = group scores above overall average; −Red = below. Values in percentage points (pp).

Emotion Summary Table

Emotion	Avg (% of matched tokens)	Std. Dev.

NRC Interpretation

Run the analysis to generate an interpretation.

Record Inspector — Token-Level Emotion Coding

Pick any record to see which tokens matched the NRC lexicon and which emotions each one carries.

Select a record:

Choose a record above to inspect token-level coding.

DIMENSIONAL AFFECT RESULTS

Reading VAD output

Each record receives three scores in [0, 1]: Valence (how pleasant the language is), Arousal (how activated or calm), and Dominance (how in-control vs. controlled the tone is). Scores near 0.5 are emotionally neutral on that dimension; scores far from 0.5 indicate strong displacement in one direction.

The Valence × Arousal scatter plot places each record in Russell's circumplex. The four quadrants reveal qualitatively different emotional states invisible to sentiment scoring— excited (high-V, high-A), anxious (low-V, high-A), relaxed (high-V, low-A), and sad / bored (low-V, low-A).

Dominance in marketing context: Customer complaints often show low valence and low dominance ("I felt helpless"). Recommendation texts often combine high valence with high dominance ("I confidently recommend this to everyone"). These V–D combinations tell a story that valence alone cannot.

Run the NRC-VAD analysis to score your text on Valence, Arousal, and Dominance.

FULL COMPARISON — ALL THREE METHODS

How Mode 4 works

All three engines run on the same dataset. Use the comparison table to see where methods converge (strong evidence) and where they diverge (interesting complexity worth investigating). The combined CSV download exports all scores for every record in a single flat file.

📊 VADER Pending

🎭 NRC Pending

🧭 VAD Pending

CROSS-METHOD COMPARISON

Methods Compared

Method	Key Result 1	Key Result 2	What it captures	What it misses

Worked Example: One Record, Three Methods

The record with the most extreme VADER compound score is used as the demonstration case. Compare how the three methods characterize the same text differently.

Cross-Method Insights

Where methods agree, conclusions are more robust. Where they diverge, the gap is itself informative. The analysis below flags the most notable convergences and divergences in your data.

Combined Export

A single CSV with VADER compound + pos/neg/neu + ambivalence, all 8 NRC emotion scores, and V/A/D values for every record—all methods side-by-side.

👨‍🏫 Professor Mode: Guided Learning Experience

OVERVIEW & WHAT THIS TOOL SOLVES

🎯 Questions It Answers

📂 What Data You Need

💡 What You'll Discover

Sentiment Analysis (VADER)

Emotion Profiling (NRC EmoLex)

Dimensional Affect (NRC-VAD)

DATA SOURCE

Use a Case Study

📊 Pre-loaded Text Datasets

Upload Your Data

CONFIGURE YOUR DATA

SELECT ANALYSIS METHOD

SENTIMENT RESULTS

Reading VADER output

Sentiment summary

Labels distribution

DETAIL & WORKED EXAMPLE

Export Per-Record Results

AI-Aided Semantic Coding & Benchmark

⚖️ Head-to-Head Model Audit: Lexicon Heuristic (VADER) vs. Contextual LLM (Claude Haiku)

⚠️ Disagreement Diagnostic: Where VADER & Claude Diverge (|Δ| ≥ 0.40)

🎯 Aspect-Based Sentiment Analysis (ABSA)

How VADER scored two examples

Record Inspector — Token-Level Valence Coding

Sentiment Character Analysis

ANALYSIS REPORT

APA-Style Statistical Reporting

Managerial Interpretation

Grouped Sentiment Comparison

Summary of Estimates

Sentiment Distribution: Histogram

Sentiment Distribution: Box Plot (By Group)

EMOTION PROFILING RESULTS

Reading NRC output

Dominant Emotion

Emotion Profile (Radar)

Emotion Scores (Bar Chart)

Emotion Profile by Group

Emotion Deviation from Overall Average

Emotion Summary Table

NRC Interpretation

Record Inspector — Token-Level Emotion Coding

DIMENSIONAL AFFECT RESULTS

Reading VAD output

Dimensional Scores

Valence × Arousal Scatter (Russell's Circumplex)

VAD Profile by Group

VAD Interpretation

Record Inspector — Token-Level VAD Coding

FULL COMPARISON — ALL THREE METHODS

How Mode 4 works

CROSS-METHOD COMPARISON

Methods Compared

Worked Example: One Record, Three Methods

Cross-Method Insights

Group Comparison Across Methods

Combined Export