Calculator Inputs
Enter scores, configure classification rules, and choose an analytical plot.
Formula Used
The calculator combines weighted classification, distribution, and uncertainty formulas.
Weighted class share
P(c) = Σ wᵢ I(cᵢ = c) ÷ Σ wᵢ
Weighted mean
μw = Σ(wᵢsᵢ) ÷ Σwᵢ
Weighted variance
σ²w = Σ[wᵢ(sᵢ − μw)²] ÷ Σwᵢ
Net sentiment
N = (Wpositive − Wnegative) ÷ Wtotal
Normalized entropy
Hn = −Σ P(c) ln P(c) ÷ ln(3)
Density estimate
f(x) = ΣwᵢK((x − sᵢ) ÷ h) ÷ (hΣwᵢ)
How to Use This Calculator
- Paste one sentiment score on every line.
- Add weights after commas when records have unequal importance.
- Set score limits matching your model output range.
- Choose negative and positive thresholds for class assignment.
- Select bins, smoothing bandwidth, scaling, and plot style.
- Press the calculation button and review results above.
- Use the plot toolbar to download a report image.
Understanding Sentiment Distribution Plots
Why Distribution Shape Matters
Sentiment distribution plots summarize how model scores spread across emotional classes. They reveal balance, uncertainty, skewness, and unusual scoring patterns within predictions. Clear distributions help teams diagnose datasets before important deployment decisions.
Each observation receives a score representing negative or positive orientation. Thresholds divide scores into negative, neutral, and positive reporting groups. Flexible boundaries support different domains, languages, products, and labeling policies.
Weights, Bins, and Density
Weighted analysis gives important records greater influence during statistical aggregation. This approach helps when examples represent unequal traffic or business value. Unweighted results remain useful for ordinary balanced evaluation datasets.
Histograms display score frequency across evenly spaced numerical intervals. More bins expose detail, while fewer bins emphasize broad structure. Excessive bins may create noise and misleading local variations.
Density curves estimate a smooth shape from observed sentiment scores. Bandwidth controls smoothing strength and visible local detail. Small bandwidths fluctuate, while large bandwidths may hide meaningful clusters.
Reading Summary Metrics
Class percentages provide a direct summary for operational reporting. Entropy measures how evenly sentiment mass spans all available classes. Lower entropy indicates dominance, while higher entropy suggests greater balance.
The weighted mean describes overall polarity across all submitted observations. Standard deviation measures dispersion around that central sentiment level. Skewness highlights asymmetric tails caused by unusual score concentrations.
Net sentiment compares positive weight against negative weight directly. Values near one indicate strong positive dominance throughout evaluated samples. Values near minus one indicate strong negative dominance instead.
Confidence intervals communicate uncertainty around the estimated weighted mean. Larger effective samples usually produce narrower and steadier intervals. Highly unequal weights can reduce effective sample size substantially.
Alternative Plot Views
Box plots reveal quartiles, medians, ranges, and possible outliers clearly. Violin plots add density shape around those summary positions. These views help compare compactness without hiding extreme observations.
Cumulative curves show the proportion falling below each score. Their slopes reveal crowded regions and sparse score intervals. Steep sections indicate many observations sharing similar sentiment values.
Validation and Interpretation
Use plots together with summary statistics and original text examples. Investigate extreme scores, threshold sensitivity, and unexpected class changes. Reliable interpretation requires context, representative data, and careful validation.
Threshold testing shows whether classifications remain stable near boundaries. Small boundary changes can expose fragile labels and uncertain predictions. Stable outputs improve confidence during monitoring and stakeholder communication.
Data quality still determines whether any visualization supports sound conclusions. Duplicate, translated, or automated records may distort the visible pattern. Review source coverage before trusting distribution changes over time.
Frequently Asked Questions
What score range should I use?
Use the exact range produced by your sentiment model. Common ranges include minus one through one, or zero through one. The calculator rejects values outside your selected limits.
How are neutral scores classified?
Scores at or above the negative threshold remain neutral until reaching the positive threshold. This interval lets you represent uncertain, mixed, or weakly polarized predictions.
When should weights be enabled?
Enable weights when records represent different traffic volumes, customer values, or sampling probabilities. Every weight must be positive. Unweighted analysis assigns one unit to every valid score.
What does net sentiment mean?
Net sentiment subtracts negative weight from positive weight, then divides by total weight. Results approach one for positive dominance. They approach minus one for negative dominance.
How should bandwidth be selected?
Start with a moderate value relative to your score range. Lower values show more local detail. Higher values produce smoother curves but may hide separate clusters.
Why does effective sample size decrease?
Highly unequal weights concentrate influence among fewer observations. The effective sample size reflects that lost information. More balanced weights usually produce stronger uncertainty estimates.
What does normalized entropy show?
Normalized entropy measures class balance from zero through one. Zero means one class dominates completely. Values near one indicate similar weighted shares across all classes.
Can this calculator detect model bias?
It can reveal imbalance, shifts, and unusual score patterns. Bias conclusions require subgroup analysis, representative labels, fairness metrics, and domain review. Treat this plot as a diagnostic starting point.
Can I save the generated plot?
Use the camera button inside the interactive toolbar. Exported browser images support reports, reviews, and model documentation.