A sentiment score is a numeric or categorical output produced by a stated sentiment-analysis method for a defined unit of text. It can represent polarity, a predicted class, confidence in that class, or a custom aggregate, so its scale and calculation must be documented before interpretation.
There is no universal sentiment scale. A positive value in one system may not be comparable with a positive value in another.
What a sentiment score can represent
“Sentiment score” is used for several different outputs:
| Output type | Example | What the number or label represents |
|---|---|---|
| Categorical label | Positive | The selected sentiment class |
| Polarity score | 0.7 on a -1 to 1 scale | Direction and strength under one model’s scale |
| Class confidence | Positive: 0.82 | Model confidence or likelihood assigned to a class |
| Aspect score | Delivery: -0.6 | Sentiment attached to a specific target or aspect |
| Aggregate index | Weekly score: 64 out of 100 | A custom formula across multiple records |
Two systems can analyze the same text and return different output structures without either being internally inconsistent. The essential question is not “What is the sentiment score?” but “What exactly does this score encode?”
Common sentiment score scales
Negative to positive polarity
Some systems place sentiment on a continuous scale. Google Cloud Natural Language, for example, defines its current sentiment score from -1.0 for negative sentiment to 1.0 for positive sentiment. It separately returns non-negative magnitude, which represents the overall strength of sentiment regardless of direction. See Google’s official Sentiment reference.
That range belongs to Google’s implementation. It should not be presented as the default for all tools.
Per-class confidence
Other systems return a score for each possible class. Amazon Comprehend returns positive, negative, neutral, and mixed class scores, then selects the sentiment with the highest confidence. AWS describes those scores as the likelihood that each sentiment was correctly detected. See the official Amazon Comprehend sentiment documentation.
A confidence of 0.90 for positive does not mean the text is “90% positive.” It means the model assigned high confidence to the positive class under that system.
Custom normalized indexes
Dashboards may transform model outputs into 0-to-100 scales, net scores, weighted averages, or proprietary indexes. A normalized number is useful only when the documentation states:
- Source records and eligibility rules
- Unit of analysis
- Underlying model output
- Mapping or normalization formula
- Weighting and treatment of neutral or mixed records
- Handling of missing, duplicate, or low-confidence data
Do not reverse-engineer business meaning from a chart’s axis label.
Sentiment score calculation example
Assume one documented model returns polarity scores on a -1 to 1 scale for five eligible reviews:
| Review | Polarity score |
|---|---|
| 1 | 0.8 |
| 2 | 0.4 |
| 3 | 0.0 |
| 4 | -0.3 |
| 5 | 0.6 |
A simple unweighted average is:
(0.8 + 0.4 + 0.0 – 0.3 + 0.6) / 5 = 0.3
The result means the average model output for this exact five-record set was 0.3 on that model’s scale. It does not prove that customers were 30% positive, that 30% of reviews were positive, or that a different model would return the same result.
Only average scores that have the same scale, model version, unit, and preprocessing. If review 1 represents one document while another value represents an aggregate topic, the inputs are not comparable.
Sentiment score versus confidence score
This is the most important distinction in sentiment reporting.
- Polarity score: Estimates the direction or degree of sentiment on a defined scale.
- Confidence score: Estimates how strongly the model supports a predicted label.
A sentence can receive a mildly positive polarity with high classification confidence. Another can receive strongly positive polarity with low confidence because the language is ambiguous or unfamiliar. Plotting both values on the same axis creates a misleading comparison.
Also separate score from magnitude, volume, and certainty. Google’s magnitude is not a confidence value, and mention count is not sentiment intensity.
How to interpret a sentiment score
Before using a score, record:
- Source: Which reviews, posts, comments, tickets, or other records were eligible?
- Target: Is the sentiment about the brand, one product aspect, or the entire document?
- Scale: What are the minimum, maximum, neutral point, and class meanings?
- Method: Which model, prompt, lexicon, or rule set produced the score?
- Aggregation: How were multiple records weighted and combined?
- Version: Did the model, query, or source coverage change?
- Validation: How did the method perform on representative human-labeled examples?
Compare trends only when those conditions remain stable. If the score changes after a model update, rerun a fixed benchmark set to separate measurement drift from a real change in audience language.
Common sentiment score errors
- Treating a confidence value as emotional intensity
- Calling any value above zero “good” without checking the scale
- Averaging outputs from different models or units
- Hiding the effect of neutral, mixed, duplicate, or missing records
- Reporting an aggregate without the underlying mention count
- Comparing languages without language-specific validation
- Inferring revenue, loyalty, or churn from sentiment correlation alone
- Ignoring mixed opinions and aspect-level disagreement
Retain the source text behind every aggregate. Review a sample when a score moves materially, and explain whether the change came from volume, topic mix, model output, or data coverage.
Sentiment analysis is the parent method; a sentiment score is one possible output. BrandJet’s sentiment analysis feature page is the source for any current product-specific scoring method.
Frequently asked questions
What is a good sentiment score?
There is no universal good score. Interpret the value using the system’s scale, baseline, source set, and decision threshold. Compare like-for-like periods and inspect the underlying records.
Is a sentiment score a percentage?
Not necessarily. It may be a polarity value, class confidence, categorical label, or custom index. A score should be described as a percentage only when its documented calculation produces a percentage.
Can sentiment scores from different tools be compared?
Usually not directly. Tools may use different sources, labels, models, units, scales, and aggregation rules. Compare them only after mapping the methods and validating a shared test set.
Should neutral records be included in an aggregate sentiment score?
That depends on the metric. Include or exclude them according to a documented formula, and report the rule. Changing neutral treatment can move the result even when the underlying text does not change.