Sentiment Analysis
What sentiment analysis is, how it works, and how to interpret the results.
Every public conversation carries a collective opinion, often distributed across thousands or millions of individual posts. Sentiment analysis makes that collective opinion measurable by classifying the emotional tone of text at scale, turning scattered reactions into a coherent picture of how people feel about a subject. Once measured, that opinion can be tracked over time and compared across entities.
What is sentiment analysis?
Sentiment analysis is a form of text analysis that identifies the emotional tone of written content. Given a piece of text, a sentiment model classifies it as positive, negative, or neutral based on the language used. The input can be anything written: social media posts, product reviews, news articles, survey responses, forum threads. The output is a classification, either a label (positive, negative, neutral) or a numerical score representing where the text falls on a spectrum from strongly negative to strongly positive.
When applied to thousands or millions of posts, individual classifications become an aggregate picture of how people feel about a subject. An investor might track sentiment around an asset to gauge collective conviction before making a decision. A researcher might follow sentiment around a policy proposal as public reaction takes shape. An analyst might compare sentiment across related subjects to see which are gaining and which are losing public confidence. A trader might weigh the conversation around a prediction market event against the price the market quotes.
The subject of analysis can be anything people discuss publicly: a company, a person, a product, a political issue, an industry trend, a cultural moment. Sentiment analysis does not require the subject to have a stock ticker or a brand marketing team behind it. If people are talking about it, their sentiment can be measured.
At its core, sentiment analysis answers a straightforward question: when people talk about this subject, are they expressing positive or negative feelings? At scale, individual classifications aggregate into a measure of public opinion across millions of posts, making it possible to track how people feel about a subject over time and compare perception across entities.
How does sentiment analysis work?
Understanding the mechanics behind sentiment analysis helps set expectations. Knowing what the system does and does not consider when classifying text makes it easier to interpret results and recognize edge cases.
Sentiment analysis processes text in two stages: preparation and classification. In the preparation stage, the system normalizes the input. Potential sources of noise such as URLs and formatting artifacts are removed, and platform-specific conventions such as hashtags are handled. The goal is to isolate the words and phrases that carry meaning.
In the classification stage, the system evaluates the prepared text and assigns a sentiment label. How it does so depends on the approach.
Rule-based systems use predefined word lists where each word carries an associated sentiment score. "Excellent" might score +3, "terrible" might score -3, and "okay" might score 0. The system tallies the scores across all words in the text and produces an overall classification. These systems are transparent and straightforward, but they struggle with context. The sentence "the plot was not terrible" contains a negative word used in a positive context, and a simple rule-based approach is likely to misinterpret it.
Trained classifiers learn patterns from labeled training data. Given a large number of text samples that humans have classified as positive, negative, or neutral, the model identifies which combinations of words, phrases, and structural patterns are associated with each category. These models handle context better than rule-based systems because they learn from examples rather than individual word scores.
Transformer models are the third approach. Where trained classifiers are built specifically for sentiment, transformer models learn how language works broadly before being applied to any specific task. They can recognize that "this phone is sick" is positive in casual speech and that "not bad at all" is a mild endorsement rather than a negative statement. They handle nuance, slang, and indirect expression more effectively than earlier approaches, and the difference shows up in application. In one peer-reviewed comparison, trading strategies built on sentiment scored by large language models outperformed the same strategies built on dictionary scores (Sentiment trading with large language models).
What are the types of sentiment analysis?
Sentiment analysis takes several forms depending on the granularity and focus of the classification.
| Type | What it classifies | Example output |
|---|---|---|
| Polarity | Overall direction of feeling | Positive, negative, or neutral |
| Fine-grained | Direction and intensity on a scale | Very positive through very negative |
| Aspect-based | Feeling toward each attribute of a subject | Positive on food, negative on service |
| Emotion classification | The specific feeling expressed | Joy, anger, sadness, or fear |
Polarity analysis
Polarity analysis is the most common form, classifying each piece of text as positive, negative, or neutral. A post like "I love this new feature" is classified as positive, while "This update broke everything" registers as negative and "The company announced quarterly results today" reads as neutral. This three-way classification captures the direction of sentiment and is sufficient for most use cases that involve tracking or comparing public opinion at scale. When someone refers to "sentiment analysis" without qualification, they usually mean polarity analysis.
Fine-grained sentiment analysis
Fine-grained sentiment analysis extends polarity to a scale, typically five levels: very positive, positive, neutral, negative, and very negative. This captures intensity alongside direction. "This is the best purchase I have ever made" would be classified as very positive, while "it is decent, nothing special" would register as mildly positive. Both are positive, but the strength of feeling is different, and fine-grained analysis preserves that distinction. It is common in contexts where the degree of sentiment matters, such as product review analysis or customer satisfaction scoring.
Aspect-based sentiment analysis
Rather than assigning a single sentiment to an entire text, aspect-based analysis identifies sentiment about specific features or attributes mentioned in the text. A restaurant review might be positive about the food, neutral about the ambiance, and negative about the service. Assigning a single overall sentiment to this review would mask the fact that the reviewer had distinctly different feelings about different aspects of the experience. Aspect-based analysis separates these dimensions and scores each independently.
This type of analysis is particularly useful for multi-faceted entities where overall sentiment obscures important variation. A smartphone might have overwhelmingly positive sentiment about its camera but consistently negative sentiment about battery life. The aggregate score would land somewhere in the middle, telling you very little. The aspect-level breakdown tells you exactly what people like and what they want improved.
Emotion classification
Emotion classification goes beyond positive and negative to identify specific emotions: joy, anger, sadness, fear, love, anticipation, disgust. Where polarity analysis tells you the direction of feeling, emotion classification tells you the quality of feeling. A post classified as negative could be driven by anger, by sadness, or by fear, and each carries different implications. This type of analysis is covered in depth in the Emotion Classification guide.
How are sentiment scores calculated?
Not all sentiment analysis systems produce scores the same way. Understanding the scoring approach clarifies what the numbers represent.
| Approach | Output | Common context |
|---|---|---|
| Categorical | A label: positive, negative, or neutral | Basic tools and quick reads |
| Fine-grained categorical | One of five labels, very negative to very positive | Review and satisfaction analysis |
| Continuous | A decimal score, typically -1.0 to +1.0 | Academic models and cloud APIs |
| Ratio-based | A score from -100 to +100 across a set of posts | Aggregated social media analytics |
Categorical scoring assigns a label: positive, negative, or neutral. This is the simplest output and the most common in basic tools. It captures the direction of sentiment effectively, though it cannot distinguish between a mildly positive review and an enthusiastic one.
Fine-grained categorical scoring extends polarity to a five-level scale: very positive, positive, neutral, negative, and very negative. The additional levels capture intensity alongside direction, making it possible to distinguish a mild preference from strong enthusiasm. This approach is common in product review analysis and customer satisfaction scoring.
Continuous scoring produces a decimal value on a scale, typically from -1.0 to +1.0, where 0 represents neutral. Rather than placing text into a fixed number of buckets, continuous scoring positions it along a spectrum, allowing finer distinctions. This approach is common in academic models and cloud APIs.
Ratio-based scoring calculates the balance between positive and negative posts across a set of results. The formula takes the count of positive posts, subtracts the count of negative posts, divides by the total of positive and negative combined, and multiplies by 100. The result is a score from -100 to +100.
An important detail in ratio-based scoring is that neutral posts are typically excluded from the calculation, so the score measures the balance of opinionated conversation rather than overall polarization. A subject where 90% of posts are neutral and the remaining 10% are positive will produce a high positive score, even though the vast majority of people had nothing emotional to say. The exclusion is deliberate. The score is designed to capture the balance of opinion among the people who expressed one, and understanding that distinction is essential for reading scores accurately.
The scoring approach determines what the numbers mean. A +0.4 on a continuous scale places a single post moderately toward the positive end, while a +40 on a ratio-based system means that among posts expressing an opinion, positive posts outnumber negative ones by a meaningful margin. Knowing which system produced a score is the first step in reading it correctly.
How do you interpret sentiment analysis results?
There are several factors that can affect the interpretation of results generated by sentiment analysis.
The distribution behind the score matters more than the score itself. Consider two scenarios that produce an identical score of +40. In the first, 90% of posts are neutral, 7% are positive, and 3% are negative. In the second, 20% of posts are neutral, 56% are positive, and 24% are negative. Both produce the same ratio when neutral posts are excluded, but they describe fundamentally different conversations. The first is a subject people rarely form opinions about, but when they do, those opinions lean positive. The second is a subject people feel strongly about, with positive posts outnumbering negative ones roughly two to one across a large and opinionated conversation. The score is identical in both cases, but the conversations behind them have nothing in common.
This is why it is important to look at the distribution alongside the headline score. The score gives you the direction, while the distribution gives you the character of the conversation: how opinionated people are, how divided they are, and how much of the conversation is neutral observation rather than emotional expression.
Different distribution shapes carry different analytical meaning. A conversation dominated by neutral posts with a small positive tail indicates mild, uncontroversial reception. A U-shaped distribution with clusters at both positive and negative ends indicates a polarized conversation, one where people hold strong opinions and disagree. A distribution that skews heavily negative with almost no positive posts is a different signal from one that is evenly split between negative and neutral. Each of these shapes tells you something the score alone cannot.
Volume gives sentiment its weight. A sentiment shift from +30 to -10 is a data point, and whether it is meaningful depends on how many posts it represents. A 40-point swing across 50 posts could reflect a handful of vocal critics in a thin conversation, while the same swing across 50,000 posts suggests a genuine shift in public perception.
Scores need volume context
A sentiment score without volume context is a fraction without a denominator. Read every score alongside the number of posts behind it, and weight scores from thin conversations accordingly.
The relationship works in the other direction too. If a product launch generates 30,000 posts with a score of +55, that represents a broadly held positive reception. If a niche library has 40 posts with a score of +55, it means a few people had good things to say. Both scores are +55, but the confidence each deserves is different.
Time range shapes the analysis. The same query analyzed over three days and over thirty days can produce different sentiment scores because the analysis window determines which conversation is included. A short window captures recent reactions, making the score sensitive to individual events. A longer window smooths daily fluctuations and reveals the broader trajectory. Neither is more correct, since the two windows answer different questions. A three-day range can tell you what people are expressing right now, whereas a thirty-day range can tell you how the conversation has evolved.
There is no universal baseline for a "normal" score. Sentiment levels vary structurally by subject type. Consumer brands often sit in mildly positive territory, in part because most posts come from customers voluntarily sharing experiences they chose to have. Political figures and polarizing topics tend to sit near zero or in negative territory because the conversation is inherently more divided. Sentiment around sports teams tends to swing with recent performance. Comparing scores across fundamentally different types of subjects requires understanding these structural differences. A score of +15 for a political figure might represent remarkably positive reception, while the same score for a consumer brand might indicate a problem.
Trend direction can matter more than the absolute score. A score of +20 that was +50 last week tells a different story from a score of +20 that was -10 last week. The first suggests deterioration; the second, recovery. Many experienced analysts watch the rate of change rather than the level. A stable score suggests a settled conversation, even if it sits in negative territory, while a score that is moving rapidly in either direction may warrant investigation.
What are the challenges of sentiment analysis?
No text analysis system is perfectly accurate. Understanding where sentiment analysis works well and where it struggles helps calibrate how much weight to place on the results.
Sarcasm and irony
Sarcastic text expresses the opposite of its literal meaning. "Great, another delayed flight" uses a positive word in a negative context. Rule-based systems will often misread this because they score words independently. Transformer models are better at detecting sarcasm than earlier approaches, but no system catches every instance.
At scale, occasional sarcastic misclassifications tend to wash out. In a dataset of 10,000 posts, a few dozen misread sarcastic comments are unlikely to move the aggregate score meaningfully. The risk increases for subjects where sarcasm is culturally prevalent. Certain types of discourse, particularly around technology products and public figures, tend to attract more sarcastic commentary than others. For these subjects, consider that the aggregate score may run slightly more positive than actual sentiment.
Negativity bias in social data
People who are dissatisfied tend to be more motivated to post about it than people who are satisfied. Negative experiences prompt public expression more readily than positive ones, so social media data can skew negative relative to actual opinion.
This bias is a property of the data source, and sentiment analysis reflects it accurately. The system captures what people say publicly, and satisfied customers who stay silent are absent from the dataset. A mildly negative score for a consumer product may say more about who chose to post than about how most customers feel.
Context and cultural variation
The same words carry different emotional weight in different communities. Different groups have their own norms for tone and directness, and language that is routine in one community may be misread by a model trained on data from another. Certain industries use jargon that carries sentiment in context but is neutral in isolation.
Regional and linguistic variation adds another layer. Models trained primarily on one regional dialect may misread expressions, idioms, or rhetorical conventions common in others, even when the text is written in the same language.
Ambiguity and mixed sentiment
Some posts genuinely contain both positive and negative signals. A review that praises the design of a product but criticizes its price is mixed rather than cleanly positive or negative. The model must either pick a dominant label or assign a neutral classification, neither of which fully captures the nuance.
At the individual post level this is a limitation, but at the aggregate level mixed-sentiment posts tend to distribute across categories and contribute to a more moderate overall score, which often reflects the genuine complexity of opinion around multi-faceted subjects.
What is sentiment analysis used for?
Sentiment analysis applies wherever public opinion matters and the volume of conversation exceeds what a person can read manually.
Comparative analysis
Analyzing sentiment for multiple entities side by side reveals how public perception compares across them: candidates in the same race, competing technologies, rival assets, or the possible outcomes of a prediction market event. Sentiment divergence, where one entity's score rises while another's falls, can signal a shift in relative standing before it appears in polls, prices, or adoption figures.
Baselines and early signals
Tracking sentiment over time establishes a baseline for how people talk about a subject. Deviations from that baseline are often the first visible sign that something has changed. A sudden negative shift might follow a controversial statement or a breaking story. A gradual positive trend might reflect improving reception or growing adoption. The pattern over time is often more revealing than the overall score.
The research record on sentiment as a leading market signal shows both the appeal of the idea and the care it requires. A widely cited study reported that mood measured from Twitter posts predicted daily moves of the Dow Jones Industrial Average with 86.7% directional accuracy (Bollen, Mao & Zeng, 2011), while a later replication study was unable to validate that headline claim and documented unexplained data exclusions in the original (Lachanski & Pav, 2017). Read together, the two papers support treating sentiment shifts as prompts for closer investigation rather than as standalone forecasts.
Research on public opinion
Sentiment data reveals how the public feels about policies, technologies, institutions, and cultural developments. Unlike surveys and polls, which capture stated preferences at a single point in time, sentiment analysis captures expressed opinions continuously. It can surface growing frustration, rising enthusiasm, and shifting attitudes as they happen rather than on a polling or research cycle.
What is the difference between sentiment and emotion analysis?
Sentiment analysis and emotion analysis are related techniques that answer different questions. Sentiment analysis classifies text by direction: positive, negative, or neutral. Emotion analysis classifies text by the specific feeling expressed: joy, anger, sadness, fear, love, anticipation, or disgust, among others. Sentiment tells you the direction people lean, and emotion tells you what is driving that lean.
The practical difference matters when a sentiment score alone does not tell the full story. Consider a subject with a strongly negative sentiment score. If the dominant emotion is anger, the conversation is likely driven by frustration, with people upset about a specific decision or event that already happened. If the dominant emotion is sadness, the conversation has a different character, marked by disappointment or a sense of loss. If the dominant emotion is fear, the conversation reflects uncertainty about what might happen next.
Each of these scenarios describes a different conversation. The same number can sit on top of an acute reaction to a specific event, a durable sense of disappointment, or an unresolved worry about what comes next. A sentiment score of -40 is identical in all three cases, but the interpretation differs significantly.
Many text analytics platforms offer both sentiment and emotion analysis as complementary views of the same data. Sentiment provides the quick read on direction, while emotion classification identifies the specific feelings behind it.
In practice, the two work together as a triage process. Sentiment provides the first read: is the conversation positive, negative, or mixed? If the answer is straightforward, with stable positive sentiment and no unusual patterns, sentiment alone may be sufficient for the task at hand. If the score is negative, shifting rapidly, or otherwise unexpected, emotion analysis offers the next level of understanding. Examining the emotion breakdown reveals whether negativity is driven by anger (often temporary and event-driven), sadness (often longer-lasting and tied to disappointment), or fear (often forward-looking and tied to uncertainty about what happens next). Starting with the broad sentiment read and drilling into emotions when the situation warrants it is more efficient than analyzing the full emotional breakdown every time.
Related guides
Sentiment analysis connects to several neighboring topics covered in their own guides.