Contents
Attitude: the missing link to intent
Why polarity alone falls short
The rise, and the limits, of emotion classification
A different signal: attitude
Attitude analysis in practice
Where attitude analysis adds value
FAQ: Attitude analysis
Author bio: about Jasper Snyder, Chief Client Officer at Converseon
Attitude: the missing link to intent
Sentiment analysis provides an assessment of the tone of a piece of content and, measured over time, that can give you an idea of how that tone is changing. But conventional sentiment analysis only provides a very simplistic picture, categorizing content into one of three categories: positive, negative, or neutral.
A decade ago, the industry's response to that limitation was emotion classification: frameworks such as Plutchik's Wheel of Emotion, which sort text into categories including joy, anger, fear, and trust. The approach represented a genuine advance, but it was adapted from psychology research built around facial expression and physiological response, not written language - a mismatch that produces ambiguous category boundaries, forced single-label tagging, and results that skew heavily toward one or two catch-all emotions.
Attitude analysis asks a different question. Rather than determining what emotion a statement communicates, it asks what stance the author is taking, and - critically - what that stance typically predicts. Two negative comments, for example one frustrated, and one disparaging, call for totally different responses: frustration generally reflects a specific, fixable lapse, while disparagement may in fact signal an eroding relationship with the brand. Attitude labels, on the other hand, correlate more directly with what an author is likely to do next: purchase, churn, advocate, disengage - the actions that every marketer cares about.
Sentiment and emotion can establish how an audience feels. Attitude establishes why they feel that way, and what follows from it.
Why polarity alone falls short
Sentiment classification assigns an entire message a single label, typically positive, negative, or neutral. That can be enough for a directional trend line, but often ignores the elements that would inform how a brand should respond.
The cost shows up quickly at scale. A team triaging by polarity treats a mildly annoyed comment identically to an urgent complaint, since both register as "negative." Response templates default to generic language, absent a more precise signal on what actually went wrong. In parallel, early indicators of a developing issue - e.g., a shift in tone that precedes any movement in volume or aggregate sentiment - go undetected until the trend is already visible in the topline numbers, by which point the advantage of early detection has already been lost.
The rise, and the limits, of emotion classification
Emotion analysis represented the industry's first serious attempt to add texture, or color, to sentiment. Its adoption, and its eventual ceiling, both trace back to the same origin.
Plutchik's model, which organizes emotion into eight core types arranged in opposing pairs, gained traction in social and media analysis beginning around 2010, reinforced by broader interest in Daniel Kahneman's distinction between fast, intuitive System 1 judgment and slower, deliberate System 2 reasoning. I recall developing a text classification framework based on the wheel; it’s relatively intuitive, neat, and symmetrical, and social posts and reviews appeared to be a natural fit. Plutchik-derived models became the industry default for the decade that followed.
The underlying challenge, however, is structural rather than incidental. Plutchik's Wheel - like most emotion taxonomies - was designed to describe facial expression and physiological state, not written language, and several practical problems follow directly from that origin. First, categories like anger versus disgust, or joy versus trust, are abstract enough that human annotators frequently disagree, which affects both consistency and model accuracy. Similarly, sarcasm and mixed tone rarely map cleanly onto a fixed emotional category.
Of course, forcing a single label per message misrepresents how people actually write, given that most posts mix and qualify sentiment within the same few sentences. An emotion label such as "anger" carries no indication of cause - e.g., pricing, service, or something completely unrelated to the brand - leaving no clear next step attached to the classification. Finally, in practice, results skew heavily toward one or two dominant categories, commonly joy and disgust, leaving you with a distribution often not meaningfully more informative or actionable than the polarity label it was meant to replace.
A different signal: attitude
Attitude analysis reframes the underlying question, from what a statement resembles emotionally to what stance it reflects and moreover what that stance predicts. The distinction changes what the resulting data is actually useful for.
Attitude correlates more closely with intent and downstream behavior than either emotional category or raw sentiment. A laudatory tone tracks with purchase likelihood; an outraged tone tracks with churn risk; a grateful tone signals advocacy potential; a frustrated tone predicts reduced future engagement even where the sentiment score itself reads only mildly negative.
Attitude also preserves layering that emotion classification tends to flatten. Consider a comment such as "thanks for finally fixing this." Sentiment registers it as positive. Emotion registers it as joy. Both readings miss that the underlying attitude is more accurately described as grateful and frustrated at once; the issue has been resolved, but the delay leaves the customer at continued risk of churning. That distinction, invisible to a single-label emotion model, is exactly the kind of signal that might trigger a follow-up rather than a closed ticket.
The nuance problems that undermine emotion models - sarcasm, contradictory statements within a single message, meaning that shifts according to what preceded it - remain genuinely difficult for attitude analysis as well; no framework resolves these outright. What attitude analysis does provide is more room to represent complexity, since its labels are not confined to a small, fixed set of emotional states, and more than one attitude can apply to a single message - disappointed and disparaging, for example - in a way a forced single-emotion choice cannot accommodate.
Attitude analysis in practice
Once your analysis or triage follows intent, rather than tone alone, the whole engine runs more effectively. A frustrated comment and a disparaging one may carry identical polarity, but they don’t belong in the same queue: the former calls for faster support, the latter for a more difficult conversation about trust. Routing by attitude directs resources according to what a given case actually requires, rather than treating all negative feedback as uniformly urgent.
Similarly, an inquiring comment requires information; a confused one requires clarification; an aggrieved one requires acknowledgment before anything else follows. Tying response playbooks to attitude, rather than to a coarser emotional category, produces a more direct line from signal to action.
Perhaps most consequentially, attitude tends to move before sentiment does. Aggregate sentiment can remain flat for weeks while attitude signals - for example a rise in cautionary or disparaging language - are already shifting beneath it.
Consider a brand that’s tracking consumer interest or reaction to one of its products over an extended period. Sentiment alone shows negative mentions ticking upward modestly, but neutral mentions are rising in parallel, and net sentiment barely moves; nothing in the topline figures appears alarming. Attitude analysis applied to the same underlying data tells a materially different story: a cluster of inquiring and cautionary mentions grows well ahead of any sentiment-driven alert, alongside a smaller but rising share of disparaging posts. Read together, the pattern typically indicates one thing - audiences are encountering conflicting or incomplete information about the product, a portion are asking questions to close the gap, and a smaller group is already repeating and amplifying the confusion. That is a totally different signal than "sentiment is holding steady," and it is one a team can act on before a narrative sets in: correcting the record, addressing the misinformation directly, and re-engaging the inquiring segment before it migrates toward disparagement.
Where attitude analysis adds value
Attitude contributes most wherever an organization needs to understand not just how an audience feels, but what action that feeling is likely to turn into. Let’s now look at how attitude analysis can support a wide range of different use cases.
- In brand perception and positioning, attitude distinguishes a laudatory mention from a celebratory one, but also identifies feedback that reads as positive on its surface while carrying a frustrated or skeptical undertone beneath it - an early indicator of reputational fragility a simple sentiment score wouldn’t surface.
- In crisis detection, a rise in frustrated or outraged attitudes frequently precedes broader backlash, allowing teams to prioritize genuinely urgent criticism over louder but less consequential complaints, and to register a shift in tone even before it appears as a spike in volume.
- From a campaign perspective, attitude clusters such as ‘desiring’ or ‘anticipatory’ reveal which marketing and communications an audience actually responds to, a more direct measure of message fit than sentiment direction alone.
- In churn and loyalty prediction, a frustrated attitude paired with continued engagement reads very differently from a disparaging one - the former might still be recoverable, the latter typically indicates a customer who has already disengaged.
- In competitive intelligence, comparing the attitude composition of conversation around a brand against its competitors - trusting and enthusiastic versus curious and wary, for example - surfaces positioning gaps that a side-by-side sentiment comparison simply wouldn’t be able to show.
- Finally, in customer experience analysis, distinguishing a frustrated-but-grateful customer from an outraged one changes the appropriate follow-up considerably, a distinction that neither a CSAT score nor a sentiment label alone will reveal.
FAQ: Attitude analysis
What is attitude analysis in text and sentiment analytics?
Attitude analysis classifies text according to the author's stance or intent - frustrated, inquiring, grateful, disparaging, among others - rather than by emotional category or simple polarity, with the aim of capturing what a statement predicts rather than merely how it sounds.
How does attitude analysis differ from emotion analysis?
Emotion analysis assigns a psychological state, such as joy, anger, or fear. Attitude analysis assigns a stance tied to intent and likely behavior, and permits more than one label per message where emotion analysis typically forces a single choice.
Why did emotion models such as Plutchik's Wheel become the industry standard, and where do their limitations lie?
Plutchik's model gained adoption in the early 2010s as social analytics matured beyond polarity alone. Its limitations stem from its origin in facial expression and physiological research rather than written text, which produces ambiguous categories and forced single labels.
Can attitude analysis handle sarcasm or mixed sentiment within a single message?
Attitude analysis can handle mixed sentiment by assigning multiple attitudes to the same message. For example, “Thanks for finally fixing this” may express gratitude about the resolution and frustration about the delay. Sarcasm is harder because the intended meaning often conflicts with the literal wording. Context-aware models improve accuracy by analyzing surrounding language, punctuation, and conversation history, but short comments, cultural references, and private jokes may still require human review.
What should organizations look for in an attitude-based approach?
Organizations should look for an attitude-analysis approach that understands informal language, mixed meanings, and conversational context. It should identify multiple attitudes within one message and explain why each label was assigned. Its categories should also correlate with meaningful outcomes such as purchase intent, advocacy, disengagement, or churn. Before adoption, organizations should test the model using their own content, audiences, languages, and channels. Customizable categories and human review are especially important for specialized terminology and high-stakes decisions.
Author bio: about Jasper Snyder, Chief Client Officer at Converseon
Jasper Snyder is Chief Client Officer at Converseon, where he works with brands on turning unstructured social data into actionable insight. Before joining Converseon in 2026 he led research at the Advertising Research Foundation - including its landmark How Advertising Works study.
