Aspect based sentiment analysis blog header
Resources / Blog

Aspect-based sentiment analysis: understand what customers feel about each topic

Posted: Feb 03, 2025
Read time: 5 minutes
#Customer Experience #Artificial Intelligence #ROI & Analytics #Voice of Customer VoC
Bottom Curve

Customer feedback is rarely simple. A customer may love a product's functionality while being frustrated by its onboarding experience. They may praise customer service but criticize delivery times. They may even express positive and negative opinions in the same sentence.

Traditional sentiment analysis often summarizes feedback using a single positive, negative, or neutral score. Document-level sentiment provides a useful overview, while aspect-based sentiment analysis reveals the specific experiences influencing customer perceptions.

By evaluating sentiment at the topic level rather than the document level, businesses can gain more actionable customer insights and understand exactly which parts of an experience are creating satisfaction or frustration.

As organizations process growing volumes of customer feedback across reviews, surveys, chat conversations, and contact center interactions, AI is playing an increasingly important role in transforming unstructured data into actionable customer insights. Stanford HAI describes human-centered AI as technology designed to augment human capabilities and improve decision-making, reinforcing the importance of combining machine intelligence with human expertise when interpreting language and context. 

Within Journey Insights Hub, aspect-based sentiment analysis helps organizations move beyond overall sentiment scores to understand how customers feel about specific journey touchpoints, products, services, and experiences.

 

What is aspect-based sentiment analysis?

Aspect-based sentiment analysis is a method of evaluating sentiment toward specific topics, features, or aspects mentioned within customer feedback.
Instead of assigning one sentiment score to an entire review or conversation, the analysis identifies the distinct subjects being discussed and evaluates sentiment for each one separately.

Consider this customer review:

"The software is intuitive and easy to use, but implementation took much longer than expected."

A traditional sentiment model might classify this review as neutral because it contains both positive and negative language.

Aspect-based sentiment analysis produces a more detailed result:

  • Software usability: Positive
  • Implementation experience: Negative

By separating sentiment according to topic, organizations gain a far more accurate understanding of customer feedback sentiment.

This level of analysis is particularly valuable when working with open-ended customer feedback, where multiple experiences are often discussed simultaneously.

The need for deeper sentiment understanding is growing as organizations scale customer listening programs. According to Forrester's research on customer feedback management and analytics, one of the biggest challenges facing customer feedback programs is bringing together fragmented feedback from multiple channels to create a holistic understanding of customer experiences. AI-powered analysis helps organizations uncover patterns and relationships that would be difficult to identify manually.

 

Document-level, aspect-level and concept-level sentiment compared

Several approaches exist for sentiment analysis, each offering a different level of detail.

Document-level sentiment

Document-level sentiment assigns a single sentiment score to an entire document, review, survey response, or conversation.

Example:

“Great product but poor customer support.”

Output:

  • Overall sentiment: Neutral

While useful for tracking overall sentiment trends, this approach summarizes customer feedback at a broader level.

Aspect-level sentiment

Aspect-level sentiment identifies individual aspects and determines sentiment for each one.

Example:

“Great product but poor customer support.”

Output:

  • Product: Positive
  • Customer support: Negative

This provides significantly more actionable insight.

Concept-level sentiment analysis

Concept level sentiment analysis takes the analysis one step further.

Rather than focusing only on explicit topics, concept-level approaches attempt to understand the broader meaning behind related words and phrases.

For example:

  • Delivery
  • Shipping
  • Fulfillment
  • Arrival time

May all represent the same underlying concept.

Similarly:

  • Website navigation
  • Site usability
  • Finding information
  • Browsing experience

May reflect a common Customer Experience theme.

Concept-level sentiment analysis helps organizations uncover relationships between customer comments, particularly when different language is used to describe similar experiences.
 

How one review can contain positive and negative sentiment

Mixed sentiment presents an interesting challenge for many sentiment analysis approaches because customers often express multiple opinions within a single comment.

Customers rarely give entirely positive or entirely negative feedback.

The following sentiment analysis example shows how a single customer review can contain multiple opinions about different aspects of the experience.

Consider the following review:

“The website was easy to use and checkout was quick. Unfortunately, my order arrived late and support couldn't provide a clear update.”

Annotated mixed sentiment example

AspectSentiment
Website usabilityPositive
Checkout experiencePositive
DeliveryNegative
Customer supportNegative


Document-level result

  • Overall sentiment: Neutral

Aspect-based result

  • Website usability: Positive
  • Checkout experience: Positive
  • Delivery: Negative
  • Customer support: Negative

The second output provides significantly more value.

Instead of seeing a neutral review, the organization can identify exactly which experiences are creating positive and negative perceptions.

This ability to handle mixed sentiment makes aspect-based sentiment analysis particularly valuable for Customer Experience teams seeking to understand the root causes of customer satisfaction and dissatisfaction.

 

From aspect extraction to sentiment classification

Aspect-based sentiment analysis generally involves two stages.

Aspect extraction 

The first stage identifies the topics, entities, features, or concepts being discussed.

Examples include:

  • Product quality
  • Customer service
  • Pricing
  • Delivery
  • Onboarding
  • Mobile application performance

This process can be straightforward when customers explicitly mention a topic.

For example:

“Customer support resolved my issue quickly.”

However, aspect extraction becomes more challenging when different customers use different language.

Examples:

  • Customer support
  • Service desk
  • Help team
  • Technical support

All refer to the same underlying concept.

Modern systems therefore use contextual analysis and semantic understanding to group related expressions together.

Sentiment classification

Once aspects have been identified, sentiment is assigned to each one.

Typical classifications include:

  • Positive
  • Negative
  • Neutral

More advanced models may also assess:

  • Mixed sentiment
  • Sentiment intensity
  • Confidence scores
  • Emotional tone

The result is a detailed understanding of how customers feel about specific topics, experiences and concepts.

Effective sentiment analysis is not simply about automating classification. It is about understanding customer meaning within context. Forrester notes that the most reliable customer feedback analytics approaches combine AI and machine learning with contextual intelligence, business rules, taxonomies, and governance mechanisms to improve consistency and trust in analytical outputs, rather than relying on AI alone.
 

How sarcasm, cultural context and domain language affect sentiment

One of the most difficult challenges in sentiment analysis is context.

Intended versus literal meaning

Consider this statement:

“Great. Another update that broke everything.”

The word "great" is generally associated with positive sentiment.

The intended meaning, however, is clearly negative.

A model relying primarily on individual words may reach the wrong conclusion.

This challenge reflects a broader limitation in language understanding. Human communication relies heavily on context, shared knowledge, cultural norms, and implied meaning. Research and initiatives from Stanford HAI emphasize that AI systems are most effective when designed around human behavior and real-world communication patterns rather than purely literal interpretations of language.

The importance of surrounding text

Meaning is often influenced by nearby phrases and sentences.

For example:

“The new feature looked promising. After three days of troubleshooting, I disabled it completely.”

The overall sentiment is negative despite the initially positive language.

Cultural and regional language

Language changes across regions and communities.

Consider this statement:

“That feature is wicked.”

Depending on context, this may indicate criticism or enthusiastic praise.

The same challenge occurs with slang, informal language and emerging terminology.

Domain-specific language

Industries often use specialized language that differs from everyday speech.

For example:

“The treatment was aggressive.”

In healthcare, this may indicate an effective approach rather than a negative outcome.

Without domain knowledge, sentiment classification can become unreliable.

Common failure examples

Organizations frequently encounter sentiment errors involving:

  • Sarcasm
  • Irony
  • Slang
  • Technical jargon
  • Ambiguous references
  • Multi-topic comments

These challenges highlight why contextual sentiment analysis is substantially more difficult than identifying positive or negative words.

The complexity of contextual understanding explains why organizations should regularly validate outputs against real customer language. Sarcasm, irony, slang and industry-specific terminology continue to present challenges across even the most advanced language models.

Evaluating contextual understanding

To assess performance, organizations should test:

  • Aspect identification accuracy
  • Sentiment classification accuracy
  • False positives
  • False negatives
  • Mixed-sentiment handling
  • Context interpretation

Human review remains an important part of validating sentiment models and improving results over time.

 

How to test contextual errors across languages and channels

Organizations increasingly collect feedback from diverse channels and markets.

These include:

  • Surveys
  • Reviews
  • Contact center conversations
  • Online communities
  • Social media
  • Messaging platforms

Each channel introduces unique language patterns.

A customer writing a formal survey response may use very different language from a customer posting a product review or sharing an experience on social media.

Cross-channel validation

Organizations should evaluate sentiment performance across:

  • Different feedback channels
  • Different customer segments
  • Different product lines
  • Different journey stages

Cross-language testing

Global organizations should also validate sentiment across multiple languages.

Key considerations include:

  • Translation effects
  • Regional expressions
  • Cultural references
  • Local slang
  • Market-specific terminology

Benchmark testing

Strong sentiment programs regularly evaluate performance using examples that include:

  • Sarcasm
  • Mixed sentiment
  • Industry jargon
  • Multiple concepts
  • Ambiguous language

Organizations should also evaluate transparency and explainability alongside accuracy. MIT Sloan Management Review highlights the growing importance of explainable AI and effective human oversight, particularly when AI-generated insights influence operational or strategic decisions.

Trust increases when teams understand how analytical conclusions have been reached and can validate them against business knowledge.

The focus should be on delivering consistent, transparent and reliable interpretation that supports business decision-making across channels and markets.

 

How topic sentiment supports journey decisions

When combined with journey intelligence, aspect-based sentiment analysis helps organizations identify where experiences are succeeding, where friction occurs and where improvement efforts will have the greatest impact. By mapping topic-level sentiment to journey stages, organizations gain a clearer understanding of how specific interactions influence customer perceptions.

For example:

Journey StageTopicSentiment
AwarenessProduct informationPositive
ConsiderationFeature comparisonsNeutral
PurchaseCheckout processPositive
FulfillmentDelivery experienceNegative
SupportResponse timesNegative

This view immediately highlights which areas deserve attention.

Rather than simply reporting sentiment trends, organizations can identify the experiences influencing customer perceptions at each stage of the journey.

The value of aspect-based sentiment analysis ultimately lies in its ability to support better decisions. Research from MIT Sloan Management Review indicates that organizations derive the greatest value from AI when it improves decision-making, strategic alignment and business outcomes rather than operating as a standalone analytics capability.

Part of the Cemantica Journey Value Management Platform, Journey Insights Hub helps organizations combine aspect based sentiment analysis with journey intelligence, enabling teams to track topic level sentiment, monitor customer feedback sentiment over time and uncover the root causes of changing customer perceptions. 

By pairing AI-driven analysis with human expertise and business context, organizations can develop a more accurate and actionable understanding of what customers feel about each topic and use those insights to guide journey improvements.

Request a demonstration of the Cemantica platform.

Similar Resources