Use case · ABSA
Flat sentiment is not enough.
“Your overall sentiment is +0.3.”
That’s what most tools give you. A single number that says customers are “moderately positive.” It tells you nothing about what to fix, what to protect, or where to invest.
Aspect-Based Sentiment Analysis (ABSA) breaks that number apart. Instead of one score, you get sentiment for every aspect of your experience:
- Product Quality+0.70
- Customer Support−0.50
- Pricing+0.10
- Onboarding+0.60
- Shipping−0.40
Now you know exactly where to act.
Aspect sentiment heatmap
5 quarters
The taxonomy problem
Why most ABSA implementations fail
ABSA isn’t a new concept. The problem has always been the taxonomy — the list of aspects you score sentiment against.
The free-invention problem
Most AI tools let the model invent aspects on the fly. This quarter it identifies “Pricing.” Next quarter the same data produces “Cost.” The quarter after, “Value for Money.” The aspects shift, the numbers aren’t comparable, and the trend lines are meaningless.
The fixed-list problem
Some tools use a hardcoded list. “Pricing, Quality, Support, Shipping.” Rigid. If your business has unique aspects (e.g. “Battery Life” for hardware, “API Reliability” for developers), the fixed list misses them.
| Free invention | Fixed list | InsightNarrator | |
|---|---|---|---|
| Aspects | Model-generated, inconsistent | Hardcoded, static | Governed: AI drafts, humans approve |
| Consistency | Varies every run | Always the same | Always the same |
| Relevance | Adapts to data | May miss key aspects | Tailored to your business |
| Auditability | No control | Predictable | Full approval workflow |
The InsightNarrator solution: governed taxonomies
AI drafts the taxonomy
Upload a dataset. Scoty reviews the feedback and proposes a taxonomy based on what customers actually talk about — “Pricing, Customer Support, Onboarding, Shipping, Feature Requests, Documentation, Billing, Integration.”
Humans approve
Your CX team reviews the proposed taxonomy. Add aspects. Remove aspects. Rename them. This is now your governed taxonomy.
ABSA runs against the taxonomy
Every piece of feedback is scored against your approved aspects. “Pricing” always means “Pricing.” Wave after wave, quarter over quarter, the comparison is valid.
Taxonomy evolves
Need a new aspect for a product launch? Edit the taxonomy, re-run. Old data can be re-scored against the updated taxonomy.
How it works
The analysis pipeline
- Step 1
Text Extraction
Parse each response.
- Step 2
Aspect Detection
Identify which governed aspects are mentioned.
- Step 3
Sentiment Scoring
Score sentiment per detected aspect (−1.0 to +1.0).
- Step 4
Aggregation
Roll up to aspect-level averages, segment by metadata.
- Step 5
Confidence Scoring
Each score carries a confidence level based on signal strength.
Raw feedback in → aspect sentiment matrix out.
Example input
“The product itself is excellent and really well-built, but dealing with support has been a nightmare. Took three emails to get a response, and then they didn’t even solve my problem. On the plus side, pricing is fair for what you get.”
Example output
- Product Quality+0.85high confidence
- Customer Support−0.72high confidence
- Pricing+0.60medium confidence
One verbatim. Three aspect scores. Multiply by thousands of verbatims, and you have a complete experience diagnostic.
Worked example
An e-commerce company analysing 8,400 reviews
The setup
An online retailer with 50,000 monthly customers. Reviews collected from Trustpilot, Google, and post-purchase surveys. Uploaded to InsightNarrator as a single normalised dataset.
The taxonomy (governed)
- Product Quality
- Pricing
- Shipping
- Packaging
- Returns Process
- Website Experience
- Customer Support
- Product Range
- Stock Availability
- Loyalty Program
| Aspect | Sentiment | Trend | Volume |
|---|---|---|---|
| Product Quality | +0.72 | +0.08 | 3,200 mentions |
| Pricing | +0.45 | +0.01 | 2,800 mentions |
| Shipping | −0.38 | −0.22 | 2,100 mentions |
| Packaging | +0.51 | +0.03 | 1,400 mentions |
| Returns Process | −0.29 | −0.15 | 980 mentions |
| Website Experience | +0.34 | +0.05 | 1,200 mentions |
| Customer Support | +0.12 | +0.09 | 890 mentions |
| Product Range | +0.58 | +0.02 | 760 mentions |
| Stock Availability | −0.44 | −0.31 | 420 mentions |
| Loyalty Program | +0.22 | +0.01 | 310 mentions |
The story the data tells
Product Quality and Pricing are strengths — protect them. But two aspects are deteriorating rapidly: Shipping (−0.22 vs. last quarter) and Stock Availability (−0.31). Stock Availability has the worst sentiment score of all aspects at −0.44 — customers are frustrated about items being out of stock.
Without ABSA, the overall sentiment (+0.23) would look “fine.” The shipping and stock problems would be invisible — drowned out by strong product quality scores.
With ABSA, the team knows exactly where to act. And because the taxonomy is governed, next quarter’s analysis will use the same aspects. The trend lines will be real.
Side by side
What you see with flat sentiment vs. ABSA
Flat sentiment
+0.30
Trend ↑ +0.05
“Customers are moderately positive and getting slightly more positive.”
You don’t know why. You don’t know where to act. You don’t know what’s hiding beneath the surface.
ABSA
+0.30
- Product Quality+0.70
- Support−0.50
- Pricing+0.10
- Onboarding+0.60
- Shipping−0.40
You know product and onboarding drive the positive overall score. You know support and shipping drag it down. You know where to invest and what to fix.
Don’t have a taxonomy? Scoty builds one.
When you upload a new dataset, Scoty (our analysis agent) scans every response, identifies the most frequently mentioned aspects, groups them semantically, and presents a draft for your approval.
You can also start from an industry template — E-commerce, SaaS, Healthcare, Financial Services, Hospitality — and customise it. The taxonomy is yours: not ours, not the AI’s. You control it.
Once approved, the taxonomy applies to all current and future analyses on that dataset. And it’s portable — export it, share it, version it.
Stop averaging away the signal.
Your customers are telling you exactly what they love and what they hate. Aspect by aspect. Upload a dataset and see the full picture in minutes.
EU/Swiss hosting availableGDPR-readyNo credit card required