A definition of aspect-based sentiment analysis (ABSA): how it works, why overall sentiment misleads, and how teams use per-topic scoring. Full disclosure: unitQ publishes this guide and appears as one example among the tools that do this work. 1
Aspect-based sentiment analysis (ABSA) identifies the specific things a piece of feedback talks about, called aspects, and scores the sentiment toward each one separately. 2 A review reading “love the new design, but checkout fails every time” isn’t averaged into “neutral”; it registers as positive on design and negative on checkout. ABSA exists because customers routinely praise and complain in the same breath, and a single overall score erases exactly the detail a product team needs.
How it works
ABSA runs in two steps, whatever the underlying model.
- 1
Aspect extraction finds what the text is about. Aspects can be discovered from the text itself (“delivery,” “battery,” “login”) or mapped onto a predefined feedback taxonomy: the tree of categories a team uses to organize all incoming feedback. Mapping onto a taxonomy is what makes results comparable over time, since freeform aspect names drift.
- 2
Sentiment assignment attaches a polarity to each extracted aspect, using the words around it. This is harder than whole-document scoring because the model must decide which opinion belongs to which aspect, especially in sentences like “the driver was great even though the app kept freezing,” where positive and negative sit a few words apart. LLM-based systems handle this attribution far better than older methods, which is largely why ABSA moved from research papers into production feedback tools.
The output is a matrix rather than a number: aspects on one axis, sentiment and volume on the other, trended over time.
“Love the new design, but checkout fails every time.”
One review, two signals, instead of a single “neutral.”
Why it matters
Overall sentiment answers “how do customers feel?” ABSA answers “how do customers feel about each part of the product?”, which is the question a roadmap actually needs. Three consequences follow.
Mixed feedback stops canceling out
A product where users adore the content but despise the search reads as bland “average sentiment” at document level, hiding both truths.
Trends localize
A two-point drop in overall sentiment is a mystery; a collapse in the “payments” aspect while everything else holds steady is a lead.
Routing gets automatic
Aspect-negative feedback about billing can flow to the billing team without a human triaging it first. Platforms built on an AI taxonomy, unitQ among them, do this aspect-to-owner routing as a standard step. 1
A worked example
An e-commerce app tracks overall review sentiment and sees a flat line for a quarter, which reads as stability. When the team switches on aspect-level scoring, the flat line decomposes: sentiment about product selection and pricing is improving while sentiment about delivery speed is deteriorating at almost the same rate. The two movements had been netting out to zero. Armed with the split, the team takes the delivery data to its logistics partner with three months of evidence, quantified by week and by region. Nothing about the underlying feedback changed; only the resolution of the analysis did.
See how unitQ compares on your data
A short demo, run on your own feedback.
Related terms
The document- or sentence-level parent technique.
Grouping similar feedback into the topics ABSA scores.
The structure aspects map onto for comparable trending.
The tools that do this at scale.
FAQ
See sentiment broken out by topic
Look up any app’s free public unitQ scorecard, or take a demo to see per-aspect sentiment on your own feedback.
Sources 2 references
unitQ, “AI taxonomy with per-topic sentiment, aspect-to-owner routing, real-time monitoring, the unitQ Score.” unitq.com. Accessed August 2026.
IBM, “What is sentiment analysis? — aspect based sentiment analysis (ABSA).” ibm.com/think/topics/sentiment-analysis. Accessed August 2026.