Aspect-based sentiment analysis: sentiment per theme, not per review

What aspect-based sentiment is

Aspect based sentiment analysis (ABSA) breaks a review into the aspects it talks about, such as staff, price, waiting or product, and assigns sentiment to each one separately. "The designs were stunning and the staff patient, but billing took forever" yields three aspects: product (positive), staff (positive) and billing wait (negative).

In business terms, an aspect is simply a topic in your feedback taxonomy, at whatever level you choose.

Why it matters

• Mixed reviews keep both sides. Review-level sentiment would call the example above positive and lose the billing complaint.
• It points to a cause. "Negative" says nothing; "negative about billing wait" says where to look.
• It makes comparison possible. Stores, months and channels can be compared aspect by aspect.
• It shows strengths. Aspects with high positive share are what customers value.

How it is done

1. Define the aspects: a fixed list of themes and sub-themes with definitions and examples.
2. Find the mentions: identify the words in each review that refer to each aspect.
3. Assign sentiment to each mention: positive, negative or neutral.
4. Handle overlaps: one sentence can touch two aspects; one aspect can be mentioned twice.
5. Keep the evidence: store the words that support each label.

Language models make all five steps practical for large volumes and mixed languages, provided they choose from your list of aspects and quote the supporting words.

A worked example

• Staff: positive ("Staff bahut helpful tha").
• Product range: positive ("Collection is good").
• Price: negative ("prices are a bit high").
• Parking: negative ("Parking was a nightmare").

Review-level sentiment would likely call this mixed or neutral. Aspect-based analysis gives four usable signals, two of them complaints with different owners.

Reporting aspect sentiment

• Negative share by aspect, with the mention count.
• The same, per store and per month.
• Changes against the previous period.
• Positive share for strengths.
• Two or three quotes behind every highlighted number.

Choosing the level of aspect

Aspects can be broad (Staff) or narrow (Staff explaining purity and making charges). Broad aspects give stable numbers; narrow ones point to specific fixes. A three-level structure gets both: report at the theme and sub-theme level, and act at the level of specific issues within them.

Pitfalls

• Letting the model invent aspects each run, which breaks trends.
• Counting a review twice for the same aspect when it mentions it twice.
• Treating neutral mentions as positive.
• Reporting aspect sentiment without counts.

Common questions

Aspect-based sentiment vs review-level sentiment

• Unit: each mention of a topic, versus the whole review.
• Output: sentiment per topic, versus one label.
• Mixed reviews: both sides kept, versus averaged away.
• What it tells you: what to fix, versus how happy customers were overall.
• Relation to the rating: adds new information, versus mostly repeating it.

Checklist

• A fixed list of aspects with definitions.
• Each mention keeps the words that support it.
• Unclear texts counted, not forced into an aspect.
• Negative share reported with counts.
• Results compared month to month for the same sources.

Doing this with Falcon

Falcon measures sentiment for every theme a review mentions, not one score per review. Each mention keeps the exact words that support its label, negative share is reported by topic and month with small samples flagged, and unclear texts are kept visible rather than forced into a category. Every number opens the reviews behind it.