Aspect-Based Sentiment Analysis, With Customer Review Examples
Aspect-based sentiment analysis scores each thing a review talks about, not the review as a whole. Using public hotel, gym and restaurant reviews, here is how it works and why aspect sentiment is not the same as impact o…
Aspect-based sentiment analysis (ABSA) finds the specific things a piece of text talks about, such as the food, the price or a member of staff, and gives each one its own sentiment. A review that says the room was lovely but the bill was wrong gets a positive score for the room and a negative score for billing, instead of one blended score for the whole review.
Below are worked examples from public reviews of a hotel group, a gym chain and a restaurant chain, followed by the limit of the method: knowing how customers feel about an aspect does not tell you how much that aspect moves the rating.
How it differs from document-level sentiment
Most sentiment tools score a whole review, sentence or ticket as positive, negative or neutral. IBM describes aspect-based sentiment analysis as one of the three most widely used types of sentiment analysis, the one that narrows the analysis to a single aspect of a product, service or experience.
The SemEval-2014 shared task, which set up a common benchmark for the method on restaurant and laptop reviews, gives a simple example. "The restaurant was expensive, but the menu was great." is labelled with price as negative and food as positive. A document-level model has to call that sentence one thing, and whatever it picks loses half of what the customer said.
| Document-level sentiment | Aspect-based sentiment | |
|---|---|---|
| Unit scored | The whole review, ticket or sentence | Each thing the text talks about |
| Output | One label or score per text | Several aspect and sentiment pairs per text |
| A mixed review becomes | Neutral or mixed, or whichever side is louder | Positive on one aspect, negative on another |
| What you can report | Share of positive reviews over time | Sentiment for each aspect, location, product or role |
| Common use | Brand tracking, alerting on a spike | Finding which part of the experience is praised or blamed |
Brand trackers often stay at document level, which is fine for watching a trend but not for deciding what to change. Our post on brand sentiment analysis covers that use.
A worked example: three reviews from our briefs
These three reviews are quoted word for word from our published hotel and restaurant briefs. The aspect labels are our own reading, added here to show what the method returns.
| Review | Stars | Document-level | Aspect-level (our reading) |
|---|---|---|---|
| “The hotel is fine, just be careful if you make and pay for your reservation through a platform like Hotels.com or Expedia because the hotel doesn't honor this and will charge you again for the room, even if you explain the situation, they won't refund anything.” Orlando hotel | 1 | Negative | Hotel overall: neutral. Booking through a third-party site: negative. Billing (charged twice): negative. Refund: negative. |
| “Waited for 20 minutes just for them to get the order wrong.” Naperville café | 1 | Negative | Wait time: negative. Order accuracy: negative. |
| “Fantastic hotel. Great pool area. Excellent service. Comfortable bed. Good food, including breakfast. Would definitely stay again. Thumbs up!.” Orlando hotel | 5 | Positive | Pool: positive. Service: positive. Bed: positive. Food: positive. Plus a separate signal that is not sentiment: says they will return. |
The first review shows why aspects matter. The guest says the hotel is fine. What made it a one-star review is a double charge and a refused refund, and a document-level score of negative would file it next to complaints about dirty rooms. The third review shows the other thing worth keeping apart: "would definitely stay again" is an intent to return, which is worth its own field rather than being folded into tone.
Across a whole chain: who gets praised and who gets blamed
Aspects do not have to be topics. In our analysis of 1,995 public Google reviews of 24 Panera cafés in the Chicago metro, one aspect was the staff role a customer names. When a customer names a cashier or counter employee, it is praise 35 times and blame 3 times. When they name a manager, it is 46 praises against 53 complaints, the only role blamed more often than praised.
From the report: Chicago café diagnostic
A document-level read of the same reviews would report that staff comments are mostly positive and stop there. Splitting the sentiment by role points at one layer of the store, which is a different fix (coaching shift managers) from the one a general "improve service" goal suggests.
Why aspect sentiment is not the same as impact on a rating
This is where aspect sentiment is most often misread. A negative aspect is not necessarily the one costing you stars. In the same Panera data, reviews whose main topic is price average 2.86 stars, and reviews about a dirty café average 2.89, so customers sound about equally unhappy about both.
The report then measured what each issue does to the chance of a one- or two-star review on its own, comparing reviews at the same café in the same year and holding review length and reviewer type fixed. A dirty café adds about 12 bad reviews per 100, a directional estimate the report flags as less certain. High prices add about 3, which the report could not tell apart from zero. A wrong or missing order adds about 20 and a long, understaffed wait about 19.
From the report: Chicago café diagnostic
The same gap shows up in hotels and gyms. In 8,504 rated public Google reviews of 45 Marriott-family hotels in Orlando, front desk and check-in is the most-mentioned topic, in 27.2% of comments, and its measured effect on a one- or two-star review is about 2 points, with a confidence interval from -1 to +4. A dirty room adds 34 points and rude staff 44, though staff attitude comes up in only 3.6% of comments. In 9,186 public reviews of 33 Planet Fitness clubs in metro Chicago, broken equipment is the complaint members raise most (122 reviews) and adds 4.7 points to the chance of a bad review; front desk and staff problems add 36.3.
Dimension Labs analysis
In hotels and gyms, the most-mentioned aspect has the smallest effect
Effect of each aspect on the chance of a one- or two-star review, in percentage points, with the other aspects and context held fixed.
Two things cause the gap. Aspects everyone touches, like the front desk or the food, come up in good reviews and bad ones, so they are common without being decisive. And aspects ride along in long, angry reviews: a guest who had a terrible stay describes the whole stay, so check-in sits next to low scores without causing them. Aspect sentiment describes the reviews an aspect appears in. Impact needs a comparison that holds the other things fixed. Our post on which complaints cause bad reviews walks through that comparison for all three industries.
Our research paper What sentiment analysis can't see found a related problem one level down. On 70,450 support conversations, the tone of what customers wrote and whether they were satisfied disagreed in 44% of conversations, so even a well-built sentiment score measures how customers sound more than how things went.
How to run aspect-based sentiment on your own reviews
Write the aspect list from your own operation. Walk the customer journey (booking, arrival, the room or the order, billing, leaving) and name the aspects at the level someone owns. "Service" is too broad to act on; "front desk", "housekeeping" and "manager on duty" are not.
Allow several aspects per review, each with its own polarity. Include mixed and neutral. Forcing one topic per review loses the second problem in a review like the hotel one above.
Keep non-sentiment signals as their own fields. Intent to return, a named employee, the ordering channel and whether a problem was fixed are not positive or negative, and they are often the most useful columns.
Check the labels against a hand-coded sample before you report on them. Our post on checking AI tagging accuracy covers how.
Measure impact separately. Model the rating with every aspect as an input, hold location, period and review length fixed, and test that each effect survives a placebo check. Only then rank aspects by what they cost. See causal inference for the method.
FAQ
What is an aspect in aspect-based sentiment analysis?
An aspect is a part or feature of the thing being reviewed that a customer can have an opinion about, such as the price, the food, the room, the app or a staff role. Each aspect mentioned in a text gets its own sentiment.
Is aspect-based sentiment better than overall sentiment?
It tells you more, because it says which part of the experience the feeling is about. Overall sentiment is still useful for tracking a trend or spotting a spike, and it is cheaper to run.
Can aspect-based sentiment tell me what to fix first?
Not on its own. It tells you how customers feel about each aspect, not how much each aspect changes the rating. In our Panera data, price talk was about as negative as complaints about a dirty café but had no measurable effect on bad reviews. Ranking fixes needs an impact estimate.
Dimension Labs reads every review, ticket, chat and survey answer, writes each aspect and signal into its own Dimension, and joins them to your ratings, revenue and location data so you can see which aspects move the numbers, with confidence intervals on the Prove plan and up. See review intelligence, or start with one file of reviews.