Sentiment analysis, applied to reviews you can act on
Two different jobs travel under one name, and only the second one is worth paying for.
No credit card required. 30-day free trial.
Sentiment analysis, in one sentence
Sentiment analysis is the automatic classification of a piece of writing as positive, negative or neutral, and increasingly into finer categories than that.
Applied to reviews it does two jobs. The first is a polarity label on the whole review, which is close to redundant when a star rating is already attached. The second is aspect level, which pulls out what the review was actually about, so a run of complaints about waiting is separable from a run about the parking.
The aspect level job is the one that earns its keep. A rating tells you a customer was unhappy. An aspect summary tells you that eleven customers were unhappy about the same thing in the same fortnight, and that is a decision rather than a mood.
Modern implementations use a language model rather than a keyword dictionary, which is why negation and sarcasm are handled better than they used to be, and why the output is a judgement rather than a lookup. It is still wrong sometimes, and it is wrong in ways that are hard to predict in advance.
Where the label and the stars disagree
Take the twenty 4 star reviews from one month and look at what the text was actually about rather than at the rating attached to it.
- 4 star reviews in the month
- 20
- Text classified positive overall
- 17
- Containing a specific complaint about waiting
- 9
- Containing a specific complaint about price
- 2
By star rating this is a good month. By aspect, 9 of those 20 reviews mention waiting, which is a service problem hiding inside a flattering number. The star rating could not have told you that, and neither could a plain positive or negative label.
Two caveats on any figure produced this way. The classifier is imperfect, so treat a count of nine as roughly nine rather than exactly nine. And the sample is only the customers who wrote something, which is never everybody who waited.
Why the value is in the aggregate
Reading forty reviews yourself is perfectly possible, and you will remember the two rudest ones rather than the pattern. A count across a quarter has no such preference, which is the entire argument for doing it by machine.
It also changes what a reply can say. A reply that names the specific thing the reviewer raised reads completely differently from one that thanks them for their feedback, and knowing the aspect before you start writing is what makes that possible at any speed.
GetSetReply analyses each review it pulls in and uses that analysis to draft your reply. The draft is yours to edit and yours to post, which is deliberate: an automated reading of a customer's words is a starting point rather than a verdict.
The misconception: the sentiment score is the number to watch
It is tempting to average the sentiment scores into a single trend line and report it every month.
A polarity average tracks your star average almost perfectly, so it is a second copy of a number you already had with added error. Nothing is learned by watching it move.
Watch the aspects instead. The count of reviews mentioning waiting, cleanliness or one particular member of staff is the output with a decision attached, and it moves before the star rating does.
Sentiment analysis questions
For owners deciding whether any of this beats reading the reviews themselves.
How accurate is sentiment analysis on reviews?
Accurate enough to count with and not accurate enough to quote individually. Treat every category count as approximate, and read the underlying reviews before acting on a small number.
Do I need it if I only get thirty reviews a month?
Not for polarity, which your star ratings already give you. Aspect grouping still saves time at thirty a month, and saves a great deal more at three hundred.
Can it tell me a review is fake?
No. It classifies what a text says, not whether the person who wrote it was real. Suspected fakes go to the platform, which holds the account signals nobody outside it can see.
Does GetSetReply run sentiment analysis?
It analyses each review before drafting a reply, and that analysis is what makes the draft specific to what the customer actually said. You still read the draft and post it yourself.
Is a language model better than a keyword list?
For review text, generally yes, because negation and sarcasm break keyword lists badly. The trade is that the output varies slightly between runs and cannot be traced back to a single rule.
Related terms for reading reviews at volume
Read once, reply well
A draft that names what they said
GetSetReply analyses each review and drafts a reply in your voice. You read it, adjust it, and post it yourself.