How dumub scores
Most listening tools give every comment a mood and average the lot. Fans of the creator, jokes between commenters and a row of laughing emoji all end up in the brand’s score. dumub asks what a comment is about before it asks how it feels, and every sentiment score tells you how many comments it rests on.
Each comment is read in its own language, slang and emoji included, and answered three ways.
What matters is what the comment takes a side on, not who it’s addressed to. Agreeing with a reviewer that the price is too high is a comment about the brand.
A row of emoji on its own is “can’t tell”. We don’t guess whether 😭 means heartbroken or thrilled. Once a comment names something, the emoji carry its meaning.
The action is whatever the client wants from people: buying or using, voting, going, booking, signing up. Most comments say nothing about it. Positive means they’ll do it or already have; negative means they won’t or have stopped, and switching to a rival counts as negative; held back means they would, but something outside their choice stops them, such as the price, not being eligible or not being able to get there; undecided means they’re waiting to see. From the ones that do, the report makes one call: positive when positive clearly outweighs negative and held back, negative the other way, and even when it’s too close to call. Undecided counts toward neither.
Brand sentiment is worked out on the comments that are about the brand and take a clear side. Everything else is still counted, so you can see how much of the conversation it was, but it doesn’t move the score.
Real comments from a dumub read on Grab, with names removed, and where each one lands.
| Comment | About · side | Counts toward |
|---|---|---|
$350 notchet cut rental, petrol, parking, speeding fines. End up earm $50…on a post about what drivers earn | brand negative | Brand sentiment |
te mo Grab, gracias por consentir a las interfans“love you Grab, thanks for spoiling the fans” | brand positive | Brand sentiment |
Udah males pake ojol, mending bluebird“done with ride-hailing, I’d rather take Bluebird” | brand negative won’t buy | Brand sentiment + intent |
aku bahkan udah 3 kali dapet kejadian kurang mengenakan sama gojek 🥲“I’ve had three bad experiences with Gojek” | competitor negative | Competitor mentions |
nice to see gorgeous Freen againon Grab Thailand’s own post with Freen Sarocha | the post positive | Counted, not scored |
Sukses bang angga“all the best, Angga”, to the video’s creator | poster positive | Counted, not scored |
kak up jg di threads 🙏“post this on Threads too” | something else | Counted, not scored |
❤❤❤👍👍👍 | something else can’t tell | Counted, not scored |
A tool that scores every comment would fold the last four into Grab’s score. None of them is about Grab. Love for the ad or its star shows up in the report as comments about the post, not as love for the brand.
The report also shows the subjects the conversation turns on: the price, the camera, a rival. dumub picks them up from the run’s own comments, not from a list written in advance, and gives each comment up to two. Shares are shares of topic mentions, so they add up to 100%. Comments that name no subject (replies, tags, jokes) aren’t in them.
Each topic says whose read it shows: the side that leads toward whoever most of its comments are about, once at least 30 take a side, and the client’s own read beside it when enough of the client’s comments sit in that topic. Below 30 it says too few were read. The specific things people single out in comments about the client or a rival (a model, a person, a show, a policy) are shown as a share of mentions once at least 30 comments name one.
dumub reads every comment it collects and groups similar ones into conversations. Within each conversation it scores a sample in full and scales the result to the conversation’s size, so a small, loud thread can’t outvote a big, quiet one. Each theme in a report shows its share of the discussion, how many comments it holds and how many were scored for tone, and how its comments about the brand split between negative, mixed and positive. A theme with no comments about the brand shows its size and no split.
“Sick” and “insane” are compliments on a product video. Comments are read for meaning, in the language they were written in.
Emoji on their own are counted as “can’t tell” and kept out of the score.
Love for the creator, the song or the celebrity isn’t counted as love for the brand.
Brand sentiment and every theme say how many comments they rest on, so a score from 40 comments never passes for one from 4,000.
Because much of any comment section isn’t about the brand: replies between commenters, praise for the creator, tags, emoji. dumub reads all of it and tells you how much there was. Only the comments about the brand with a clear side go into the score.
A topic gets a side once at least 30 of its comments about the brand take one. Some topics are big but mostly about something else, like a rival’s phone or the reviewer, and some were read too thinly to call. The report says which.
Yes. Comments are read in the language they were written in, code-switching and local slang included, and quotes in the report carry an English translation.
A comment counts for what it means, not what it says. “So cheap, only 51 million” about a phone is negative.
The formula is the standard one, positive minus negative over the total. The base is different: dumub leaves out comments that aren’t about the brand, so its scores are usually based on fewer comments and are sharper for it. Compare direction and gaps, not decimals.