Zovda
Sentiment is the field buyers ask for first and trust last, because everyone has been burned by a label that did not match the sentence.
The tag is not a property of the text. It is a property of the text plus a guideline plus whoever applied it, and only one of those three is frozen.
What the field actually stores
A three-way positive, neutral, negative label collapses several different questions: how the author feels, how the author talks about the subject, and how the audience receives it.
These come apart constantly. A post can be furious about a product and still produce enthusiasm in the replies. Decide which one you are storing, then name the field for it.
Agreement is the real number
Before the distribution, report how often two human raters agree on a sample. A sentiment set with weak agreement has a distribution shaped by noise.
If the number is low, the guideline is the place to fix it, not the model.
Sarcasm is not an edge case
In some verticals a large share of the text is ironic, and the words carry the opposite of the meaning. A set that treats this as a corner case will score well and fail in the feed.
Collect these deliberately. They are also the rows buyers test with.
Refresh labels, keep the text
When the guideline changes, re-label a sample rather than the whole set, and keep both versions. The old label is data about how the standard moved.
The text itself should never be edited. A cleaned-up post is no longer evidence of anything.
Zovda builds these sets through Datuva, where each label keeps its guideline version, so a change in the distribution can be explained by a change in the rules. The social media data collection shows the fields involved.
Tell us what you are trying to measure. We reply within one business day.