Custom content classification
A generic toxicity model was trained on somebody else's idea of harm. It flags the comments your community treats as ordinary and misses the ones written in the language your community actually uses. Sence does not ship a default classifier: yours is built from your own moderation decisions and your own policy taxonomy.
HideANYPersonal HateORHarm WishANDRaceORReligionANDNegativecomments
Applies to: 2 channels · News section
HideANYSpamORScamcomments
Decision
Hidden · 9:14 pm
Matched rule: Identity Based Abuse
Trusted by teams using Sence
How your taxonomy gets built
Start from decisions you have already made.
Sence trains on your own content-policy decisions rather than a generic toxicity corpus. Dedicated workshops with your team establish where your line sits, on your topics, in the language your community writes in.
Your categories, your thresholds.
The classification taxonomy is customer-defined. A section covering sensitive health topics can carry different categories and different thresholds from one covering sport or entertainment, and both can run at once.
It learns from every correction.
Once deployed, every correction your team makes feeds back in. The model gets closer to your policy the longer it runs, rather than staying where a vendor left it.
HideANYPersonal HateORHarm WishANDRaceORReligionANDNegativecomments
Applies to: 2 channels · News section
@terracecal9:12 pm
We were robbed, that ref needs glasses 🤓
NegativeOpponent Hostility
@grimjab_779:14 pmHidden
NegativePersonal HateRace
Matched rule: Identity Based Abuse
What a customer-defined taxonomy lets you do
A toxicity score gives you one number and one threshold. A taxonomy gives you categories you named, thresholds you set, and tags that mean something to your team.
Sence moderates across eight categories of harmful comment, and what belongs inside each of them for your community is yours to define rather than inherited from a vendor.
The same word is not the same comment under a health story and under a match report. Categories and thresholds can differ between sections of your own site and channels.
Custom tags run beside the harm categories, so a campaign, a product line or a fixture can be tracked in the same pass that moderates it.
One taxonomy applies to Facebook, Instagram, YouTube and TikTok, so the same comment is treated the same way whichever network it arrives on.
Human review is available at any classification point, and manual, semi-automatic and fully automatic modes all use the same taxonomy.
Every classification is logged with the action taken and the reasoning, so a decision can be explained by the category it came from rather than defended as a black box.
Accuracy is a question about your taxonomy, not about a benchmark
A model that is 95% accurate at detecting toxicity may be 40% accurate at detecting coordinated harassment targeting your journalists. Both numbers are real and only one of them is about your problem, which is why a vendor accuracy figure quoted against somebody else's categories tells you almost nothing.
Sence does not ship a default classifier. Every classifier is built from your own moderation decisions and your own policy taxonomy. Only what your organisation has trained the system to recognise is applied, which is what produces fewer false positives and fewer missed violations than a generic approach.
- Eight categories to build on: hate, harassment, spam and scams, violence, extremist content, illicit activity, self-harm and sexual content
- Categories and thresholds defined per customer, and per section where you need it
- Custom tags alongside the harm categories
- One policy applied consistently across every connected channel
- Human oversight available at any classification point
What it runs with
Protect
The job the taxonomy exists for: keeping a comment section open and safe rather than switching it off.
Learn moreComment & DM Moderation
The taxonomy applied to every comment and message as it arrives, across four networks, at peak volume.
Learn moreTalent & Team Protection
The categories that matter most when the target is a named person rather than the organisation.
Learn moreNewsrooms feel this first, because a comment that is ordinary under a sports story is not ordinary under a health story. There is a version of this written for news and media.
Frequently asked questions
We take a bespoke, enterprise partnership approach, conducting dedicated workshops to deeply understand your brand strategy and tone of voice. Once deployed, our AI continuously learns and improves with every interaction.
Yes. Sence can moderate automatically in real time, or your team can review flagged comments and decide whether to hide them. Every correction your team makes feeds back into your classifier.
The comparison does not hold, which is the point. A model that is 95% accurate at detecting toxicity may be 40% accurate at detecting coordinated harassment targeting your journalists, so an accuracy figure measured against somebody else's categories cannot be read across to yours. What is measurable is the outcome your team sees: fewer false positives and fewer missed violations than a generic approach.
Yes. The classification taxonomy is customer-defined, so a publisher covering sensitive health topics can set different thresholds and categories from one covering sport or entertainment, and both can run under one policy across every connected channel.
Perspective closes to new API request submissions in February 2026 and reaches full end of life on 31 December 2026, so the deadline is real. The change is not like for like: instead of a toxicity score you tune a threshold against, you get categories your team defined, thresholds you set, and a classifier trained on your own moderation decisions.