How Taimi, the worlds largest LGBTQI+ dating community, used Nyckel for custom content moderation.
Taimi is the world’s largest LGBTQI+ dating community, with an app made specifically for people who identify with that community. With a community this large, content moderation becomes hard to manage.
- Highlights
- 3x reduction in content moderation time.
- 4x increase in auto:moderation coverage.
- 10x cost reduction compared to manual curation.
- 96% curation accuracy.
- Coverages and quality increase over time with Nyckel auto-retrain.
Taimi’s moderation team were dealing with thousands of images and pieces of text per day, both moderating content and developing and optimizing their moderation tools. They wanted a solution that was faster and better quality than keyword-based moderation, and which would take into account the context of the content within the platform. They knew they needed complex moderation with machine learning.
Taimi’s Moderation Manager, Vladislav, came across Nyckel’s content curation solutions via an internet search for ML and AI solutions and was immediately pleased with the simplicity and usability of the interface. The ready-to-use free plan meant that Vladislav’s team could get started straight away.
Taimi’s moderation doesn’t simply involve classifying content as acceptable or unacceptable. They use more than 10 categories during the moderation process, including contextual considerations, for example, whether the content contains external links or social media.
The first approach with Nyckel was to use a single model across half a dozen moderation categories. This was successful, but Vladislav’s team needed more coverage and greater accuracy. They decided to train separate models for their different categories. The next step was to train the first of these stripped-down models.
The model was able to moderate 45% of Taimi’s content with 95% accuracy, and with every training cycle, the numbers improved. Today, most of Taimi’s content is moderated automatically. The moderation time, the time from when the content goes live to when the content has been moderated, has reduced to a few seconds. The moderation team checks a subsample of auto-moderated content to review in order to monitor the model’s continued performance as user behavior changes over time.
Using Nyckel’s Text Classification API, Taimi now automates 60% of all content moderation with an accuracy of 96%. Average moderation time is down 3x and auto-moderation coveration up 4x. Apart from continuing to improve the current model, Vladislav plans to apply Nyckel’s machine learning to the problem of scammers and spammers in future.
