Cyberbullying is depressingly prevalent. Legal advice and psychological first-aid are of great value to victims, but thinking of the volume of online communication, offering them at scale requires support from smart tools. Then, expanding such a service by language specific models would not be feasible.
SomeBuddy had developed tools to assist their lawyers and psychologists in their jobs. One of these tools is a solution that predicts crime labels such as “defamation” or “child sexual usage” based on user input. Having been trained in Finnish, the model was not directly scalable to other languages, and simply translating the training data would result in unnecessary complexity and numerous models.
A language-agnostic version of the original tool was developed in cooperation with Reaktor to allow utilising the model regardless of input language as well as using new data in any supported language to train it further. As the underlying model supports over 90 languages, tackling cyberbullying in new areas was made a lot easier.