The Result

Thanks to the cooperation with ATL, TrollWall was able to build and launch entirely new language models. These models have been built on the data delivered exclusively by ATL, so I can safely say that they wouldn’t exist if it wasn’t for ATL work. Additionally, ATL has significantly contributed to the creation and optimization of our existing language models, by delivering several datasets that have been used for the re-training of existing models.

Currently, ATL supports us with training and optimization of 5 language models: Czech, German, English, Polish and Romanian.

This allows TrollWall to expand internationally and to meet our internal KPIs. We have ambitious goals for this year, and I see ATL as a key partner in making these plans reality for us. That will certainly include: continued work on the language data for the existing models, providing complete datasets for the purpose of the implementation of new languages, as well as supporting TrollWall AI with any other linguistic tasks we may need.

Building a new language model is a complex effort from several different teams, and scaling up quickly would be much more difficult. Using high-quality annotated linguistic data is crucial in the process and ATL takes over a big part of our work on it. Instead of managing the freelance annotation teams for all languages internally, we can free up our hands by outsourcing this effort to ATL. Without ATL’s support, we’d probably need to employ 1-2 coordinators who could deal with workload management, hiring, training, and managing the annotation teams – that would be quite a significant cost for a startup company like TrollWall. So far, ATL has delivered over 200 000 labels which we use in the process of language model training. We see good accuracy ratio for the newly created models and the good performance of the model proves that the data annotated by ATL was of high quality.