The National Library of Sweden has been awarded development access to the Vega EuroHPC JU system making it the first public administration actor in Sweden to access the system. The successful application was a joint effort between KB expert Dr. Love Börjeson and his group, and ENCCS expert Dr. Hossein Ehteshami.
Sentiment analysis of texts and speech-to-text transformation are active areas of research and development in the field of Artificial Intelligence (AI). Two main ingredients of such endeavor are high-quality training data and a suitable deep neural network (NN) model, which uses the training data to tune its parameters. The reward is a system that not only can turn (almost) any speech to text but also “understand” the context and sentiment in it. Modern phones, laptops, and other gadgets are already using this technology to serve their owners. Nonetheless, most of the development in this field emerged around the English language model.
Currently, there is a void for a Swedish counterpart. As a response to this void, the data lab (KBLab) at the National Library of Sweden (Kungliga Biblioteket) developed the KB-BERT model, the Swedish trained transformer model based on Google BERT architecture. KB-BERT, trained on the vast amount of high-quality data solely available at KB, proved to be a game-changer in this area.
With the awarded HPC time on the Vega (https://doc.vega.izum.si/) EuroHPC JU petascale system within EuroHPC JU development call(https://prace-ri.eu/hpc-access/eurohpc-access/eurohpc-ju-benchmark-and-development-access-calls/), training, and deploying of next generation of language models can be significantly facilitated.