IGARRITZ: adapted web environment for text prediction in Basque

Abstract

Students with limited mobility, for example, those caused by brain paralysis, have adapted tools for writing texts, such as eye-tracking hardware, to select letters and predict words. For instance, they can use eye-tracking hardware to select letters and choose words predicted by the system. These systems offer resources for writing in Basque, and predictionscan be customized by inputting Basque word lists. The primary aim of text prediction is to alleviate the effort involved in typing and to facilitate faster or increased text production. However, writing with Iris is slower and more challenging compared to conventional typing with ten fingers. Furthermore, predictive text functionality in Basque is comparatively less effective than in other languages, offering minimal quality output. Thus, the objective of this study is to develop an adapted web environment for Basque text prediction employing artificial intelligence techniques. To achieve this goal, we have developed a web interface named IGARRITZ based on the HiTZ/roberta-eus-euscrawl-base-cased language model, utilizing a Transformer architecture. It was re-trained with an educational Basque corpus sourced from student texts, educational texts from Gizapedia, Wikipedia, and Berria. Finally, we evaluated the tool and compared it with another currently available system using texts produced by a secondary school student with cerebral palsy. The results indicate that IGARRITZ enhances text prediction in Basque. The student, who writes using eye-tracking technology, reported that the writing process has become significantly easier and more efficient. Additionally, our automatic evaluation demonstrated improved results compared to the existing system.