Artificialy developed a custom language model using Mixtral, an open-source “mixture of experts” architecture particularly effective for European languages. The approach combined cultural expertise with advanced NLP techniques.

Key components of the solution included:

• Custom OCR development. After testing open-source solutions, we built a proprietary OCR system capable of handling Ticinese-specific graphemes and optimizing digitization conditions.

• Strategic dataset construction. Italian was used as a linguistic bridge, progressively enriching the model with Ticinese content from literary works, theater scripts, poetry, and songs to capture diverse registers and formats.

• Linguistic standardization. Official variants and declensions, such as those defined by the Centro di dialettologia di Bellinzona, were adopted to resolve regional ambiguities.

• Iterative training cycles. Over more than ten training iterations, we refined the model to address verb tenses, auxiliary verbs, and the filtering of overly colorful expressions through data augmentation techniques developed with subject-matter experts.

• Targeted validation. Test sets containing previously unseen sentences were created to quantitatively assess translation quality, achieving satisfactory results after multiple refinement cycles.

The final dataset reached tens of thousands of accurately mapped word pairs between Italian and Ticinese.