Regionalized models for Spanish language variations based on Twitter.

Area of Science:

• Computational Linguistics
• Natural Language Processing
• Sociolinguistics

Background:

• Spanish is a globally significant language with substantial regional variations in communication.
• Existing NLP models often overlook these linguistic nuances, impacting performance on region-specific tasks.
• Understanding regional language differences is crucial for developing accurate and context-aware AI.

Purpose of the Study:

• To develop and present a comprehensive set of regionalized Spanish language resources.
• To facilitate improved performance of NLP models on tasks sensitive to regional language variations.
• To analyze lexical and semantical similarities across different Spanish-speaking regions.

Main Methods:

• Utilized a 4-year dataset of geotagged public Twitter messages from 26 Spanish-speaking countries.
• Developed regionalized word embeddings using FastText and language models based on BERT.
• Created per-region sample corpora and conducted comparative analyses of linguistic features.

Main Results:

• Introduced novel regionalized FastText word embeddings and BERT language models for Spanish.
• Provided detailed comparisons of lexical and semantical similarities across various Spanish-speaking regions.
• Demonstrated the utility of regional resources in enhancing message classification task performance.

Conclusions:

• The developed regionalized resources effectively capture Spanish language variations.
• These resources significantly improve NLP model performance on region-specific tasks.
• The findings underscore the importance of linguistic regionalization in NLP for global languages.