Spanish-language text classification for environmental evidence synthesis using multilingual pre-trained models.

Area of Science:

• Environmental Science
• Computer Science
• Information Science

Background:

• Evidence syntheses in environmental science are often monolingual (English-dominant).
• This linguistic bias can lead to skewed results and flawed policy decisions.
• Artificial intelligence (AI) presents a potential solution to incorporate non-English evidence.

Purpose of the Study:

• To develop and evaluate AI-powered text classifiers for screening non-English evidence in environmental syntheses.
• To assess the effectiveness of machine learning models in identifying relevant Spanish-language biodiversity conservation papers.
• To reduce the manual effort and potential bias in evidence synthesis screening.

Main Methods:

• Utilized a corpus of Spanish-language peer-reviewed papers on biodiversity conservation.
• Developed supervised machine learning models for text classification.
• Employed a pre-trained multilingual model for text encoding and class-weights for imbalanced datasets.

Main Results:

• The best-performing model achieved 100% recall, missing no relevant studies.
• Over 70% of irrelevant documents were filtered based on title and abstract alone.
• The approach effectively handled a highly imbalanced dataset (0.79%).

Conclusions:

• AI, specifically multilingual language models and class-weights, can create effective non-English language classifiers.
• This method significantly reduces the time and effort in document screening for evidence syntheses.
• Future work can expand this approach to include diverse non-English scientific literature in global evidence syntheses.