A Multi-Method Validation Framework for Large-Scale Multilingual Text Analytics

Jan 15, 2026·
Stefano Blando
Stefano Blando
,
Domenica Fioredistella Iezzi
· 0 min read
Abstract
To distinguish genuine findings from methodological artifacts, this paper proposes a validation framework based on method-invariant patterns. Analyzing 999,152 multilingual reviews across 18 independent techniques (from classical clustering to Transformers), we demonstrate that substantive content accounts for 95.4% of variance, while methodological choice explains less than 3%. The study confirms that robust patterns transcend specific algorithms and implementations. Furthermore, while BERT achieves peak accuracy (91.3%), classical approaches like SVM offer comparable performance (89.1%) with a 29-fold reduction in computational cost.
Type
Publication
JADT 2026, Palermo, Italy (in review)
publications
NLP Methodology BERT Graph Neural Networks Deep Learning Validation Framework
Stefano Blando
Authors
Stefano Blando (he/him)
PhD Student in Artificial Intelligence
Stefano Blando is a PhD student in the National PhD Program in Artificial Intelligence at Scuola Superiore Sant’Anna and the University of Pisa. His research lies at the intersection of AI, agent-based modeling, and economics. He studies adaptive multi-agent systems, statistical verification of economic simulations, and robust quantitative methods for financial and socio-economic data.
Authors
Domenica Fioredistella Iezzi