The paper focuses on linguistic complexity and language universals, which are two important and controversial issues in language research. A Fuzzy Property Grammar for determining the degree of universality and complexity of a natural language is introduced. In this task, the Fuzzy Property Grammar operated only with syntactic constraints. Fuzzy Natural Logic sets the fundamentals to express the notions of universality and complexity as evaluative expressions. The Fuzzy Property Grammar computes the constraints in terms of weights of universality and calculates relative complexity. We present a proof-of-concept in which we have generated a grammar with 42B syntactic constraints. The model classifies constraints in terms of low, medium, and high universality and complexity. Degrees of relative complexity in terms of similarity from a correlation matrix have been obtained. The results show that the architecture of a Universal Fuzzy Property Grammar is flexible, reusable, and re-trainable, and it can easily take into account new sets of languages, perfecting the degree of universality and complexity of the linguistic constraints as well as the degree of complexity between languages.

This research analyses the potential and limitations of two artificial intelligence models (ChatGPT 4 and Claude 3.7) as feedback tools for texts written by students of Spanish as a foreign language. Using selected texts from the CEDEL2 corpus, the study evaluates these models' ability to provide pedagogically appropriate corrections. The qualitative analysis focuses on four criteria: accuracy in error detection, clarity of explanations, pedagogical adequacy, and potential problems. Results reveal that both models show high precision in detecting basic errors but present significant limitations such as overcorrection and difficulties adapting explanations to the student's level. Claude stands out for its systematic structure, while ChatGPT excels in humanizing feedback. Based on these findings, we propose an integrated model to implement these tools in the Spanish language classroom, where the teacher maintains a crucial role as mediator, and we offer practical recommendations to maximize their benefits while minimizing their risks.

In this paper, we introduced a fuzzy model for calculating complexity based on universality, aiming to measure the complexity of natural languages in terms of the degree of universality exhibited in their rules. We validated the model by conducting experiments on a corpus of 143 languages obtained from Universal Dependencies 2.11. To formalize the linguistic universals proposed by Greenberg, we employed the Grew tool to convert them into a formal rule representation. This formalization enables the verification of universals within the corpus. By analyzing the corpus, we extracted the occurrences of each universal in different languages. The obtained results were used to define a fuzzy model that quantifies the degree of universality and complexity of both the Greenberg universals and the languages themselves, employing the mathematical theory of evaluative expressions from fuzzy natural logic (FNL). Our analysis revealed an inversely proportional relationship between the degree of universality and the level of complexity observed in the languages. The implications of our findings extended to various applications in the theoretical analysis and computational treatment of languages. In addition, the proposed model offered insights into the nature of language complexity, providing a valuable framework for further research and exploration.

This research analyses the evolution of feedback capabilities between specific artificial intelligence for different Claude Sonnet language models in versions 3.7 and 4.0 as correction tools for texts written by students of Spanish as a foreign language. Through a qualitative comparative analysis of 15 texts from the CEDEL2 corpus, the study evaluates different items: accuracy in error detection, explanatory clarity, pedagogical appropriateness, and problems detected. Claude 4.0 increases error detection by 17% (189 vs 161) and develops greater sophistication in level adaptation, focusing on fundamental errors for beginners while providing comprehensive analysis for advanced students. The latest version introduces improvements in structural organisation through a tripartite format: ‘error → correction → explanation’. However, it presents worrying pedagogical setbacks: it eliminates complementary activities characteristic of Claude 3.7, degrades motivational feedback to generic third-person comments, and maintains biases towards peninsular varieties and hypercorrection. More problematic are the interlinguistic interferences generated by proposals in Spanish and English, resulting in inappropriate Spanglish. The analysis confirms that neither version can function autonomously without teacher mediation, establishing their optimal role as complementary tools with active pedagogical supervision. The findings show that technological evolution in educational AI does not constitute linear improvement, revealing complex exchanges between technical sophistication and pedagogical adequacy. Esta investigación analiza la evolución de las capacidades de retroalimentación entre la inteligencia artificial específica para diferentes modelos de lenguaje Claude Sonnet en sus versiones 3.7 y 4.0 como herramientas de corrección para textos de estudiantes de español como lengua extranjera. Mediante análisis comparativo cualitativo de 15 textos del corpus CEDEL2, el estudio evalúa diferentes ítems: precisión en detección de errores, claridad explicativa, adecuación pedagógica y problemas detectados. Claude 4.0 incrementa la detección de errores en 17% (189 vs 161) y desarrolla mayor sofisticación en adaptación por niveles, concentrándose en errores fundamentales para principiantes mientras proporciona análisis exhaustivos para estudiantes avanzados. La versión más reciente introduce mejoras en organización estructural mediante formato tripartito "error → corrección → explicación". Sin embargo, presenta retrocesos pedagógicos preocupantes: elimina actividades complementarias características de Claude 3.7, degrada la retroalimentación motivacional a comentarios genéricos en tercera persona, y mantiene sesgos hacia variedades peninsulares e hipercorrección. Más problemáticas resultan las interferencias interlingüísticas que generan propuestas en español e inglés, generando un spanglish inadecuado. El análisis confirma que ninguna versión puede funcionar autónomamente sin mediación docente, estableciendo su rol óptimo como herramientas complementarias con supervisión pedagógica activa. Los hallazgos evidencian que la evolución tecnológica en inteligencia artificial educativa no constituye mejora lineal, revelando intercambios complejos entre sofisticación técnica y adecuación pedagógica. Esta investigação analisa a evolução das capacidades de retroalimentação entre a inteligência artificial específica para diferentes modelos de linguagem Claude Sonnet nas suas versões 3.7 e 4.0 como ferramentas de correção para textos de estudantes de espanhol como língua estrangeira. Por meio de uma análise comparativa qualitativa de 15 textos do corpus CEDEL2, o estudo avalia diferentes itens: precisão na deteção de erros, clareza explicativa, adequação pedagógica e problemas detetados. O Claude 4.0 aumenta a deteção de erros em 17% (189 vs 161) e desenvolve maior sofisticação na adaptação por níveis, concentrando-se em erros fundamentais para iniciantes, ao mesmo tempo que fornece análises exaustivas para estudantes avançados. A versão mais recente introduz melhorias na organização estrutural por meio do formato tripartido “erro → correção → explicação”. No entanto, apresenta retrocessos pedagógicos preocupantes: elimina atividades complementares características do Claude 3.7, degrada o feedback motivacional a comentários genéricos na terceira pessoa e mantém vieses em relação às variedades peninsulares e à hipercorreção. Mais problemáticas são as interferências interlinguísticas que geram propostas em espanhol e inglês, criando um spanglish inadequado. A análise confirma que nenhuma versão pode funcionar de forma autónoma sem a mediação do professor, estabelecendo o seu papel ideal como ferramentas complementares com supervisão pedagógica ativa. As conclusões evidenciam que a evolução tecnológica na IA educativa não constitui uma melhoria linear, revelando interações complexas entre sofisticação técnica e adequação pedagógica.

Antoni Brosa Rodríguez, M. Dolores Jiménez López. Proceedings of the 7th Workshop on Research in Computational Linguistic Typology and Multilingual NLP. 2025.