Publication: Retroalimentación enriquecida de la dinámica del aula utilizando IA
Loading...
Date
2026-05-19
Authors
Pardo García, Federico
item.page.secondaryauthor
Escuela Internacional de Doctorado
item.page.director
Cánovas Reverte, Óscar ; García Clemente, Félix Jesús
Publisher
Universidad de Murcia
publication.page.editor
publication.page.department
DOI
item.page.type
info:eu-repo/semantics/doctoralThesis
Description
Abstract
Tradicionalmente, la evaluación y mejora de la práctica docente se ha basado en la observación humana directa, un enfoque que, pese a ser el estándar de oro, conlleva altos costes logísticos y una inherente subjetividad. El campo de las Analíticas de Aprendizaje Multimodal (MMLA) ha experimentado un avance significativo en la capacidad de monitorizar y modelar entornos educativos. Sin embargo, existe una tendencia predominante en la literatura a perfeccionar las técnicas algorítmicas sin proporcionar mecanismos efectivos de retroalimentación hacia el docente, dejando un vacío crítico en la aplicación práctica de estos avances. Esta tesis doctoral aborda dicha desconexión, proponiendo una arquitectura computacional diseñada para el análisis de clases síncronas, con mecanismos de retroalimentación docente.
La investigación se estructura en cuatro fases metodológicas incrementales. En primer lugar, se realizó una revisión sistemática de la literatura (2014-2024) que permitió taxonomizar el uso de las características de audio y analizar tendencias en el campo de investigación. Esto llevo a la detección de una ausencia de sistemas de retroalimentación docente. En segundo lugar, se modeló la práctica docente utilizando exclusivamente características paralingüísticas derivadas de la diarización de hablantes (como la gestión de turnos y los silencios), capaces de clasificar metodologías (clase magistral, trabajo en grupo y uso de sistemas de respuesta de estudiantes (SRS)) con alta precisión. En tercer lugar, se implementó una fusión multimodal que integra características paralingüísticas con el análisis semántico de transcripciones mediante modelos de lenguaje, aplicando técnicas de Inteligencia Artificial Explicable (XAI) para desambiguar intervenciones docentes complejas en entornos SRS. Finalmente, los hallazgos se integraron en una plataforma web que permite a los docentes visualizar las métricas extraídas.
Los resultados demuestran que el audio contiene niveles estratificados de información pedagógica, donde elementos habitualmente descartados, como el ruido ambiental y el solapamiento de voces, actúan como indicadores válidos de colaboración en el aula. Asimismo, se confirma que la fusión de conocimiento experto (ingeniería de características) con modelos de aprendizaje profundo supera el rendimiento de enfoques puramente textuales en tareas de clasificación de intervenciones. Finalmente, el desarrollo de la plataforma proporciona indicios prometedores de que la entrega de métricas objetivas y transparentes tiene el potencial de activar procesos de reflexión docente.
Traditionally, the evaluation and improvement of teaching practice have been based on direct human observation, an approach that, despite being the gold standard, entails high logistical costs and an inherent subjectivity. The field of Multimodal Learning Analytics (MMLA) has experienced a significant advancement in the capacity to monitor and model educational environments. However, there is a predominant trend in the literature toward perfecting algorithmic techniques without providing effective feedback mechanisms for the teacher, leaving a critical gap in the practical application of these advances. This doctoral thesis addresses this disconnection, proposing a computational architecture designed for the analysis of synchronous classes, with teacher feedback mechanisms. The research is structured into four incremental methodological phases. First, a systematic literature review (2014-2024) was conducted, which allowed for the taxonomization of the use of audio features and the analysis of trends in the research field. This led to the detection of an absence of teacher feedback systems. Second, teaching practice was modeled using exclusively paralinguistic features derived from speaker diarization (such as turn-taking management and silences), capable of classifying methodologies (lecturing, group work, and the use of Student Response Systems (SRS)) with high precision. Third, a multimodal fusion was implemented that integrates paralinguistic features with the semantic analysis of transcriptions through language models, applying Explainable Artificial Intelligence (XAI) techniques to disambiguate complex teaching interventions in SRS environments. Finally, the findings were integrated into a web platform that allows teachers to visualize the extracted metrics. The results demonstrate that audio contains stratified levels of pedagogical information, where elements usually discarded, such as ambient noise and overlapping voices, act as valid indicators of classroom collaboration. Likewise, it is confirmed that the fusion of expert knowledge (feature engineering) with deep learning models outperforms purely textual approaches in intervention classification tasks. Finally, the development of the platform provides promising indications that the delivery of objective and transparent metrics has the potential to activate teacher reflection processes.
Traditionally, the evaluation and improvement of teaching practice have been based on direct human observation, an approach that, despite being the gold standard, entails high logistical costs and an inherent subjectivity. The field of Multimodal Learning Analytics (MMLA) has experienced a significant advancement in the capacity to monitor and model educational environments. However, there is a predominant trend in the literature toward perfecting algorithmic techniques without providing effective feedback mechanisms for the teacher, leaving a critical gap in the practical application of these advances. This doctoral thesis addresses this disconnection, proposing a computational architecture designed for the analysis of synchronous classes, with teacher feedback mechanisms. The research is structured into four incremental methodological phases. First, a systematic literature review (2014-2024) was conducted, which allowed for the taxonomization of the use of audio features and the analysis of trends in the research field. This led to the detection of an absence of teacher feedback systems. Second, teaching practice was modeled using exclusively paralinguistic features derived from speaker diarization (such as turn-taking management and silences), capable of classifying methodologies (lecturing, group work, and the use of Student Response Systems (SRS)) with high precision. Third, a multimodal fusion was implemented that integrates paralinguistic features with the semantic analysis of transcriptions through language models, applying Explainable Artificial Intelligence (XAI) techniques to disambiguate complex teaching interventions in SRS environments. Finally, the findings were integrated into a web platform that allows teachers to visualize the extracted metrics. The results demonstrate that audio contains stratified levels of pedagogical information, where elements usually discarded, such as ambient noise and overlapping voices, act as valid indicators of classroom collaboration. Likewise, it is confirmed that the fusion of expert knowledge (feature engineering) with deep learning models outperforms purely textual approaches in intervention classification tasks. Finally, the development of the platform provides promising indications that the delivery of objective and transparent metrics has the potential to activate teacher reflection processes.
publication.page.subject
Citation
item.page.embargo
Collections
Ir a Estadísticas
Este ítem está sujeto a una licencia Creative Commons. http://creativecommons.org/licenses/by-nc-nd/4.0/