The main goal of this PhD thesis is to test, through two empirical studies, the reliability of a method aimed at automatically assessing Critical Thinking (CT) manifestations in Higher Education students’ written texts. The empirical studies were based on a critical review aimed at proposing a new classification for systematising different CT definitions and their related theoretical approaches. The review also investigates the relationship between the different adopted CT definitions and CT assessment methods. The review highlights the need to focus on open-ended measures for CT assessment and to develop automatic tools based on Natural Language Processing (NLP) technique to overcome current limitations of open-ended measures, such as reliability and costs. Based on a rubric developed and implemented by the Center for Museum Studies – Roma Tre University (CDM) research group for the evaluation and analysis of CT levels within open-ended answers (Poce, 2017), a NLP prototype for the automatic measurement of CT indicators was designed. The first empirical study was carried out on a group of 66 university teachers. The study showed satisfactory reliability levels of the CT evaluation rubric, while the evaluation carried out by the prototype was not yet sufficiently reliable. The results were used to understand how and under what conditions the model works better. The second empirical investigation was aimed at understanding which NLP features are more associated with six CT sub-dimensions as assessed by human raters in essays written in the Italian language. The study used a corpus of 103 students’ pre-post essays who attended a Master's Degree module in “Experimental Education and School Assessment” to assess students' CT levels. Within the module, we proposed two activities to stimulate students' CT: Open Educational Resources (OERs) assessment (mandatory and online) and OERs design (optional and blended). The essays were assessed both by expert evaluators, considering six CT sub-dimensions, and by an algorithm that automatically calculates different kinds of NLP features. The study shows a positive internal reliability and a medium to high inter-coder agreement in expert evaluation. Students' CT levels improved significantly in the post-test. Three NLP indicators significantly correlate with CT total score: the Corpus Length, the Syntax Complexity, and an adapted measure of Term Frequency- Inverse Document Frequency. The results collected during this PhD have both theoretical and practical implications for CT research and assessment. From a theoretical perspective, this thesis shows unexplored similarities among different CT traditions, perspectives, and study methods. These similarities could be exploited to open up an interdisciplinary dialogue among experts and build up a shared understanding of CT. Automatic assessment methods can enhance the use of open-ended measures for CT assessment, especially in online teaching. Indeed, they can support teachers and researchers to deal with the growing presence of linguistic data produced within educational 4 platforms. To this end, it is pivotal to develop automatic methods for the evaluation of large amounts of data which would be impossible to analyse manually, providing teachers and
L'obiettivo principale di questa tesi di dottorato è testare, attraverso due studi empirici, l'affidabilità di un metodo volto a valutare automaticamente le manifestazioni del Pensiero Critico (CT) nei testi scritti da studenti universitari. Gli studi empirici si sono basati su una review critica della letteratura volta a proporre una nuova classificazione per sistematizzare le diverse definizioni di CT e i relativi approcci teorici. La review esamina anche la relazione tra le diverse definizioni di CT e i relativi metodi di valutazione. Dai risultati emerge la necessità di concentrarsi su misure aperte per la valutazione del CT e di sviluppare strumenti automatici basati su tecniche di elaborazione del linguaggio naturale (NLP) per superare i limiti attuali delle misure aperte, come l’attendibilità e i costi di scoring. Sulla base di una rubrica sviluppata e implementata dal gruppo di ricerca del Centro di Didattica Museale – Università di Roma Tre (CDM) per la valutazione e l'analisi dei livelli di CT all'interno di risposte aperte (Poce, 2017), è stato progettato un prototipo per la misurazione automatica di alcuni indicatori di CT. Il primo studio empirico condotto su un gruppo di 66 docenti universitari mostra livelli di affidabilità soddisfacenti della rubrica di valutazione, mentre la valutazione effettuata dal prototipo non era sufficientemente attendibile. I risultati di questa sperimentazione sono stati utilizzati per capire come e in quali condizioni il modello funziona meglio. La seconda indagine empirica era volta a capire quali indicatori del linguaggio naturale sono maggiormente associati a sei sottodimensioni del CT, valutate da esperti in saggi scritti in lingua italiana. Lo studio ha utilizzato un corpus di 103 saggi pre-post di studenti universitari di laurea magistrale che hanno frequentato il corso di "Pedagogia sperimentale e valutazione scolastica". All'interno del corso, sono state proposte due attività per stimolare il CT degli studenti: la valutazione delle risorse educative aperte (OER) (obbligatoria e online) e la progettazione delle OER (facoltativa e in modalità blended). I saggi sono stati valutati sia da valutatori esperti, considerando sei sotto-dimensioni del CT, sia da un algoritmo che misura automaticamente diversi tipi di indicatori del linguaggio naturale. Abbiamo riscontrato un'affidabilità interna positiva e un accordo tra valutatori medio-alto. I livelli di CT degli studenti sono migliorati in modo significativo nel post-test. Tre indicatori del linguaggio naturale sono 5 correlati in modo significativo con il punteggio totale di CT: la lunghezza del corpus, la complessità della sintassi e la funzione di peso tf-idf (term frequency–inverse document frequency). I risultati raccolti durante questo dottorato hanno implicazioni sia teoriche che pratiche per la ricerca e la valutazione del CT. Da un punto di vista teorico, questa tesi mostra sovrapposizioni inesplorate tra diverse tradizioni, prospettive e metodi di studio del CT. Questi punti di contatto potrebbero costituire la base per un approccio interdisciplinare e la costruzione di una comprensione condivisa di CT. I metodi di valutazione automatica possono supportare l’uso di misure aperte per la valutazione del CT, specialmente nell'insegnamento online. Possono infatti facilitare i docenti e i ricercatori nell'affrontare la crescente presenza di dati linguistici prodotti all'interno di piattaforme educative (es. Learning Management Systems). A tal fine, è fondamentale sviluppare metodi automatici per la valutazione di grandi quantità di dati che sarebbe impossibile analizzare manualmente, fornendo agli insegnanti e ai valutatori un supporto per il monitoraggio e la valutazione delle competenze dimostrate online dagli studenti.
Defining and Assessing Critical Thinking: toward an automatic analysis of HiEd students’ written texts / Amenduni, Francesca. - (2021). [10.14274/amenduni-francesca_phd2021]
Defining and Assessing Critical Thinking: toward an automatic analysis of HiEd students’ written texts
AMENDUNI, FRANCESCA
2021-01-01
Abstract
The main goal of this PhD thesis is to test, through two empirical studies, the reliability of a method aimed at automatically assessing Critical Thinking (CT) manifestations in Higher Education students’ written texts. The empirical studies were based on a critical review aimed at proposing a new classification for systematising different CT definitions and their related theoretical approaches. The review also investigates the relationship between the different adopted CT definitions and CT assessment methods. The review highlights the need to focus on open-ended measures for CT assessment and to develop automatic tools based on Natural Language Processing (NLP) technique to overcome current limitations of open-ended measures, such as reliability and costs. Based on a rubric developed and implemented by the Center for Museum Studies – Roma Tre University (CDM) research group for the evaluation and analysis of CT levels within open-ended answers (Poce, 2017), a NLP prototype for the automatic measurement of CT indicators was designed. The first empirical study was carried out on a group of 66 university teachers. The study showed satisfactory reliability levels of the CT evaluation rubric, while the evaluation carried out by the prototype was not yet sufficiently reliable. The results were used to understand how and under what conditions the model works better. The second empirical investigation was aimed at understanding which NLP features are more associated with six CT sub-dimensions as assessed by human raters in essays written in the Italian language. The study used a corpus of 103 students’ pre-post essays who attended a Master's Degree module in “Experimental Education and School Assessment” to assess students' CT levels. Within the module, we proposed two activities to stimulate students' CT: Open Educational Resources (OERs) assessment (mandatory and online) and OERs design (optional and blended). The essays were assessed both by expert evaluators, considering six CT sub-dimensions, and by an algorithm that automatically calculates different kinds of NLP features. The study shows a positive internal reliability and a medium to high inter-coder agreement in expert evaluation. Students' CT levels improved significantly in the post-test. Three NLP indicators significantly correlate with CT total score: the Corpus Length, the Syntax Complexity, and an adapted measure of Term Frequency- Inverse Document Frequency. The results collected during this PhD have both theoretical and practical implications for CT research and assessment. From a theoretical perspective, this thesis shows unexplored similarities among different CT traditions, perspectives, and study methods. These similarities could be exploited to open up an interdisciplinary dialogue among experts and build up a shared understanding of CT. Automatic assessment methods can enhance the use of open-ended measures for CT assessment, especially in online teaching. Indeed, they can support teachers and researchers to deal with the growing presence of linguistic data produced within educational 4 platforms. To this end, it is pivotal to develop automatic methods for the evaluation of large amounts of data which would be impossible to analyse manually, providing teachers andFile | Dimensione | Formato | |
---|---|---|---|
Tesi Amenduni con firma digitale.pdf
accesso aperto
Tipologia:
PDF Editoriale
Licenza:
Dominio pubblico
Dimensione
5.96 MB
Formato
Adobe PDF
|
5.96 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.