ICT Adoption in Education: Unveiling Emergency Remote Teaching Challenges for Students with Functional Diversity Through Topic Identification in Modern Greek Data

Abstract

This study explores topic identification using text analysis techniques in Modern Greek interviews with parents of students with functional diversity during Emergency Remote Teaching. The analysis focused on identifying key educational themes and addressing challenges in processing Greek educational data. Machine learning models, combined with Natural Language Processing techniques, were applied for topic identification, utilizing cross-validation and data balancing methods to enhance reliability. The findings revealed the impact of linguistic complexity on topic modeling and highlighted the educational implications of analyzing qualitative data in this context. Among the models tested, the Naïve Bayes (Kernel) algorithm performed best when combined with lemmatization-based preprocessing, confirming that text normalization significantly enhances classification accuracy in Greek educational data. The proposed framework contributes to the analysis of qualitative educational data by identifying key parental concerns related to Emergency Remote Teaching. It demonstrates how text analysis techniques could support data-driven decision-making and help guide policy development for the inclusive and effective integration of Information and Communication Technology in education.

1. Introduction

During the COVID-19 pandemic, education systems worldwide transitioned to Emergency Remote Teaching (ERT), transforming in-person instruction into online formats on short notice. This abrupt shift disproportionately impacted students with disabilities, often referred to as students with functional diversity (FD), who encountered barriers in accessing and engaging with remote instruction [1]. According to the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) of the American Psychiatric Association [2], students with FD can be diagnosed with Autism Spectrum Disorder (ASD), dyslexia, Attention Deficit Hyperactivity Disorder (ADHD), physical disabilities, specific learning disorders, or speech and language disorders. With the appropriate support, their needs can be met, and their educational achievements can be promoted, even in educational environments where technology is integrated [3,4]. Due to their unique physical, cognitive, or sensory needs, students with FD require tailored educational strategies. In inclusive settings, many such students struggled to participate meaningfully when their specific needs were not adequately supported. Early parental reports highlighted negative effects of ERT on learning progress and well-being [5]. To address this gap, the present study applies Natural Language Processing (NLP) and Machine Learning (ML) techniques to analyze Modern Greek interview data collected from parents of students with functional diversity. The use of Modern Greek—a low-resource language in NLP—constitutes a novel contribution of this work, advancing both linguistic and educational research [6]. Parents can provide invaluable insights into the ERT experience for FD students to ensure an inclusive and equitable learning environment. By dividing their views into four topics and employing text analysis techniques, this paper aims to reveal the intricacies and implications of the educational shift to ERT and highlight the unique linguistic aspects of a Modern Greek dataset. Four research questions (RQ) were defined, namely:

• RQ1: Which challenges and benefits do parents identify regarding ERT for FD students in Greece?
• RQ2: How effectively can NLP techniques model the primary concerns of parents regarding ERT?
• RQ3: What idiosyncrasies of Modern Greek affect topic identification performance?
• RQ4: How efficiently can ML methods extract key information from educational data to guide policy development for the successful adoption of Information and Communication Technology (ICT) in education?

This study contributed to educational research by analyzing parental perspectives on ERT for students with FD, identifying key topics in qualitative data, and offering insights into the challenges and effectiveness of ERT. By applying NLP and ML, the study provided a structured approach to understanding parental experiences. The study further emphasized the need for targeted interventions to optimize ERT practices and support students with diverse learning requirements.

To the authors’ knowledge, this research is the first to perform topic identification in interviews of parents of students with FD regarding ERT. The text corpus was collected and analyzed in this unique context, setting a precedent for future research. The creation of a dataset in Modern Greek enables a more thorough examination of linguistic complexities. The Greek language, with its intricate grammatical and syntactic rules, poses certain challenges for computational analysis. The innovative approach adopted in this research offers a more detailed understanding that conventional methods may overlook. By incorporating preprocessing and data cleaning procedures, the integrity and quality of the dataset were significantly improved. This preparation ensures that the analysis is not only robust but also sensitive to the subtleties within the text corpus. Such preprocessing techniques are crucial for reducing noise and normalizing the data, thereby providing a solid foundation for deploying sophisticated ML algorithms that can detect and interpret patterns within the corpus more effectively. Also, our novel contribution involved systematically implementing Synthetic Minority Over-sampling Technique (SMOTE) to balance class distributions across the dataset. This process was conducted to address class imbalances, particularly between the minority and majority classes. Topic identification using ML, establishes a foundation for Aspect-Based Sentiment Analysis, aiming at uncovering the diverse emotional responses of parents for each topic and enhancing the significance and value of the research efforts. This study aims to address a substantial gap in the predominantly English-focused research by proposing an innovative method tailored specifically for the Modern Greek language. The present work is an extended version of the prior study [7]. Compared to the previous research approach, the following were performed:

(i)

topic-specific annotation by manual labeling of data to one of four predefined topics,

(ii)

identification of Modern Greek idiosyncrasies and unique linguistic features,

(iii)

feature-value text vectorization based on unigram models,

(iv)

text normalization techniques, lemmatization, words replacement and stemming to refine text preprocessing,

(v)

implementation of SMOTE to address class imbalance and

(vi)

employment of ML techniques for topic identification

(vii)

support for data-driven decision-making for inclusive ICT integration in education.

These findings have the potential to support inclusive education by providing valuable insights for educators and policymakers on improving digital learning strategies, communication, and the educational and psychological aspects of remote learning. They may also contribute to the refinement of instructional approaches, ensuring that digital tools are better aligned with the diverse needs of students. Furthermore, these insights could inform data-driven decision-making processes, facilitating the development of policies that promote equal access to quality education.

2. Related Work

It is important to distinguish ERT from traditional online learning (e-learning or distance learning), as ERT refers to a temporary shift in instructional delivery due to crisis conditions, lacking the systematic design and preparation of conventional online education [8]. Existing studies emphasize the disproportionate challenges faced by students with disabilities during ERT. E-learning further marginalized students with special needs, underscoring the necessity of tailored interventions. In the U.S., e-learning was found to be effective for students with developmental disabilities only when highly individualized support was available [9]. While much of the early research emphasized teacher perspectives, limited work focuses on students and parents at the K–12 level. Methodologically, there is a growing trend in applying NLP and ML to qualitative data. Fang et al. [10] demonstrated that NLP can accurately classify thematic content in patient interviews, achieving over 90% AUC. However, the application of such techniques in languages beyond English remains limited. Greek, in particular, presents challenges due to its morphology and underdeveloped NLP resources [11]. These linguistic constraints highlight the need for developing tools and corpora tailored to low-resource languages like Modern Greek.

Several studies have investigated the impact of ERT on learners with disabilities, with parental testimonies highlighting significant obstacles in adapting to remote learning environments. For example, Abuhammad [12] conducted a qualitative review from parents’ perspectives, revealing that distance learning exacerbated existing inequalities for children with special educational needs. NLP and ML techniques are increasingly applied to qualitative data in educational research, offering scalable alternatives to manual coding. Fitkov-Norris and Kocheva [13] emphasize that thematic analysis supported by NLP and ML can enhance the efficiency and consistency of qualitative insights, particularly in large-scale studies. Similarly, Bani et al. [14] explore the integration of NLP approaches for text processing, demonstrating their applicability in structuring and analyzing unstructured data across domains. However, most existing applications focus on English or other high-resource languages, leaving a considerable gap in NLP research for low-resource languages such as Modern Greek. The present study addresses this gap by applying NLP techniques to Greek-language interview transcripts, integrating computational methods with qualitative educational analysis to generate interpretable and reproducible findings.

2.1. ERT and Students with FD

Due to the Covid-19 crisis, approximately 100 countries closed their schools leaving 1.5 billion students without traditional classroom learning [15]. Most of the educational institutions transitioned to urgent distance education through various methods [16]. Hodges et al. [8] termed this as ERT, distinguishing it from structured online education. Transition to ERT was not smooth [17]. In countries where the educational sector was already inclined towards technological integration, the pandemic merely accelerated the move towards digital education [18]. The success of online instruction was contingent upon several factors, namely the educators’ familiarity with technology and the accessibility of Internet and digital tools [19]. However, in Greece the pre-pandemic plan for K-12 distance education was limited. ERT was hastily implemented to ensure educational continuity, despite general unpreparedness and lack of thorough planning [20,21]. Certain researchers highlighted the limited preparedness in using digital tools and the notable lack of official support; students faced interrupted learning and social isolation, and teachers had to rapidly acclimate to new online platforms and methodologies [22,23]. They employed quantitative, qualitative, and mixed methods to analyze data of surveys from university students, interviews with teachers, administration, students, and parents. Students with FD faced difficulty in adaptation, technical issues, emotional and psychological stress, and lack of personalized support [1,24]. Additionally, due to large online classes and the students’ young age, the presence and support of parents during remote teaching became essential [25,26]. Their collaboration with the teachers, combined with their hands-on guidance at home, improved the efficacy of ERT. This underscored the need for tailored strategies to cater to the diverse needs of all students. Investigating and understanding ERT’s challenges are important, offering a valuable research contribution towards better preparedness in education [27].

2.2. Topic Identification and Text Categorization

Topic identification and text categorization are fundamental tasks in NLP and information retrieval [28]. Topic identification extracts the main idea from a text by classifying or grouping it into predefined categories [29,30]. Its outcome provides an understanding of the content, facilitating tasks like summarizing and recommendation. Text categorization is used to analyze and organize text into categories based on their content [31,32]. Its application is crucial for email filtering, fake news detection, domain identification, and sentiment analysis [33,34,35,36]. Topic identification in interviews refers to the automated categorization of interview transcripts into predefined categories [37]. It can be performed with statistical approaches, latent semantic approaches, traditional ML algorithms, rule-based systems, and Deep Learning (DL) frameworks [29,38,39]. NLP techniques develop the ML models’ ability to categorize text [40,41]. Regarding ML techniques in topic identification, approaches range from statistical (e.g., Term Frequency-Inverse Document Frequency), latent semantic (e.g., Latent Dirichlet Allocation, Latent Semantic Analysis) [42,43], to traditional ML algorithms like Naive Bayes, for its simplicity and computational efficiency in large datasets, and Support Vector Machines for their ability to effectively handle complex data [44,45,46]. Rule-based systems and decision trees also play a significant role, offering either predefined sets of rules or tree-like models for topic identification, respectively [47]. Moving to DL approaches, frameworks like Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) and Transformers have become prominent [48]. Their capability to capture intricate contextual relationships within the text and complex patterns, improves the performance and reliability of topic identification [49,50].

2.3. Topic Identification for Data-Driven Decision Making

Recent studies have demonstrated the effectiveness of topic identification in enhancing decision-making across diverse fields. For instance, Cao et al. [51] introduced a method for large-group emergency decision-making that incorporates topic sentiment analysis to better assess risk, illustrating its value in time-sensitive scenarios. Similarly, Ahne et al. [52] developed a topic discovery and identification framework to support clinical decision-making by facilitating access to relevant information within diabetes-related biomedical literature. These works underscore the potential of topic-based approaches for organizing and interpreting large volumes of text in support of informed decisions, particularly in high-stakes environments such as healthcare and emergency response. This research aims to contribute to the educational domain by applying topic identification techniques to qualitative parental narratives, with the objective of supporting data-driven decision-making in the context of ERT.

2.4. NLP in Low-Resource Languages

Most languages other than English, Spanish, and Chinese have limited resources for NLP, classifying them as low-resource languages indicating that the effectiveness of general NLP techniques decreases significantly when applied them. Furthermore, studies on cross-lingual methods frequently test their models on high-resource languages like Spanish and Chinese, emphasizing the necessity for focused research on low-resource languages [53]. This highlights the critical need for advancements in NLP tailored to those languages, as discussed in the survey on challenges and advances in NLP with a focus on legal informatics and low-resource languages [54].

2.5. NLP for Modern Greek

While the majority of NLP research has focused on English and Chinese, text mining [55,56,57,58] and sentiment analysis on Modern Greek data has also been applied to some extent [59,60,61,62]. However, research remains limited in applying text categorization to Modern Greek interview data and more particularly to education-related data. Existing work concentrates on assigning Modern Greek news articles to a list of different topics [63], classifying legal texts [64], identifying offensive language [57,65], and genre and authorship attribution [66]. Considering the past literature, this study attempts to fill a significant gap in the largely English-centered research by introducing a novel approach specifically designed for the Modern Greek language. Modern Greek, with its distinct linguistic characteristics, poses challenges for text analysis [67]. Its rich inflectional nature means words change their form depending on their grammatical features, making root word recognition crucial. The Modern Greek language contains a vocabulary with many synonyms, complex compound words, and is relatively free in phrase ordering [11]. For example, in our dataset, it is noteworthy that the word “θέμα” (topic) could carry a negative sentiment, translating to “πρόβλημα” (problem). Diverse dialects and varieties of Modern Greek or even “slang” phrases/expressions like “μπούρου μπούρου” (blah blah), or “…είχε αυτό που λέμε στην αργκό ‘τικ’” (…had what we call slangily as ‘tic’) introduce additional intricacies. Idiomatic expressions, deeply rooted in the culture, can be hard for algorithms to interpret [65].

2.6. Preprocessing in Modern Greek

Text preprocessing is an essential process that significantly impacts the accuracy and effectiveness of NLP models [56,68]. Limited resources for NLP in the Modern Greek language hinder the development of effective text processing systems. Existing tools for lemmatization and stemming have various imperfections, struggling with the language’s complex morphology, including infinitives and compound words [69]. Unlike English, where stemming algorithms are mature and effective, Modern Greek NLP tools often fail to handle unique linguistic structures and patterns where the suffix significantly influences the formation of words across different tenses, such as the infinitive, and in cases of irregular or compound verbs. Tools must efficiently identify and process these variations, as suffixes play a crucial role in the morphological structure of the language [70]. This capability is essential for accurate lexical analysis and subsequent applications in NLP tasks. Effective preprocessing is crucial as it directly impacts the performance of NLP and ML models, necessitating improvements to overcome current limitations and develop overall application effectiveness. In this research, existing lemmatization and stemming techniques were utilized, but corrections had to be made to address the mistakes generated by these tools (see Section 3.1.3).

3. Materials and Methods

3.1. Dataset

3.1.1. Research Sample

This research introduces a novel dataset comprising of interviews with parents of students with FD, filling a gap in existing research that has primarily focused on hotel reviews, tweets, product reviews, and Greek legislation corpora [56,60]. The dataset contains 12 semi-structured interviews from parents of FD students, reporting their views regarding the ERT process. This dataset has been also used in the research of [1].

Table 1. Children’s profiles.

Interviewee	Gender	Age	Level of Education	Functional Diversity
I-1	Male	5	Primary	Speech Disorder (stuttering, dysarthria), Physical Disability
I-2	Male	7	Primary	General Learning Difficulties (GLD)
I-3	Male	8	Primary	Attention Deficit Hyperactivity Disorder (ADHD)
I-4	Male	8	Primary	Dyslexia, Developmental Dyscalculia
I-5	Female	9	Primary	General Learning Difficulties (GLD)
I-6	Male	13	Secondary	Dyslexia, Speech Disorder (stuttering)
I-7	Male	5	Primary	General Learning Difficulties (GLD)
I-8	Male	10	Primary	General Learning Difficulties (GLD)
I-9	Male	11	Primary	Attention Deficit Hyperactivity Disorder (ADHD), Aggressiveness
I-10	Male	10	Primary	General Learning Difficulties (GLD)
I-11	Female	10	Primary	Vision Disability
I-12	Female	14	Secondary	Attention Deficit Hyperactivity Disorder (ADHD)

outlines the characteristics of the interviewees’ children, specifying their gender, age, level of education and identified functional diversities. The selection of 12 interviewed parties aimed at a comprehensive representation focusing on: (i) capturing a wide range of FD, (ii) a representative regional distribution, and (iii) a varying range of school grades. To qualify for participation, their children had to: (i) have a confirmed diagnosis of FD and (ii) have participated in ERT classes. The participating group consisted of parents of 9 boys and 3 girls. Throughout the ERT phase, these parents actively supported their children. Each interview lasted approximately for 45 min and was conducted either remotely or in person.

3.1.2. Data Collection and Ethics

Data collection and management consisted of the following steps: Step 1: Conduction of interviews, consisting of 14,827 words of qualitative data in Modern Greek. During in-person interviews, the health guidelines were followed. Recordings were made via a mobile phone app. Step 2: Extraction of oral speech attributes. All pauses, laughter, errors, verbal irony, characteristic and idiomatic phrases, and repetitions (interview noise) were preserved via the transcription (

Table 2. Frequently paralinguistic examples.

Word	Example Type	Count
όχι όχι (“no no”)	Repetition	17
ναι ναι (“yes yes”)	Repetition	10
Ε βέβαια (“um certainly or of course”)	Filler	9
εεεε (“um”)	Filler	7
χαχαχα (“hahaha—laughter”)	Laughter	6
Ε ναι (“um yes”)	Filler	4

), adhering to the methodology employed by other researchers [71,72]. Those attributes are often considered as paralinguistic, though they may confound the performance of NLP algorithms in topic identification tasks.

Prior to the interviews, participants signed consent forms to ensure compliance with General Data Protection Regulation (GDPR) guidelines for the use of the dataset, thereby guaranteeing that the data handling processes respect individual privacy and data rights. They were informed about the anonymization of any personal details and that the data would be used for research purposes. All practices followed GDPR guidelines, with approval from the Ethics Committee of the Ionian University. It is noteworthy that the value of this dataset is not solely determined by its size in number of words but by the profound and detailed qualitative insights it presents. The dataset is publicly accessible and has been released under the CC BY-NC-ND 4.0 [73]. Datasets analyzed in this study are publicly available [74]. The repository provides open access to the data, including metadata, data files, and all associated protocols, supporting transparency, replicability, and further exploration of ERT experiences during the Covid-19 pandemic.

3.1.3. Corpus Linguistic Characteristics

During the transcription process, language attributes unique to oral speech were detected, such as repetitiveness and redundancy, informal or slangy expressions, expressions of hesitation, metaphors, rhetorical questions, spontaneous expressions, words and phrases of intimacy, filler words, ambiguities, interjections, etc. Some instances are shown in

Table 3. Instances of oral speech attributes.

Type	Instance	Translation
Informal	κακά τα ψέματα	truth be told
Metaphor	είχαμε γίνει αυτοκόλλητοι	we have become joined at the hip
Idiom	τα τσαμένα τα παιδιά	the poor fellows
Spontaneity	να πάρω τα βουνά;	should I head for the hills?
Intimacy	κοίτα να δεις	lo and behold
Hesitation	τι να πω τώρα;	now, what can I tell?
Ambiguity	αυτό με τα μικρόφωνα	that one with the microphones
Interjection	εεεε	uhmmm
Repetition	ε βέβαια πολύ, πολύ	eh sure by far, by far
Question	τι να κάναμε;	what should we have done?
Diminutive	ερωτησούλες	little questions
Bad syntax	για εμένα το πρόβλημα μου ήταν ότι εντάξει δεν ξέρω έχει να κάνει και με τον δάσκαλο αυτό πιθανόν	for me the problem was that okay I don’t know it might also have to do with the teacher

. Other features of oral speech such as elongated syllables, e.g., “ναιιιι” (yeaaaah) and laughter notations, e.g., “Χααααχα” (haaaaha) were also noted, as in other research on informal Modern Greek text [65]. ML algorithms encounter challenges in accurately discerning the topic behind certain oral expressions, particularly those imbued with cultural nuances and contextual subtleties.

Consequently, to gain further insights into the characteristics of each class, the top 10 most frequently occurring words for every class are presented in in

Table 4. Linguistic information for the top 10 most frequent words per class.

Class A	Freq.	Class B	Freq.
πρόβλημα (problem)	21	παιδί (child)	55
μάθημα (lesson)	19	δασκάλα (teacher)	44
σύνδεση (connection)	18	μάθημα (lesson)	40
φορές (times)	18	μπορούσε (could)	30
σχολείο (school)	15	σχολείο (school)	22
παιδί (child)	11	τάξη (classroom)	21
κωδικούς (passwords)	10	ώρες (hours)	19
μπορούσε (could)	10	φορές (times)	18
υπολογιστή (computer)	10	έκανε (did)	16
κινητό (mobile phone)	9	σπίτι (home)	15
Class C	Freq.	Class D	Freq.
σχολείο (school)	26	παιδί (child)	61
φίλους (friends)	15	παράλληλη (support teacher)	51
φορές (times)	15	δασκάλα (teacher)	23
δασκάλα (teacher)	14	στήριξη (support)	18
έλεγε (said)	14	σχολείο (school)	15
ώρες (hours)	14	μάθημα (lesson)	12
παιδί (child)	12	τηλεκπαίδευση (ERT)	12
ήθελε (wanted)	11	μπορούσε (could)	10
σπίτι (home)	11	περίπτωση (case)	10
τηλεκπαίδευση (ERT)	11	διάσπαση (ADHD)	9

. This representation provides valuable linguistic information on the distinguishing features of each class.

More specifically, in text categorization tasks, the presence of specific tokens commonly appearing across multiple classes can contribute to misclassification. For instance, tokens like “παιδί” (child), “δασκάλα” (teacher), “σχολείο” (school), “μάθημα” (lesson), “ώρες” (hours), “φορές” (times), “μπορούσε” (could), and “σπίτι” (home) can be highly prevalent in more than two of the four classes examined. Such ambiguity is quantitatively reflected in the confusion matrices of the models. These linguistic intricacies pose challenges for the accuracy of topic identification. Additionally, the study notes that the removal of stopwords such as “δεν” (not) and “μη(ν)” (not) may further complicate matters. Specifically, the phrase “δεν την ήθελε καθόλου” (did not want her at all), contains two negative words (“δεν-did not”-“καθόλου at all”) form a so-called double negation in Modern Greek. Also, in the phrase “φοβόταν μην τον κοροϊδέψουν οι συμμαθητές του” (he was afraid of being made fun of by his classmates), “μην” (do not) is not used in its usual negative sense. It means “μήπως” (maybe) and that can be semantically altered when the stopword “μην” is removed.

3.2. Preprocessing

The preprocessing pipeline, illustrated in Figure 1, outlines the sequential steps applied to the raw textual data prior to model training. Each component of this pipeline is systematically detailed in the following subsections to provide a clear understanding of the methodological approach.

3.2.1. Dataset Preparation-Normalization

Data preprocessing is essential for effective data analysis. It involves data preparation, normalization and cleaning to convert raw data into a machine-friendly format, addressing issues like missing values, outliers, and standardization. In this research, the preprocessing was conducted with RapidMiner Studio (version 10.1) operators and Python (version 3.12) was employed specifically for dataset creation and cleaning tasks. After data collection, important steps such as transcription, anonymization, segmentation into sentences, annotation and filtering classes are essential in preparing the text corpus for analysis (Figure 2). More specifically the process started with:

• Interviews transcription, capturing all elements of orality to ensure the richness and authenticity of the spoken content were preserved. The transcription of interviews was conducted carefully to ensure accuracy and preserve the integrity of the original content.
• Anonymization and pseudonymization. Ensuring data privacy by eliminating any possibility of direct or indirect identification of personal data. Names were replaced with pronouns like “αυτός” (he) or “αυτή” (she), or general nouns such as “παιδί” (child) or “δάσκαλος” (teacher).
• Sentence segmentation. 1019 distinct sentences were identified by using end-of-sentence punctuation marks for demarcation. Segmentation into sentences was a preliminary step crucial for maintaining data integrity and enabling precise and efficient processing. This step was conducted to ensure a structured and organized dataset for analysis, preparing it for the annotation process.
• Annotation. The data was annotated at the sentence level, divided into five topics based on their semantic content by two annotators (one internal and one external). The five predefined topics were: (i) “Material and Technical Conditions”, (ii) “Educational Dimension”, (iii) “Psychological/Emotional Dimension”, (iv) “Learning Difficulties and ERT” and (v) “Not Defined”. “Not Defined” topic was created for annotators to tag phrases or words lacking clear semantic content. This category was later excluded from the final analysis. The internal annotator was a member of the research team, closely involved with the project and possessing an in-depth understanding of the annotation guidelines. The external annotator was a qualified researcher in the field, though not directly involved with the research project, who annotated the data based on provided guidelines, directing them to classify sentences into designated categories based on their semantic content. To improve understanding and ensure accuracy in the annotation process, explicit examples for each classification category were provided to the annotators. Annotators achieved a 93% agreement rate. Some annotators disagreement examples were: (i) “Έκαναν πολλές ασκήσεις στο e-class, κουραζόταν και νευρίαζε” (They did a lot of exercises in e-class, he was getting tired and angry) was annotated either as Educational Dimension or as Psychological/Emotional Dimension, and (ii) “τον αποσυνδέαμε όταν βλέπαμε ότι είχε κουραστεί από το μάθημα” (We disconnected him when we saw that he was getting tired of the lesson) was annotated either as Material and Technical Conditions or as Psychological/Emotional Dimension. In such cases, the final decision was to keep the internal annotator’s choice. This decision was made due to the internal annotator’s deeper involvement in the research, enhancing the credibility and reliability of the annotation process.
• Filter “Not Defined” Class. Then, the “Not Defined” topic, as mentioned above, was excluded, as it encompassed non-informative sentences, for example sentences solely dedicated to interrogative clarifications of the questions, phrases like “thank you”, “you are welcome”, etc. The most frequent words of this topic are presented in

  Table 5. Word occurrences in topic “Not Defined”.

  Word	Count
  ευχαριστώ (“thanks”)	9
  και (“and”)	8
  όχι (“no”)	7
  εγώ (“I”)	6
  αυτά (“that’s all”)	6
  παρακαλώ (“you are welcome”)	4

  .

After removing the “Not Defined” topic, the sentence count dropped from 1019 to 996 (14,731 words). During the preprocessing phase, which included steps such as segmentation, tokenization, lemmatization, and removal of stop words (Section 4), the total count of sentences was reduced to 972. The distribution of sentences to topics (classes) is illustrated in Figure 3, showing data imbalance with 40% of the data belonging to topic “Educational Dimension”. Hereafter, the remaining topics (classes) are referred to as: “A: Material and Technical Conditions”, “B: Educational Dimension”, “C: Psychological/Emotional Dimension”, and “D: Learning Difficulties and ERT”. From the above processes the Initial Dataset was formed.

3.2.2. Versions of Datasets Creation—Combinations of Preprocessing Techniques

Screening the “Initial Dataset” results, it was observed that the initial dataset did not have satisfactory results due to the presence of residual noise. The findings underscore the critical need for more preprocessing measures to improve the dataset’s quality and the subsequent classification accuracy. In order to achieve better results, lemmatization was performed on the corpus. Lemmatization process reduced word variants to their dictionary form, or lemma, by employing vocabulary and morphological analysis utilizing the elcorenewslg-3.7.0 model from spaCy [75]. This model was selected for its extensive vocabulary and superior linguistic accuracy. The model was integrated in the Python environment by installing spaCy and downloading the specific language model, enabling to standardize Greek words to their lemma forms efficiently. This methodology enhances text analysis by aligning with modern NLP practices and ensuring precise handling of Greek linguistic features. After lemmatization, it was observed that not all words were correctly lemmatized, necessitating further corrections. To address this, a Python script was developed to refine the dataset more effectively. This custom script completed the lemmatization process and replaced words that confused the algorithm. For instance, verbs in different tenses were standardized to the infinitive in the present tense or converted to the first-person singular. Similarly, nouns, adjectives, and participles were adjusted to the nominative singular in the masculine gender. Examples of these adjustments include “έλεγα” to “λέω” (“I was saying” to “I say”), “μαθήματος” to “μάθημα” (“of the lesson” to “lesson”), “κουρασμένοι” to “κουρασμένος” (“tired” plural to “tired” singular), and “δάσκαλοι” to “δάσκαλος” (“teachers” to “teacher”). After this, one more preprocessing task was attempted to improve the model’s accuracy by excluding some words that appeared across all four topics, as presented in Section 3.2.1 in Table 4, and created ambiguity that reduced the accuracy of the ML model. These words were “παιδί” (“child”), “δάσκαλος” (“teacher”), “σχολείο” (“school”), “μάθημα” (“lesson”), “ώρες” (“hours”), “φορές” (“times”), “μπορούσε” (“could”), “σπίτι” (“house”), “είμαι” (“I am”), “αυτός” (“he”), “γιατί” (“why”), “άλλος” (“other”), “πηγαίνω” (“I go”), “πολύς” (“many”), “κάποιος” (“someone”), “πρέπει” (“must”), “βέβαια” (“of course”), “δηλαδή” (“that is” or “i.e.,”), “εντάξει” (“okay”), “ειδικά” (“especially”), “οπότε” (“so”), “υπάρχω” (“exist”), “τηλεκπαίδευση” (“ERT”). As a final preprocessing step, was attempted to perform stemming using the Greek Stemmer tool, publicly available in a GitHub repository [76]. This step aimed to further refine the text corpus by reducing words to their root forms, thereby enhancing the model’s ability to accurately interpret and classify the data. From the processes described above,

Table 6. Dataset Versions per Processing Technique.

Dataset Version	Preprocessing Step
LEM	Lemmatization
WR	Word Replacement
WE	Word Exclusion
STEM	Stemming

details the dataset versions corresponding to each processing technique.

All versions of datasets were created and combined to investigate potential improvements in topic identification through various preprocessing tasks. By systematically applying not only one procedure but also the combination of the procedures described above with different order, the aim was to improve the accuracy and reliability of the models. This comprehensive approach generated multiple datasets, each reflecting different preprocessing combinations, to identify the most effective strategy for improving topic identification in the corpus. As a result, from the above processes 16 distinct datasets, each undergoing a minimum of one preprocessing step, were created to be analyzed. All preprocessing combinations presented in Figure 4 and Figure 5.

3.3. Data Cleaning and Vectorization

As mentioned above datasets solely and combinations of them were imported in Rapidminer Studio. The process started by reading data from a spreadsheet file. This file comprised of two columns: (i) text data segmented into sentences and (ii) the topic (class label) to which the text refers to. After reading the data, a subprocess “Text Normalization” was employed, incorporating the “Nominal to Text” operator, which was used to convert nominal features to text format and the “Process Documents from Data” operator, that was utilized to generate a Term Frequency-Inverse Document Frequency (TF-IDF) vector representation, retain metadata, and maintain the original text format. This operator was used to ensure that the tokens could be used to generate a numerical vector representation of each text. The approach involved performing tests on Term Occurrences, Binary Term Occurrences, and Term Frequency, with the TF-IDF approach yielding the best results. In this operator, term pruning was implemented by ranking, which involves retaining terms with rank scores between 0.9 and 1.0 are retained, effectively filtering out terms that are too infrequent (below 0.9) according to the chosen ranking criteria. This method helps focus on the most relevant terms for the analysis while excluding outliers. As part of the “Process Documents from Data” operator, several text processing steps were applied: (i) lowercasing, (ii) tokenization (using non-letter characters as the delimiter), (iii) replacement of specific Greek accented characters with their non-accented versions, (iv) stopword removal using a Greek stopwords dictionary (Greek stopwords), (v) removal of tokens that have a length less than 4 or more than 25 characters. In Figure 6, the Rapidminer workflow is presented. After text processing procedures, the final number of words was 5624 words. The analysis did not include bigrams or trigrams, as the occurrence of them was limited, rendering them ineffective for reliable topic identification. After the “Process Documents from Data” operator, the “Filter Examples” operator was used to exclude the empty attributes that emerged from the above analysis, ensuring they were excluded from the training and testing part. In Figure 6 the Rapidminer process regarding “Cleaning and Vectorization” is presented.

4. Results

4.1. Experimental Setup

The cross validation process involved a 10-fold stratified sampling method to ensure that each fold had a representative distribution of the classes. To address class imbalance within the four classes of the dataset, SMOTE was employed [77]. This technique improves model performance by balancing class distribution, which is crucial for improving the accuracy and generalizability of the ML model’s predictions. The SMOTE operator was configured to perform upsampling with the following parameters: two neighbors, normalization enabled, equalizing classes, automatic detection of minority class, rounding integers, a nominal change rate of 0.5, and appending the generated samples to the original dataset. It was employed to balance one minority class at a time, resulting in a total of three separate balancing processes. The procedure is presented in Figure 7.

During the experiments, the “Optimize Parameters (Grid)” operator was employed to determine the best possible parameters. This operator is a nested subprocess which executed the experiments for all combinations of selected parameter values and identified the optimal parameters based on performance metrics. The optimal parameters were determined by the performance value delivered to the inner performance port. Various algorithms were tested, but only those with the best results were retained. The RapidMiner process for model training and testing is depicted in Figure 8.

4.2. Datasets and Algorithms

All possible combinations of the dataset versions were utilized to evaluate their impact on model performance. As a result, 16 dataset versions were created, each incorporating at least one preprocessing step. The combinations of preprocessing techniques applied to each dataset are presented in the Results section (Section 4.3). This approach aimed to determine whether improvements were achieved through each preprocessing step (lemmatization, stemming, words replacement, words exclusion). By systematically testing each dataset variant or combinations of them, insights were gained into how these preprocessing techniques influenced the accuracy and reliability of topic identification within the corpus. Results are presented below in Section 4.2. The following algorithms were utilized: Generalized Linear Model (GLM), XGBoost, and Naive Bayes (Kernel). Additionally, ensemble methods were employed, specifically Adaboost combined with the Deep Learning H2O algorithm, and a Vote ensemble integrating both GLM and Naive Bayes Kernel algorithms. Various algorithms were tested, but those performed better in our evaluation. Each algorithm was tuned and tested to achieve the best results. Instances of overfitting were identified, where the metrics approached the levels expected from random chance, leading to the decision not to utilize those algorithms further in the analysis. The selected algorithms and the values of their parameters were chosen since they provided the best performance after experimentation. It is important to note that those results were obtained by running five times each experiment and calculating the mean value for each metric. In the Appendix A, the algorithms and their respective parameters are presented.

4.3. Results

The results of various preprocessing procedures on different ML algorithms are presented below in

Table 7. ML experiment results by preprocessing task.

Datasets	Naïve Bayes (Kernel)	GLM	XGBoost	Vote	Adaboost
Initial Dataset	53.18%	51.66%	49.69%	51.59%	47.16%
LEM	53.24%	53.76%	52.16%	53.86%	47.71%
WR	52.23%	51.25%	52.01%	51.23%	47.26%
WE	52.71%	49.55%	50.70%	51.22%	47.49%
STEM	53.28%	50.44%	51.80%	52.18%	46.36%
LEM + WR	53.66%	54.56%	53.54%	54.16%	51.17%
LEM + WE	51.88%	51.80%	48.84%	50.13%	49.06%
LEM + STEM	55.15%	53.26%	51.54%	52.16%	48.29%
WR + WE	50.19%	49.49%	50.24%	53.30%	47.10%
WR + STEM	51.53%	51.65%	52.35%	50.77%	47.93%
WE + STEM	53.87%	49.75%	50.67%	52.70%	46.59%
LEM + WR + WE	55.08%	51.72%	50.41%	54.38%	48.66%
LEM + WR + STEM	51.70%	51.48%	52.96%	49.48%	46.50%
LEM + WE + STEM	54.79%	52.86%	50.11%	53.86%	49.83%
WR + WE + STEM	49.82%	51.02%	49.49%	50.63%	46.84%
LEM + WR + WE + STEM	54.29%	53.42%	51.03%	52.10%	49.21%

LEM = Lemmatization, WR = Word Replacement, WE = Word Exclusion, STEM = Stemming. Bold values highlight the top-performing results.

. A detailed description of the performance for each algorithm follows. The results indicate that different preprocessing techniques improved the performance of each algorithm. Naïve Bayes (Kernel) saw a 1.97% improvement with the combination of LEM and STEM. GLM improved by 2.90% with LEM and WR, while XGBoost saw a 3.85% increase with the same combination. The Vote algorithm’s F1 average increased by 2.79% with LEM, WR, and WE, and Adaboost achieved a 3.01% improvement with LEM and WR. In Figure 9 below is the chart illustrating the comparison between the initial and best F1 scores after preprocessing for each algorithm.

The Naïve Bayes (kernel) algorithm demonstrates varied improvements across different preprocessing procedures. The initial dataset starts with an F1 average of 53.18%. Lemmatization (LEM) shows a minor increase to 53.24%, while Word Replacement (WR) and Word Exclusion (WE) slightly reduce the performance to 52.23% and 52.71%, respectively. Stemming (STEM) improves it marginally to 53.28%. The combination of LEM and WR leads to a notable increase to 53.66%, while LEM combined with WE shows a decrease to 51.88%. LEM and STEM together achieve the highest individual preprocessing improvement at 55.15%. Other combinations, such as WR + WE and WR + STEM, do not significantly enhance performance. The triple preprocessing combination of LEM + WR + WE further boosts the F1 average to 55.08%. Overall, the combination of multiple preprocessing techniques yields the best results for Naïve Bayes (kernel). GLM initially performs with an F1 average of 51.66%. Lemmatization significantly improves this to 53.76%. WR results in a slight decrease to 51.25%, and WE further reduce the score to 49.55%. STEM brings the score close to the baseline at 50.44%. The combination of LEM + WR yields the highest improvement at 54.56%. Interestingly, combining LEM with WE or STEM does not show significant improvements, with scores of 51.80% and 53.26%, respectively. The triple combination of LEM + WR + WE results in a moderate increase to 51.72%, whereas the inclusion of STEM in various combinations such as WR + WE + STEM and LEM + WR + WE + STEM achieves moderate improvements. Therefore, GLM benefits most notably from the LEM + WR combination. XGBoost begins with an F1 average of 49.69%. Lemmatization improves the performance to 52.16%, indicating a substantial benefit. WR and WE alone result in slight improvements, with scores of 52.01% and 50.70%, respectively. STEM also shows a moderate enhancement to 51.80%. The LEM + WR combination achieves the highest individual improvement at 53.54%. However, combining LEM with WE or STEM shows diminished returns, with scores of 48.84% and 51.54%, respectively. The combination of WR + STEM yields the highest performance at 52.35%. The multiple combinations such as LEM + WR + WE or LEM + WR + STEM do not significantly outperform the best individual combinations. Consequently, XGBoost shows the most improvement with the LEM + WR preprocessing combination. The Vote algorithm starts with an F1 average of 51.59%. Lemmatization increases the score to 53.86%, while WR and WE result in slight reductions to 51.23% and 51.22%, respectively. STEM shows a moderate improvement to 52.18%. The LEM + WR combination results in the highest improvement at 54.16%. Other combinations, such as LEM + WE and LEM + STEM, show varying levels of improvement, with scores of 50.13% and 52.16%, respectively. The triple combination of LEM + WR + WE further increase the F1 average to 54.38%, indicating the benefit of multiple preprocessing techniques. Therefore, Vote benefits significantly from the LEM + WR combination and its variations. Adaboost starts with the lowest initial F1 average of 47.16%. Lemmatization shows a slight improvement to 47.71%. WR and WE result in marginal changes, with scores of 47.26% and 47.49%, respectively. STEM reduces performance to 46.36%. The LEM + WR combination provides a notable improvement to 51.17%, which is the highest among the preprocessing techniques. Other combinations such as LEM + WE and LEM + STEM do not significantly enhance performance, showing scores of 49.06% and 48.29%, respectively. The combination of LEM + WR + WE improve the F1 average to 48.66%, indicating a slight benefit from multiple preprocessing techniques. Adaboost benefits the most from the LEM + WR preprocessing combination. In summary, Lemmatization generally proves to be the most effective single preprocessing technique across the algorithms, consistently improving performance. The combination of LEM with WR often results in the highest improvements, particularly for Naïve Bayes (kernel), GLM, and XGBoost. The Vote algorithm also benefits significantly from this combination, while Adaboost shows the most notable improvement with the LEM + WR combination, indicating the robustness of ensemble methods. GLM and XGBoost generally performed well, particularly in dataset combinations with lemmatization and corrections, underscoring their effectiveness in handling structured and normalized text data. It is obvious that combining multiple preprocessing techniques tends to yield better results, highlighting the importance of preprocessing in enhancing ML algorithm performance. For example, datasets with LEM + WR and LEM + WR + WE showed higher performance across most of the algorithms, suggesting that integrating multiple text normalization methods is beneficial. Also, stemming alone or in combination with other techniques (e.g., LEM + STEM) showed moderate improvements but in most cases did not outperform the combinations involving lemmatization. These findings underscore the necessity to tailor preprocessing strategies to specific algorithms to maximize their effectiveness.

Table 8. Confusion Matrix of the Top Performing Algorithm.

	Class A	Class B	Class C	Class D	Class Precision
Class A	99	44	17	6	59.64%
Class B	39	205	60	44	57.75%
Class C	37	95	148	27	48.21%
Class D	6	51	15	88	57.52%
Class Recall	54.70%	52.84%	61.67%	51.16%

below presents the confusion matrix illustrating the performance of the Naïve Bayes (Kernel) algorithm, identified as the top-performing model in this analysis. The best results were performed combining preprocessing with lemmatization and stemming procedures in the initial dataset.

Table 9. Performance comparison with related work.

Paper	Data	Classes	ML	Results
[57]	4779 Greek Twitter posts	offensive 3 (imbalanced)	LSTM and GRU with Attention	89% F1-score
[58]	11,156 Greek teachers’ reviews	5 (imbalanced)	REPTree	57.96% Accuracy
[78]	727 English interviews of cancer patients	3 (balanced)	Decision Tree	75.38% Accuracy
[79]	339 English interviews for mental health	3 (imbalanced)	Transformers	77% Accuracy
This paper	12 Greek interviews for ERT	4 (imbalanced)	Naïve Bayes (Kernel)	55.15% F1-score, 61.67% Recall

presents a comparison with relevant literature. Unlike the current study, Pitenis et al. [57] used an already existing and publicly available corpus, though their dataset consists of informal language from Greek offensive Twitter posts. In contrast, in the current study, a dataset of interviews was created and analyzed. Spatiotis et al. [58] is most closely related to the current study. Both the previous mentioned studies deal with imbalanced classes and focus on unstructured data (reviews). Patra et al. [78] used interviews from web medical articles, categorizing them into sentiment. Although their dataset is also interview-based, it is publicly available and contains domain specific using medical terms. Wouts [79] used 339 interviews for mental health and employed pretrained transformer models. Their interviews contain more extensive content and are publicly available. In the current study a novel dataset was created in Modern Greek, through an unstructured interview process. The dataset is unique in its content, that pertains to ERT for students with FD and depicts the parents’ views. It is restricted in size (due to its very specific domain), compared to the data size used in related work.

5. Discussion

The analysis revealed that the Naïve Bayes (Kernel) algorithm, particularly when applied to dataset combinations involving lemmatization and its corrections, outperformed other models with an F1 average of up to 55.15%. The addition of lemmatization and subsequent corrections consistently improved the performance across most algorithms, indicating that normalizing the text body by reducing words to their base forms and correcting any lemmatization errors significantly improves the model’s understanding of the corpus. The combination of various preprocessing techniques generally yielded better results. Furthermore, ensemble methods like the Vote algorithm also showed robust performance across multiple dataset configurations, reinforcing the value of integrating multiple preprocessing steps. These findings highlight the importance of comprehensive text preprocessing, showing that integrating multiple techniques generally enhances the accuracy and reliability of topic identification models. It is also important to note that in some cases, the results did not improve after the application of SMOTE. From the comparative analysis of the research results with relevant literature (Table 9), it is evident that various research approaches using ML and DL algorithms have reported superior performance metrics. However, this research holds significant value. It introduces a smaller yet novel dataset, unlike others that rely on publicly available datasets.

The findings suggest that linguistic complexity plays a significant role in topic modeling, particularly in the context of a morphologically rich and low-resource language such as Modern Greek. The highest classification accuracy achieved (55.15%) highlights the importance of preprocessing, especially the combination of lemmatization and stemming. The proposed framework could support the systematic analysis of qualitative data collected during ERT, offering a replicable approach for identifying recurring themes in parental narratives. This approach could also assist in data-driven decision-making by providing structured insights into the use of ICT in education, potentially informing inclusive strategies and contributing to the development of more equitable remote learning environments.

6. Conclusions

The dataset serves a dual purpose: (i) it fills a significant gap in the literature by providing a targeted corpus within the specialized context of ERT, and (ii) it constitutes a valuable resource for the Modern Greek linguistic infrastructure. It is unique in its focus on four critical dimensions: (i) Material and Technical Conditions, (ii) Educational Dimension, (iii) Psychological/Emotional Dimension, and (iv) Learning Difficulties and ERT (RQ1). Integrating multiple preprocessing techniques enhances model performance, confirming that thorough text normalization improves ML model quality (RQ3). By dividing interviews into sentences and categorizing them into the four topics, the study contributes a structured resource for Modern Greek topic identification (RQ3). Written in Greek, the dataset contains notable linguistic complexities and includes verbal-oral expressions, enhancing its richness and analytical depth. The intricate structures and diverse vocabulary serve as a strong foundation for evaluating different preprocessing and ML techniques. Despite lower performance indicators, the dataset’s novelty remains crucial. The comprehensive application of preprocessing techniques strengthens its value. In such contexts, lower metrics may offer unique insights into specific topics rather than being considered drawbacks.

The current model’s performance results should not be interpreted as ideal or conclusive. This study is intended as a benchmark investigation based on a small, purposefully selected sample of parents of children with functional diversity. The primary objective was to explore their unique perspectives within a specific educational context during ERT. Given the novelty of both the dataset and the applied methodology, this work serves as an initial step in this research direction, offering early insights and a foundation for future studies. While the present study does not adopt a longitudinal design nor incorporate a multi-stakeholder analysis, these elements are planned as part of a broader research project. Future work will focus on expanding the dataset to include a more diverse set of participants and educational settings. Additionally, the adoption of more advanced learning architectures will be explored to enhance classification performance across topic categories. To our knowledge, no prior studies have applied this methodology to such data in the Greek context. As such, this exploratory study aims to lay the groundwork for more robust and generalizable analyses in subsequent phases. Continued refinement, broader data collection, and methodological improvements are expected to substantially enhance the reliability and applicability of the findings.

Recurring themes in parental interviews revealed challenges faced by students with FD, particularly in digital learning strategies, communication, and educational and psychological aspects of remote learning. This study demonstrated how data-driven approaches could analyze accessibility barriers, improve lesson structure, and refine the teaching experience. These insights could support policymakers in developing evidence-based ICT strategies that meet diverse learner needs. ML provides a structured approach to processing qualitative data, enhancing the understanding of stakeholder perspectives. NLP and ML contributed to optimizing instructional approaches by ensuring ICT tools were adapted to diverse student needs. The extracted information could inform policies for more inclusive and effective digital learning environments (RQ4), as ML methods can transform qualitative insights into actionable knowledge. Despite the dataset’s small size, the initial success in classification shows promise for NLP methodologies in educational and psychological domains that are traditionally underexplored. The dataset enriches the existing corpus and highlights the need for interdisciplinary research involving NLP.

All the research questions were successfully addressed through the study’s findings, contributing to the advancement of topic identification in Modern Greek educational data by applying NLP techniques to analyze parental perspectives during ERT. It provided insights into the impact of linguistic complexity on topic modeling and demonstrated how data-driven approaches and the role of ICT in education can be utilized effectively. As a significant contribution of this paper, topic identification was successfully achieved by employing various preprocessing techniques, demonstrating that these methods can lead to substantial improvements in model performance.

Limitations and Future Work

As an exploratory study, this work provides initial insights based on a limited sample of parents of students with functional diversity. While model performance is not yet optimal, the study introduces a novel dataset and methodology within the Greek educational context. The lack of longitudinal design and absence of multi-stakeholder perspectives are acknowledged limitations. Future work will focus on expanding the dataset and applying more advanced learning models to enhance robustness and generalizability.

As a future research direction, the research team aims to integrate distributed vector representations with Bidirectional Encoder Representations from Transformers (BERT) to enhance semantic analysis capabilities and refine their methodology through targeted attribute selection. Insights from parental interviews highlighted challenges faced by students with FD in digital learning, communication, and psychological aspects of remote education. This study underscored how data-driven approaches could address accessibility issues and improve lesson structures. NLP and ML also helped tailor digital tools to diverse student needs. Investigating the long-term impact of ICT-driven policies on remote learning effectiveness and inclusiveness could further enhance digital education frameworks. A potential direction for future research is dataset expansion to strengthen the robustness of findings. While insights were drawn from parents’ narratives on ERT for students with FD, the intricate nature of the Modern Greek language suggests that a larger dataset could enable more comprehensive analyses. A larger dataset could be formed by incorporating narratives from parents, teachers, and school directors to capture diverse experiences more effectively. The identification of any specific linguistic pattern in Modern Greek text that consistently aligns with each of the four primary topics in the context of ERT is another potential research direction. As a next step, aspect-based sentiment analysis will provide significant utility to the current work. This preparatory step lays the groundwork for aspect-based sentiment analysis, adding further meaning and value to the research efforts. Targeted sentiment analysis for each topic could be performed to automatically estimate the polarity and provide a better understanding of the emotions and sentiments of parents regarding ERT for students with FD. Optimizing the models ensures a robust foundation for more detailed sentiment analysis, ultimately improving the overall impact and applicability of the findings.