In the past decade, the rapid spread of large volumes of online information among an increasing number of social network users is observed. It is a phenomenon that has often been exploited by malicious users and entities, which forge, distribute, and reproduce fake news and propaganda. In this paper, we present a novel approach to the automatic detection of fake news on Twitter that involves (a) pairwise text input, (b) a novel deep neural network learning architecture that allows for flexible input fusion at various network layers, and (c) various input modes, like word embeddings and both linguistic and network account features. Furthermore, tweets are innovatively separated into news headers and news text, and an extensive experimental setup performs classification tests using both. Our main results show high overall accuracy performance in fake news detection. The proposed deep learning architecture outperforms the state-of-the-art classifiers, while using fewer features and embeddings from the tweet text.

This research investigates topic-based sentiment classification in Greek educational-related data using transformer-based language models. A comparative evaluation is conducted on GreekBERT, XLM-r-Greek, mBERT, and Palobert using three original sentiment-annotated datasets representing parents of students with functional diversity, school directors, and teachers, each capturing diverse educational perspectives. The analysis examines both overall sentiment performance and topic-specific evaluations across four thematic classes: (i) Material and Technical Conditions, (ii) Educational Dimension, (iii) Psychological/Emotional Dimension, and (iv) Learning Difficulties and Emergency Remote Teaching. Results indicate that GreekBERT consistently outperforms other models, achieving the highest overall F1 score (0.91), particularly excelling in negative sentiment detection (F1 = 0.95) and showing robust performance for positive sentiment classification. The Psychological/Emotional Dimension emerged as the most reliably classified category, with GreekBERT and mBERT demonstrating notably high accuracy and F1 scores. Conversely, Learning Difficulties and Emergency Remote Teaching presented significant classification challenges, especially for Palobert. This study contributes significantly to the field of sentiment analysis with Greek-language data by introducing original annotated datasets, pioneering the application of topic-based sentiment analysis within the Greek educational context, and offering a comparative evaluation of transformer models. Additionally, it highlights the superior performance of Greek-pretrained models in capturing emotional detail, and provides empirical evidence of the negative emotional responses toward Emergency Remote Teaching.

Mining social web text has been at the heart of the Natural Language Processing and Data Mining research community in the last 15 years. Though most of the reported work is on widely spoken languages, such as English, the significance of approaches that deal with less commonly spoken languages, such as Greek, is evident for reasons of preserving and documenting minority languages, cultural and ethnic diversity, and identifying intercultural similarities and differences. The present work aims at identifying, documenting and comparing social text data sets, as well as mining techniques and applications on social web text that target Modern Greek, focusing on the arising challenges and the potential for future research in the specific less widely spoken language.

Highly-skilled migrants and refugees finding employment in low-skill vocations, despite professional qualifications and educational backgrounds, has become a global tendency, mainly due to the language barrier. Employment prospects for displaced communities are mostly decided by their knowledge of the sublanguage of the vocational domain they are interested in working. Common vocational domains include agriculture, cooking, crafting, construction, and hospitality. The increasing amount of user-generated content in wikis and social networks provides a valuable source of data for data mining, natural language processing, and machine learning applications. This paper extends the contribution of the authors’ previous research on automatic vocational domain identification by further analyzing the results of machine learning experiments with a domain-specific textual data set while considering two research directions: a. prediction analysis and b. data balancing. Wrong prediction analysis and the features that contributed to misclassification, along with correct prediction analysis and the features that were the most dominant, contributed to the identification of a primary set of terms for the vocational domains. Data balancing techniques were applied on the data set to observe their impact on the performance of the classification model. A novel four-step methodology was proposed in this paper for the first time, which consists of successive applications of SMOTE oversampling on imbalanced data. Data oversampling obtained better results than data undersampling in imbalanced data sets, while hybrid approaches performed reasonably well.