In this study, the authors present a novel methodology adept at decoding multilingual topic dynamics and identifying communication trends during crises. We focus on dialogues within Tunisian social networks during the coronavirus pandemic and other notable themes like sports and politics. We start by aggregating a varied multilingual corpus of comments relevant to these subjects. This dataset undergoes rigorous refinement during data preprocessing. We then introduce our No‐English‐to‐English Machine Translation approach to handle linguistic differences. Empirical tests of this method show high accuracy and F1 scores, highlighting its suitability for linguistically coherent tasks. Delving deeper, advanced modeling techniques, specifically LDA and HDP models, are employed to extract pertinent topics from the translated content. This leads to applying ARIMA time series analysis to decode evolving topic trends. Applying our method to a multilingual Tunisian dataset, we effectively identify key topics mirroring public sentiment. Such insights prove vital for organizations and governments striving to understand public perspectives during crises. Compared to standard approaches, our model outperforms, as confirmed by metrics like coherence score, U‐mass, and topic coherence. Additionally, an in‐depth assessment of the identified topics reveals notable thematic shifts in discussions, with the proposed trends’ identification indicating impressive accuracy, backed by RMSE‐based analysis.
A recommendation system can be used in E-learning to suggest relevant and personalized resources. It uses data about learners' behaviors, such as the courses they have taken, the content they have viewed, and the assessments they have completed, to make recommendations for additional learning opportunities. In this work, we build a hybrid educational video recommender system based on learners’ reviews and ratings. We use a Latent Dirichlet Allocation (LDA) topic model on textual data extracted from educational videos to train language models as an input to a supervised Convolutional Neural Network (CNN) model. Additionally, we use a Latent Factor Model (LFM) to extract educational video features and learners' preferences from their historical data (ratings and reviews) as an output CNN model. In our proposed technique, we use hybrid user ratings and reviews to tackle sparsity and cold start problems in the recommender system. Our recommender uses reviews to predict a new sentiment matrix. For cases where reviews are absent (represented as empty cells in the sentiment matrix) or comments are ambiguous, the system utilizes normalized user ratings from the rating matrix, employing a tailored mathematical framework specifically designed for this purpose.
This study confronts the challenge of analyzing multilingual, code-switched conversations during the COVID-19 pandemic, a context where traditional classifiers often fall short. We developed a cost-sensitive ensemble learning approach that combines a reweighted AdaBoost-SVM model with an SVM as its base learner, specifically designed to effectively manage the imbalanced dataset common in code-switched communication scenarios. A key innovation of our approach is the novel rebalancing of AdaBoost weights. By incrementally adjusting the weights of misclassified samples from both minority and majority classes, we achieve a more balanced classification in each iteration. This strategy significantly improves the accuracy for minority class classification, a common issue with existing models. In the testing phase, we employed a comprehensive selection of both machine and deep learning classifiers, including Naive Bayes, Decision Trees, SMOTEBoost, CNN, Bi-LSTM, etc. These classifiers underwent comprehensive evaluation across two different multilingual datasets, assessed using six distinct metrics, including P-mean. The results from our experiments demonstrate that our proposed ensemble learning approach, fine-tuned with optimal hyperparameters and leveraging M-BERT for feature extraction, achieved remarkable accuracies of 78.84%, 86.56% and 83.96% on the test sets of the CTSA, TUNIZI and combined CTSA-TUNIZI datasets, respectively. This performance not only surpassed traditional classification methods but also outperformed advanced deep learning models, such as Bi-LSTM.