Tanjim Mahmud

Video summarization aims to produce a short yet informative summary of a long video while reducing the amount of redundancy. Most transformer-based methods are single-temporal scale or are unconcerned with shot-level structure, limiting temporal coherence and cross-dataset generalization. To fill these gaps, we present HybridHiT-UNet, a supervised framework that combines three complementary parts: a pretrained Vision Transformer encoder to provide spatially rich frame representations, a multi-scale 1D Temporal U-Net backbone to provide hierarchical temporal modeling of frame representations, and a shot-aware hierarchical transformer scoring head to provide inter-shot context to importance prediction. Frame-level scores are summed into shot-level utilities and optimized with a knapsack selection on a fixed-length budget, and a weighted focal loss is used to address extreme class imbalance. Wide experiments using four benchmarks (SumMe, TVSum, OVP, and YouTube) under canonical, augmented, and transfer protocols demonstrate that HybridHiT-UNet achieves F1-scores of 65.8% on SumMe and 79.92% on TVSum, which is higher than the existing methods, which still achieve diversity scores of 64.98% and 48.68%, respectively. A systematic study further demonstrates that a 20% summary budget would yield a consistently superior coverage-diversity trade-off than the traditional 15% one, which provides useful evidence-based advice on the selection of summary length.

The rapid transition to online learning has intensified the need to understand the multifaceted drivers of student satisfaction in virtual education environments. Student satisfaction is a cornerstone of effective digital education, influencing academic performance, learner engagement, and institutional decision-making. As a critical predictor of retention and perceived instructional quality, it directly impacts the sustainability of online learning systems. However, existing studies often lack integrated frameworks that combine subjective expert judgment, global and local model interpretability, and subgroup-specific analysis. To address this gap, this study aims to develop a hybrid decision-analytic and explainable machine learning framework to systematically assess and interpret student satisfaction in virtual classrooms. The proposed framework adopts a layered analytical design in which statistical feature relevance (ANOVA), model-based interpretability (SHAP and LIME), and expert-driven weighting under uncertainty (FAHP) are treated as complementary but non-overlapping components, with TOPSIS used exclusively for decision-level satisfaction ranking. A structured survey has been conducted among 1,469 university students, capturing responses across 21 features encompassing demographic attributes, digital readiness, behavioral patterns, and perceived instructional quality. Feature selection has been performed using the ANOVA F-test, while SHapley Additive Explanations (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) have been applied to derive global and local feature importance. To incorporate subjective expert judgment, feature weights have been calculated using the Fuzzy Analytic Hierarchy Process (FAHP), and composite satisfaction scores have been computed using the Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS). These scores have been used to segment students via Agglomerative Hierarchical Clustering into low (n = 852), medium (n = 484), and high (n = 133) satisfaction cohorts. Cluster-specific SHAP and LIME analyses have uncovered distinct patterns of feature influence across subgroups. The identified satisfaction drivers align closely with constructs from established models of technology acceptance and learning motivation, including the Technology Acceptance Model (TAM), the Unified Theory of Acceptance and Use of Technology (UTAUT), and Self-Determination Theory (SDT). High satisfaction has been linked to emotional positivity, perceived equivalence to offline instruction, and academic self-efficacy, whereas low satisfaction has been driven by affordability barriers, psychological stress, and misalignment with preferred learning modalities. The proposed framework offers a comprehensive, interpretable, and rank-aware approach to modeling student satisfaction, equipping educational stakeholders with actionable insights for designing more inclusive, equitable, and learner-centered online education environments.

Fake news is false or misleading information shared as real news, often meant to trick or influence public opinion. Its fast spread through social media causes confusion, divides communities, and reduces trust in reliable news sources. While much research has been done on fake news detection in English, there is a lack of tools designed specifically for Bangla using a combined linguistic and psychological approach, despite the rise of fake news in Bangla-speaking regions. Our research aims to bridge this gap by developing tailored tools that effectively address the nuances of fake news in the Bangla language. By implementing a robust detection model, we contribute to a deeper understanding of misinformation in this context. In this work, we present a psycholinguistic method that combines linguistic and psychological features of the Bangla language to detect fake news. We created a dataset of 5,428 Bangla news articles, categorized as real or fake, capturing diverse topics and styles. This research highlights the importance of developing language-specific tools to combat misinformation. It also offers a valuable resource for future studies, helping to promote a more informed and responsible media environment in Bangla-speaking communities. Our hybrid model, which integrates Convolutional Neural Networks (CNN) and Multilayer Perceptron (MLP), effectively captures these features, achieving an accuracy of 96.32%. These results demonstrate the strong potential of our approach in improving the detection of fake news in Bangla.

Selecting a source language for zero-shot cross-lingual transfer is typically done by intuition or by defaulting to English, despite large performance differences across language pairs. We study whether linguistic similarity can predict transfer performance and support principled source-language selection. We introduce quantified WALS (qWALS), a typology-based similarity metric derived from features in the World Atlas of Language Structures, and evaluate it against existing similarity baselines. Validation uses three complementary signals: computational similarity scores, zero-shot transfer performance of multilingual transformers (mBERT and XLM-R) on four NLP tasks (dependency parsing, named entity recognition, sentiment analysis, and abusive language identification) across eight languages, and an expert-linguist similarity survey. Across tasks and models, higher linguistic similarity is associated with better transfer, and the survey provides independent support for the computational metrics.

Retrieval-Augmented Generation (RAG) ranks passages by their relevance to a question and passes the top ones to the generator whether or not they contain the answer. Relevance, however, does not imply sufficiency: a passage can be on topic and still lack the fact the question asks for. We place a sufficiency gate between retriever and generator that judges whether the retrieved context is adequate and then answers, abstains, or retrieves more, and we compare four scorers built without any sufficiency-labelled training data—a simple heuristic, a single-passage cross-encoder, a set-level cross-encoder that scores the retrieved passages jointly, and the generator itself prompted as a judge—on 500 HotpotQA and 500 MuSiQue multi-hop questions with Flan-T5-Large. An answer is counted as unsupported when no retrieved passage entails it according to a natural language inference model (DeBERTa-large-MNLI) at a threshold fixed once on held-out data. Because abstaining lowers any error count that ignores unanswered questions, we report the unsupported rate among answered questions alongside coverage. Vanilla RAG is unsupported on 26.2% of HotpotQA and 30.8% of MuSiQue answers. At their calibrated operating points the set-level cross-encoder is the best gate on HotpotQA (17.1% unsupported at 57.2% coverage, with 7.0% of sufficient contexts wrongly refused) and the generator-as-judge on MuSiQue (27.0% at 71.0% coverage, none wrongly refused). Only the judge beats a random abstainer matched to its own coverage on both datasets (both p<0.01), and only the judge ranks best at every threshold. Reliability is paid for in usefulness: on HotpotQA, Exact Match rises from 32.0% to 43.9% among answered questions but falls from 32.0% to 25.0% when abstentions are scored as wrong. Both cross-encoders identify insufficient context better than the heuristic by the area under the receiver operating characteristic curve, decisively on HotpotQA (0.775 and 0.675 against 0.521) but within the noise on MuSiQue (0.738 and 0.717 against 0.658); by selective risk across all operating points the heuristic is worse than not gating at all. Replacing the retrieved context with the gold supporting passages removes about two thirds of the answer-quality failures and between a third and a half of the unsupported answers, so much of the failure, though not all, originates in retrieval rather than the generator. Repeating the evaluation with Qwen2.5-3B-Instruct shifts the absolute rates but preserves the ordering of the scorers.

Log parsing is a fundamental prerequisite for automated system monitoring, anomaly detection, and root cause analysis in large-scale distributed environments. However, existing parsing approaches often rely on heuristic rules, manually engineered features, or fixed similarity thresholds, limiting their adaptability to heterogeneous and evolving log structures. To address these challenges, this study presents SemBERT, a fully unsupervised log parsing framework that integrates semantic BERT embeddings, Incremental Principal Component Analysis (IPCA), HDBSCAN clustering, and adaptive centroid-based cluster merging for robust template mining. Unlike conventional methods that employ fixed merging criteria, SemBERT adaptively determines semantic merging thresholds according to the distribution of cluster centroids, reducing parameter sensitivity across datasets. Experiments on the LogHub HDFS benchmark demonstrated that SemBERT achieved perfect Grouping Accuracy (GA = 100.00%), Parsing Accuracy (PA = 95.82 ± 0.44%), and Template Accuracy (TA = 76.59 ± 4.49%), outperforming several classical parsers in semantic template discovery while maintaining competitive parsing performance. Repeated reproducibility experiments using three different random seed initializations (42, 123, and 999) produced identical adaptive thresholds and final template counts across all evaluated data scales, confirming the deterministic behavior and reproducibility of the proposed framework. Additional scalability experiments on randomly sampled subsets ranging from 50K to 1M log entries demonstrated stable threshold estimation and controlled template growth, highlighting the effectiveness of the adaptive semantic merging strategy. To validate the practical utility of the extracted templates, downstream anomaly detection experiments were conducted using representative machine learning, deep learning, and transformer-based models. The semantic templates generated by SemBERT provide informative representations suitable for downstream log analytics. Furthermore, experiments on the BGL dataset indicate that the framework generalizes effectively to heterogeneous logging environments through lightweight dataset-specific preprocessing. These findings suggest that SemBERT provides a scalable and semantically informed solution for unsupervised log parsing in modern distributed systems.