International Journal of Web Research

psychological research, it often suffers from label interference, vocabulary-driven overfitting, and limited labeled datasets. As a result, models are brittle: they can fail with small training samples and behave inconsistently across trait ranges. To address this, we employ a practical single-trait approach that uses five independent ELECTRA-based classifiers, each corresponding to one of the big five dimensions, and trained them as separate binary tasks to prevent cross-trait interference. To reduce lexical bias and double the Pennebaker and King essay corpus from 2,467 to 4,934 samples, the team applied careful synonym-replacement augmentation using WordNet and additionally incorporated contextual augmentation generated by the Gemma model. Models were adjusted methodically to ensure fair comparisons. With test AUCs above 0.75, the ensemble achieves an average test accuracy of 0.724 on the Pennebaker and King benchmark, with per-trait accuracies of 0.72, 0.71, 0.74, 0.73, and 0.72 for openness, conscientiousness, extraversion, agreeableness, and neuroticism (OCEAN), respectively. These results substantially reduce inter-trait interference while matching or surpassing LIWC baselines and other transformer approaches.

Satellite imagery provides valuable data to address the growing demand for agricultural production. However, analyzing such vast amounts of data requires advanced artificial intelligence methods, such as deep learning. The primary challenge lies in the scarcity of labeled training data, as its preparation is both costly and time-consuming. To address this issue, this study integrates remote sensing data, deep neural networks, and transfer learning techniques to estimate the cultivated area of strategic crops in Iran. Given the diverse climates and topographies across Iran’s provinces, in addition to Sentinel-1 and Sentinel-2 satellite data, MODIS sensor and SRTM elevation data were also utilized. To compensate for data limitations, transfer learning was employed to enhance model performance in data-deficient regions (Kermanshah and Markazi). This approach resulted in an approximate 10% improvement in Cohen’s Kappa coefficient. Furthermore, the study investigated the minimum data required for fine-tuning the models. The results demonstrated that even with a reduction of over 60% in the target province's training data, transfer learning still achieved model performance comparable to scenarios where it was not applied.

Driver drowsiness detection plays a critical role in improving road safety, as drowsiness substantially increases the likelihood of traffic accidents. In this study, we propose a novel multi-modal framework within Advanced Driver Assistance Systems (ADAS) that leverages an Attention-Guided Siamese Network coupled with temporal modeling to accurately capture both spatial and temporal patterns of driver fatigue. The Siamese network processes paired facial images, enabling the extraction of discriminative features that highlight subtle changes in driver state. The attention mechanism is explicitly applied to the spatial feature maps within each branch of the Siamese network, allowing the model to focus selectively on key facial regions—such as eyes and mouth—that are most indicative of drowsiness, while also weighting complementary sensor modalities dynamically. Temporal modeling is incorporated through a sequential module (e.g., LSTM or temporal convolution) that analyzes the extracted features over time, capturing gradual and evolving signs of drowsiness that static frame-based methods often overlook. Extensive evaluations on benchmark datasets (YawDD, NTHUDDD) and a novel real-world driving dataset demonstrate superior accuracy exceeding 98.8%, along with strong cross-subject generalization. Ablation studies confirm the critical contributions of the attention mechanism in improving feature discrimination, and the temporal modeling module in enhancing sensitivity to progressive drowsiness. The proposed method surpasses traditional approaches in temporal awareness, data efficiency, and resilience to inter-subject and environmental variations, offering a robust and interpretable solution for real-time driver drowsiness monitoring in intelligent vehicles.

Storytelling is a fundamental pillar of childhood development, where visual narratives play a crucial role in enhancing engagement and cognitive processing. While Generative Artificial Intelligence (GAI) has revolutionized content creation, its application for automated story generation from book covers remains largely unexplored. This study presents an innovative pipeline that combines computer vision for genre classification with G AI to create tailored illustrated stories. After evaluating four deep learning architectures widely used in image classification tasks, ConvNeXt-Tiny was selected as the final model, achieving a Weighted F1-score of 0.6898 in categorizing children's books into 13 distinct genres through cover image analysis. To address the lack of benchmark datasets, we compiled and rigorously validated a specialized collection of 4,085 Persian children's book covers. The proposed system leverages both cover design elements and predicted genre features within structured prompts to generate coherent illustrated stories through LLMs and image-synthesis models. A sample of 26 generated stories was qualitatively evaluated by three child psychologists based on narrative coherence, genre alignment, age appropriateness, character continuity, and visual congruence. This research makes significant contributions to both Persian literary analysis and AI-driven creative systems, demonstrating how machine learning can enhance educational storytelling while preserving cultural authenticity.

Relation extraction between drugs plays a crucial role in identifying drug–drug interactions and predicting side effects. The advancement of machine learning methods in relation extraction, along with the development of large medical text databases, has enabled the low-cost extraction of such relations compared to other approaches that typically require expert knowledge. However, to the best of our knowledge, there are limited datasets specifically designed for drug–drug relation extraction currently available. Therefore, employing transfer learning becomes necessary to apply machine learning methods in this domain. In this study, we propose DREAM, a method that first employs a trained relation extraction model to discover relations between entities and then applies this model to a corpus of medical texts to construct an ontology of drug relationships. The main contribution of this study is adapting the ACORD model to the medical domain using transfer learning, enabling the extraction of domain-specific drug–drug relations and the construction of an ontology. The extracted relations are subsequently validated using a large language model. Quantitative results indicate that the LLM agreed with 71 % of the relations extracted from a subset of PubMed abstracts. Furthermore, our qualitative analysis indicates that this approach can uncover ambiguities in the medical domain, highlighting the challenges inherent in relation extraction in this field.

Social media offers a timely lens into public perceptions of emerging technologies. To assess public opinion on the Internet of Things (IoT), we analyzed a corpus of 824,845 IoT-related posts collected from X between 2013 and 2022. Using Latent Dirichlet Allocation (LDA), we identified seven primary themes of discussion: Smart Home, Business Intelligence, Artificial Intelligence, Smart City, IoT Usage, Emerging Technologies, and Blockchain. We then applied an unsupervised machine-learning technique to evaluate sentiment toward each theme. Overall, public discourse was positive: 46.78% of tweets expressed positive sentiment, 43.41% were neutral, and 9.81% were negative. Although predictable, short-term shifts in tone occurred around specific events, interest in these themes remained consistent throughout the study period. These findings suggest that the Internet of Things is generally perceived favorably and demonstrate how large-scale social media analytics can capture authentic, real-time attitudes toward complex technologies. By linking public opinion to specific topics of discussion, our results provide valuable insights for researchers, policymakers, and product teams seeking to align IoT development with societal expectations.