International Journal of Web Research

Interpretation of dental panoramic radiographs which encompass all teeth as well as portions of the jaw and facial bones is critically important for preventive care and for devising appropriate treatment plans based on clinical findings. However, a high clinical workload or the absence of a specialist may compromise the accurate interpretation of even fundamental conditions, such as the detection of abnormalities. In such cases, artificial intelligence techniques can serve as valuable tools to enhance diagnostic accuracy. This research introduces a modified detection framework based on YOLOv11, incorporating two main architectural enhancements: the addition of a module designed to increase attention to specific regions, and improvements to the multi-scale blocks in the backbone of the network. The post-processing stage also employed methods capable of effectively distinguishing overlapping teeth. Experimental results demonstrate an improvement of over 7 percent in the F1-score compared to the baseline YOLOv11 architecture. The proposed model demonstrates competitive performance compared to models with similar architectures and exhibits satisfactory generalization on an independent dataset that was not utilized during training. Furthermore, relying on the real-time processing capability inherent to the YOLO framework, the proposed method can serve as an effective deep learning engine for integration into web software platforms and tools, enabling rapid and accurate dental radiograph analysis in clinical and telemedicine environments.

The increasing adoption of large language models (LLMs) has intensified concerns about hallucinations—outputs that are syntactically fluent but factually incorrect. In this paper, we propose a method for detecting such hallucinations by evaluating the consistency of model responses to paraphrased versions of the same question. The underlying assumption is that if a model produces consistent answers across different paraphrases, the output is more likely to be accurate. To test this method, we developed a system that generates multiple paraphrases of each question and analyzes the consistency of the corresponding responses. Experiments were conducted using two LLMs—GPT-4O and LLaMA 3–70B Chat—on both Persian and English datasets. The method achieved an average accuracy of 99.5% for GPT-4O and 98% for LLaMA 3–70B, indicating the effectiveness of our approach in identifying hallucination-free outputs across languages. Furthermore, by automating the consistency evaluation using an instruction-tuned language model, we enabled scalable and unbiased detection of semantic agreement across paraphrased responses.

Out-of-domain intent detection in natural language understanding systems faces significant challenges from suboptimal threshold selection and signal degradation through inappropriate normalization techniques. This paper presents an adaptive ensemble thresholding framework that substantially extends our previous conference work by addressing fundamental limitations in existing variational autoencoder-based detection methods. Our approach combines reconstruction loss from variational autoencoders with classifier confidence scores to create a unified detection signal that captures both semantic deviation and prediction uncertainty. The framework incorporates a novel smart scaling strategy that preserves natural separation ratios between in-domain and out-of-domain samples, preventing the signal destruction caused by standard normalization approaches. Through systematic parameter optimization using grid search techniques, the method adaptively determines optimal ensemble weights and threshold selection strategies tailored to specific dataset characteristics. We evaluate our framework across multiple datasets with varying semantic complexity and domain structures, demonstrating consistent performance improvements over baseline variational autoencoder approaches and recent state-of-the-art methods. Compared to our previous VAE-based approach, the framework demonstrates an average performance gain of 3.15 percentage points across all evaluation metrics. Our analysis reveals that ensemble scaling strategy significantly impacts detection performance, with proper signal preservation being more critical than sophisticated threshold selection methods. This work provides a principled approach to adaptive ensemble learning for out-of-domain detection, offering a robust solution that generalizes effectively across diverse datasets and linguistic contexts including low-resource languages like Persian.

Multimedia content due to its inherent audio-visual capabilities plays a significant role in transferring complex/ mystical concepts into the learners. Its effectiveness, is mostly concerned with goal-oriented narration of the story from one side and the level of engagement and understanding of concept by learners from the other side. In this respect, presenting a framework to organize an e-book to facilitate the effective transmission of complex /mystical concept is the focal concern of this article. Furthermore, the proposed framework incorporates rigorous user experience (UEQ) methodology, employing established metrics and qualitative feedback mechanisms to evaluate learners' satisfaction across a comprehensive range of UEQ scales, including attractiveness, hedonic and pragmatic quality in general. Empirical results, derived from experimental studies utilizing this framework, demonstrate a statistically significant and practically noteworthy impact on the learners' ability to grasp and articulate the core meaning embedded within complex and mystical concepts, thereby validating the efficacy of this innovative approach in promoting accessible and meaningful learning experiences. In this regard, the correlation between UEQ scales with learners' learning style model based on Felder-Silverman was investigated. The assessment of the results concerning the relationship between the UEQ scales and the Felder-Silverman learning style dimensions reveals that the multimedia e-book was evaluated as more engaging and innovative by learners whose learning styles, as classified by Felder-Silverman, were characterized as intuitive and visual. This confirms that the e-book's design resonates particularly well with these learning preferences. Moreover, the User Experience Questionnaire (UEQ) results for the presented e-book, when compared to benchmark data from other products assessed using the same standardized UEQ, indicate a positive multimedia potential. Specifically, the e-book demonstrates favorable scores across dimensions such as Attractiveness, Efficiency, Stimulation, and Novelty as well, which well express the potential of proposed multimedia e-book in transferring complex/mystical concept.

Accurate and early prediction of diabetes is crucial for initiating prompt treatment and minimizing the risk of long-term health issues. This study introduces a comprehensive machine learning model aimed at improving diabetes prediction by leveraging two clinical datasets: the PIMA Indians Diabetes Dataset and the Early-Stage Diabetes Dataset. The pipeline tackles common challenges in medical data, such as missing values, class imbalance, and feature relevance, through a series of advanced preprocessing steps, including class-specific imputation, engineered feature construction, and SMOTETomek resampling. To identify the most informative predictors, a hybrid feature selection strategy is employed, integrating recursive elimination, Random Forest-based importance, and gradient boosting. Model training uses Random Forest and Gradient Boosting classifiers, which are fine-tuned and combined through weighted ensemble averaging to boost predictive performance. The resulting model achieves 93.33% accuracy on the PIMA dataset and 98.44% accuracy on the Early-Stage dataset, outperforming previously reported approaches. To enhance transparency and clinical applicability, both local (LIME) and global (SHAP) explainability methods are applied, highlighting clinically relevant features. Furthermore, probability calibration is performed to ensure that predicted risk scores align with true outcome frequencies, increasing trust in the model’s use for clinical decision support. Overall, the proposed model offers a robust, interpretable, and clinically reliable solution for early-stage diabetes prediction.

Retracted publications continue to influence scholarship long after withdrawal. This study assembles a curated set of 169 Iran-affiliated retractions in computer science, data science, and electrical engineering from 2008 to 2024, links them to citing and cited works through two complementary retrieval pipelines, and constructs an expanded citation network of 1'694 nodes and 1,703 edges. We quantify retraction reasons and timing, community structure, node centrality, self-citation patterns, author and institutional concentration, international co-authorship, and a field-adjusted national benchmark. Misconduct-related causes predominate. The average interval from publication to retraction increased into 2021 and has since begun to shorten. The citation network exhibits strong community structure with three major thematic clusters. Centrality profiling isolates five retracted works that function as hubs, often reinforced by self-citation loops. Contribution is highly concentrated among a small set of authors and institutions, while collaboration extends across multiple regions beyond Iran. A field-adjusted retraction rate places the national record among mid-tier producers. These results identify practical leverage points to reduce downstream spread of invalidated findings: persistent indexing flags on hub retractions, routine screening of citations to retracted work, and focused attention on repeat patterns in self-citation and institutional clusters. The study offers a reproducible dual-pipeline approach, a full centrality profile of an enlarged network, and actor-level diagnostics that support targeted integrity interventions.