Frontiers in Artificial Intelligence最新文献

{"title":"AI-assisted peer review: efficiency for reviewers, burden for authors-a narrative review.","authors":"Asaad Ma Babker","doi":"10.3389/frai.2026.1968385","DOIUrl":"https://doi.org/10.3389/frai.2026.1968385","url":null,"abstract":"<p><p>Artificial intelligence (AI) in scholarly peer review has the potential to simplify the work of editors and reviewers through faster triage, automated reporting completeness checks, and scalable synthesis of reviewer comments. However, it also presents practical, ethical, and technical limitations that may place implementation burdens on authors. I examine bias arising from training data and model architecture, confidentiality risks from commercial AI deployment, accountability gaps when human judgment is displaced, and the growing threat of hallucinated citations to scientific integrity. I further analyze how poorly governed AI integration shifts remediation and formatting burdens onto authors-particularly those from under-resourced or non-Western institutions. I propose hybrid human-AI workflows, mandatory disclosure standards, and equity-conscious implementation frameworks to address these harms. Responsible AI integration in peer review requires transparency, human accountability, and deliberate protection of the diversity of voices represented in scientific literature. These considerations also extend beyond individual actors to recursive feedback loops in which AI-assisted manuscripts are evaluated by AI-assisted review systems, reinforcing the need for human contextual stewardship.</p>","PeriodicalId":33315,"journal":{"name":"Frontiers in Artificial Intelligence","volume":"9 ","pages":"1968385"},"PeriodicalIF":6.7,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13680960/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149239586","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Analysis and generation of motion profiles from positive and negative human actions.","authors":"Alessio Russo, Alessio Saccuti, Eleonora Iotti, Giuseppe Di Cesare","doi":"10.3389/frai.2026.1963524","DOIUrl":"https://doi.org/10.3389/frai.2026.1963524","url":null,"abstract":"<p><p>Human actions convey not only their functional goal, the \"what,\" but also the manner in which they are performed, the \"how.\" The same action may be executed in a <i>Positive</i> or <i>Negative</i> manner depending on the agent's affective state. For example, handing someone an object carefully and politely reflects a positive manner, whereas doing so abruptly and dismissively reflects a negative one. This action modulation, known as Vitality Form (VF), is fundamental to human communication. Despite their importance, VFs remain relatively underexplored. This study investigates the kinematic signatures of <i>Positive</i> and <i>Negative</i> VFs in a pointing action and examines whether they can be reproduced through synthetic motion profiles. To this aim, kinematic data were collected from 40 right-handed participants performing three different pointing actions after recalling pleasant or unpleasant personal academic experiences. Each movement was assigned to either the <i>Positive</i> or <i>Negative</i> VF according to the emotional valence of the recalled experience. Several classifiers were evaluated for predicting VFs from wrist kinematic profiles, achieving a best accuracy of 73.97%. Grad-CAM analysis further provided insights into the kinematic basis of VF discrimination. We then investigated whether the kinematic patterns distinguishing <i>Positive</i> and <i>Negative</i> VFs could also be reproduced in synthetic motion profiles using generative models, achieving a Train-on-Synthetic, Test-on-Real accuracy of 72.50%. Finally, a perceptual study assessed whether the generated velocity profiles conveyed the intended <i>Positive</i> or <i>Negative</i> VFs. Participants identified the corresponding affective category with an accuracy of 75.80%, suggesting that affective information can be transferred to synthetic profiles and retained in a form that remains perceptually recognizable. Overall, this work advances our understanding of how affective states are encoded in human movement, demonstrating that these cues can be modeled, reproduced, and recognized in artificially generated actions. These findings have potential applications in neuroscience and social robotics, enabling artificial agents to recognize human attitudes and adapt their behavior accordingly.</p>","PeriodicalId":33315,"journal":{"name":"Frontiers in Artificial Intelligence","volume":"9 ","pages":"1963524"},"PeriodicalIF":6.7,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13680483/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149239557","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Defending against deepfake speech attacks via feature dropout and dropin regularization in large audio language models.","authors":"Yupei Li, Yixiong Fang, Jiahao Xue, Manuel Milling, Björn W Schuller","doi":"10.3389/frai.2026.1928592","DOIUrl":"https://doi.org/10.3389/frai.2026.1928592","url":null,"abstract":"<p><p>Deepfakes have raised widespread concern owing to their threats to privacy, security, and societal trust, driving growing research interest in effective detection methods. Audio deepfake detection has moved from handcrafted-feature classifiers to end-to-end deep learning architectures and, more recently, to self-supervised speech foundation models and general-purpose speech recognizers such as Whisper. The use of instruction-following speech large language models (LLMs), of which Speech Audio Language Music Open Neural Network (SALMONN) is a representative example, remains comparatively less explored, and forms the setting of this study. We argue that large language models (LLMs), with their powerful representational and information retrieval capabilities, hold significant potential for this task-provided that relevant audio features are carefully selected and prioritized. To this end, we employ a speech LLM, specifically SALMONN, for deepfake audio detection, incorporating a two-stage training strategy in which Stage 1 injects acoustic-feature awareness into the Low-Rank Adaptation (LoRA) adapters via an auxiliary feature-prediction pretext task (referred to as feature dropin), and Stage 2 performs windowed random retention on the encoder token sequence (referred to as feature dropout) before the final real/fake decision. Our approach achieves an absolute accuracy improvement of 0.188 on the FakeOrReal benchmark and attains state-of-the-art performance on the Automatic Speaker Verification Spoofing and Countermeasures (ASVspoof) 2019 LA subset with an Equal Error Rate (EER) of 0.00496, representing a 33.0% relative improvement over prior methods. Furthermore, the model transfers non-trivially to an out-of-domain artificial intelligence (AI)-generated music dataset (M6), reaching accuracy comparable to a ResNet18 baseline reported in prior study, which we interpret as evidence of transferable audio representations rather than broad cross-domain robustness.</p>","PeriodicalId":33315,"journal":{"name":"Frontiers in Artificial Intelligence","volume":"9 ","pages":"1928592"},"PeriodicalIF":6.7,"publicationDate":"2026-10-06","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13680471/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149239615","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Machine learning approaches for phishing detection and cybersecurity awareness in digital infrastructure.","authors":"Parves Alam, Md Azizul Rahaman, Jubayer Uddin Shamim, Shakhawat Hossain Refat, Wali Ahmed, Pappu Roy, Shaon Reza, Samuel Kim","doi":"10.3389/frai.2026.1875321","DOIUrl":"10.3389/frai.2026.1875321","url":null,"abstract":"<p><strong>Introduction: </strong>Phishing remains a dominant socio-technical threat to digital infrastructure, serving as an entry vector for credential-theft, ransomware, and supply-chain incidents. The machine learning literature on phishing-URL detection remains fragmented across heterogeneous datasets, inconsistent leakage controls, and explanation pipelines seldom cross-validated by independent attribution methods.</p><p><strong>Methods: </strong>An end-to-end, leakage-controlled machine learning framework was developed over a balanced corpus of 11,430 instances and 87 engineered URL-lexical, HTML-content, and external-reputation features, with all preprocessing transforms fit on the training partition only. Eleven heterogeneous classifiers were benchmarked; the best classical model (LightGBM) was repurposed as a deterministic feature extractor whose out-of-fold posterior probabilities augmented the input to a hybrid CNN-BiLSTM. Statistical significance was assessed via Wilcoxon signed-rank tests with Holm-Bonferroni correction on 10-fold cross-validation <i>F</i> ₁ scores, and two explainable-AI methods (Tree SHAP and Permutation Importance) were applied alongside ablation studies.</p><p><strong>Results: </strong>LightGBM attained <i>F</i> ₁ = 0.9699 and AUC = 0.9943, significantly outperforming every other classifier (<i>p</i> ≤ 0.002) except XGBoost (<i>p</i> = 0.375). The CNN-BiLSTM hybrid attained <i>F</i> ₁ = 0.9652 and recall = 0.9703. SHAP and Permutation Importance agreed on 11 of 15 top-ranked features, separating external-reputation indicators (Google indexing, page rank, domain age, web traffic) from URL-lexical complexity indicators (length, dots, hyphens, phishing hints).</p><p><strong>Discussion: </strong>Strict leakage control confirms that the attainable performance is intrinsically high rather than artifactual. The convergent dual-method explanations translate into six candidate cybersecurity-awareness guidelines for end-user training, connecting technical detection to the human factors that shape enterprise security posture.</p>","PeriodicalId":33315,"journal":{"name":"Frontiers in Artificial Intelligence","volume":"9 ","pages":"1875321"},"PeriodicalIF":6.7,"publicationDate":"2026-10-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13677868/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149224955","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Automated machine learning to foster AI development processes: an explorative study.","authors":"Anna Katharina Polzer, Stefan Thalmann","doi":"10.3389/frai.2026.1867883","DOIUrl":"10.3389/frai.2026.1867883","url":null,"abstract":"<p><strong>Introduction: </strong>Communication barriers between citizen developers and AI experts are a serious challenge for AI system development in the age of No and Low Code (LCNC) AI development. Automated Machine Learning (AutoML) has emerged as a promising approach to democratize AI development by enabling non-technical stakeholders to participate in the creation of AI systems. However, its role in supporting collaboration between citizen developers and AI experts remains underexplored.</p><p><strong>Methods: </strong>This study examines AI experts' perceptions of AutoML's potential as a boundary object that facilitates communication and collaboration between citizen developers and AI experts. An exploratory qualitative interview study was conducted with 14 AI experts from various industries to examine their perceptions of how GUI-based low- or no-code AutoML prototypes may support collaboration.</p><p><strong>Results: </strong>The findings indicate that AutoML can provide a shared frame of reference that supports knowledge exchange, mutual understanding, and interdisciplinary teamwork. Interviewees highlighted that AutoML-generated insights and early-stage prototypes facilitate iterative, feedback-driven development processes and help overcome data and modeling barriers. However, the effective integration of AutoML requires appropriate governance structures, including clear responsibilities, guided processes, and structured stakeholder interaction.</p><p><strong>Discussion: </strong>This exploratory study highlights AI experts' perceptions that AutoML may serve not only as a technological tool but also as a potential facilitator of collaboration in AI development. However, democratizing AI development requires more than accessible tools alone and depends on well-managed communication and engagement processes. The study contributes to a deeper understanding of AutoML's role as a boundary object and highlights the need for further research into its responsible and inclusive deployment.</p>","PeriodicalId":33315,"journal":{"name":"Frontiers in Artificial Intelligence","volume":"9 ","pages":"1867883"},"PeriodicalIF":6.7,"publicationDate":"2026-10-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13678699/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149224997","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Optimizing multi-class intrusion detection in IIoT: advanced feature selection using grey wolf optimizer and machine learning classifiers.","authors":"Ibrahim Al-Shibly, Llorenç Burgas, Joaquim Massana","doi":"10.3389/frai.2026.1932015","DOIUrl":"10.3389/frai.2026.1932015","url":null,"abstract":"<p><strong>Introduction: </strong>The rapid expansion of the Industrial Internet of Things (IIoT) has increased the cyberattack surface of industrial infrastructures, creating demand for intrusion detection systems (IDS) that combine predictive performance with deployment efficiency.</p><p><strong>Methods: </strong>This study proposes a Grey Wolf Optimizer (GWO)-guided fixed top-k feature-ranking framework for multi-class IIoT intrusion detection using candidate-dependent redundancy regularization, class-weighted learning, stratified cross-validation, repeated runs, and independent holdout evaluation. The framework was evaluated on the CIC IIoT Dataset 2025 and independently assessed on Edge-IIoTset as a secondary benchmark across six machine learning classifiers and several feature-selection baselines.</p><p><strong>Results: </strong>On CIC IIoT Dataset 2025, Random Forest and Decision Tree achieved holdout Macro-F1 scores of 0.9389 and 0.9330, respectively, using only 10 selected features, representing an 85.92% reduction from the 71-feature model-input space. Repeated training-only comparisons showed significant overall differences among GWO, BA, LightGBM, and Boruta (Friedman <i>χ</i>² = 16.920, <i>p</i> = 0.000734; Kendall's <i>W</i> = 0.564); GWO outperformed BA and LightGBM after Holm correction but did not differ significantly from Boruta. Ablation analysis further showed lower feature redundancy under adaptive regularization, although its predictive advantage was not statistically significant. Results on Edge-IIoTset showed that feature-selection effectiveness is dataset dependent.</p><p><strong>Discussion: </strong>Overall, the proposed framework provides a practical balance between predictive performance, feature compactness, and computational efficiency for resource-constrained IIoT intrusion detection.</p>","PeriodicalId":33315,"journal":{"name":"Frontiers in Artificial Intelligence","volume":"9 ","pages":"1932015"},"PeriodicalIF":6.7,"publicationDate":"2026-10-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13677896/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149224974","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Semi-supervised Bayesian deep clustering with uncertainty-guided representation and robustness.","authors":"S Krishnamoorthy, B Jaganathan","doi":"10.3389/frai.2026.1935933","DOIUrl":"10.3389/frai.2026.1935933","url":null,"abstract":"<p><strong>Introduction: </strong>In real-world data, tackling missing data, noisy observations, and uncertainty at the same time is hard since the number of labelled data is limited. Current deep clustering algorithms have considered the imputation, noise, semi-supervision, and uncertainty problems separately, such as without modelling the uncertainty in the noise, or without modelling the missingness in the missing data. There is still a need to develop an end-to-end deep clustering algorithm that incorporates all these four factors.</p><p><strong>Methods: </strong>To overcome the limitations, we propose to robust the semi-supervised Bayesian Deep Clustering (RBDC), which involves four parts, namely: The Student-t decoder does not explicitly impute for missing features and yields a robust reconstruction for heavy-tailed noise. Label-guided Gaussian mixture priors provide partial class label guidance in the learning process within the latent space, which is positioned between supervised and unsupervised learning. An uncertainty-weighted Evidence Lower Bound (ELBO) scales the contribution of each sample by the posterior uncertainty, effectively being noisy or contaminated. Sytems with Hybrid EM + SGD optimisation strategy. The mixture is trained using the hybrid EM + SGD optimisation strategies. The latent uncertainty is used in the mixture updates. The framework is tested on University of California, Irvine (UCI) Adult Income and Gym Members Fitness datasets, and the results are compared against strong deep clustering baselines.</p><p><strong>Results: </strong>Results show that RBDC outperforms baseline methods in calibration, discrimination/predictive ability and clustering quality in the low supervision regime. Ablation studies suggest the benefit of label informed priors and uncertainty weighting in enabling stable and interpretable clustering. Results also highlight the complementary benefits of semi-supervision and modelling of uncertainty.</p><p><strong>Discussion: </strong>Our results show that RBDC offers a strong and label-efficient calibration-aware framework to clustering of noisy, incomplete and partially labelled tabular data. RBDC attempts to model in a single generative framework the missingness, noise, uncertainty and weak supervision in the data, with a possible benefit to downstream applications in decision support such as risk stratification, anomaly detection and predictive modelling. Hence RBDC provides a conceptually integrated approach to semi-supervised deep clustering for challenging real-world datasets.</p>","PeriodicalId":33315,"journal":{"name":"Frontiers in Artificial Intelligence","volume":"9 ","pages":"1935933"},"PeriodicalIF":6.7,"publicationDate":"2026-10-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13677874/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149224987","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"A feature-optimized EEG framework using multiscale entropy and metaheuristic optimization for cognitive analysis.","authors":"Keerthika N, Kiruthika V","doi":"10.3389/frai.2026.1888234","DOIUrl":"10.3389/frai.2026.1888234","url":null,"abstract":"<p><strong>Background: </strong>Cognitive activity plays an important role by influencing working-memory, problem-solving, attention, response and decision making. EEG serves as a non-invasive approach to analyze neural activity associated with cognitive processing. The current study aims to identify the more contributing features to analyze the cognitive activity between sports and non-sports groups.</p><p><strong>Methods: </strong>The present study included 100 participants, consisting of 80 males and 20 females aged between 19 and 23 years. EEG signals were acquired and preprocessed using filtering methods to remove unwanted noise and artifacts. The four-entropy metrics dispersion entropy, permutation entropy, sample entropy, and fuzzy entropy were applied to the coarse-graining method across multiple temporal scales to analyze the complexity of EEG signals. To identify the most informative feature subsets and reduce redundancy, filter-based methods such as Mutual Information, Pearson Correlation, Minimum Redundancy Maximum Relevance and wrapper-based methods such as Genetic Algorithm and Particle Swarm Optimization were applied. Finally for differentiating the cognitive activity, five machine learning classifiers that include Support Vector Machine, Artificial Neural Network, <i>K</i>-nearest neighbor, Decision Tree and Random Forest were applied. Statistical methods such as <i>T</i>-test and Fleiss kappa were used to analyse the performance efficiency.</p><p><strong>Results: </strong>Feature extraction phase yielded a total of 160 multiscale entropy features. The entire features when fed into both filter-based and wrapper-based methods, Genetic Algorithm based optimization yielded a minimal contributing feature subset with high nested cross-validated accuracy of 93.0 ± 2.7% with Support Vector Machine. This contributing feature subset was used as input for five different machine learning classifiers for differentiating the cognitive activity between both groups. The five classifiers demonstrated high cross-validated agreement on held-out subjects, with a mean Fleiss' kappa of 0.874 ± 0.051 (95% CI: 0.811-0.937), depicting a good inter-classifier agreement for classification of samples into sports and non-sports individuals.</p><p><strong>Discussion: </strong>The study demonstrates that multiscale entropy features combined with metaheuristic-based feature optimization can effectively analyze the cognitive activity between sports and non-sports individuals. The proposed EEG framework provides a structured approach for real-time cognitive activity analysis using compact and informative entropy-based feature subsets.</p>","PeriodicalId":33315,"journal":{"name":"Frontiers in Artificial Intelligence","volume":"9 ","pages":"1888234"},"PeriodicalIF":6.7,"publicationDate":"2026-10-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13677936/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149225012","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Temporal phase coverage and patient cohort size in deep learning-based cardiac cine computed tomography angiography segmentation.","authors":"Andrew Heller, Li-Yueh Hsu, An-Yu Sun, Matthew Jacobs, Lin-Ching Chang, Marcus Y Chen","doi":"10.3389/frai.2026.1914255","DOIUrl":"10.3389/frai.2026.1914255","url":null,"abstract":"<p><strong>Background: </strong>Annotating all phases of cine computed tomography angiography (CTA) is burdensome. We evaluated how temporal phase coverage and patient cohort size affect deep learning-based cardiac segmentation.</p><p><strong>Methods: </strong>Twelve 3D U-Net models were trained using a 110-patient single-phase CTA dataset or a 20-patient, 20-phase cine CTA dataset. Phase-coverage ablations used one, two, or 20 phases; patient-count ablations were performed within both datasets. One cine patient was used for validation and four independent patients were used for testing. Five cardiac structures were evaluated using Dice and the 95th-percentile Hausdorff distance (HD95). Phase-wise and patient-level Dice consistency and structure-specific volumetric variability within and between subjects were assessed.</p><p><strong>Results: </strong>Full cine training produced higher overall Dice (0.89 vs. 0.87) and lower overall HD95 (4.87 vs. 9.78 mm) than the single-phase reference (both <i>p</i> < 0.05). Right ventricular (RV) Dice was higher (0.91 vs. 0.84) and HD95 was lower (5.16 vs. 22.75 mm; both <i>p</i> < 0.05). The two-phase model achieved an overall Dice of 0.88 and an overall HD95 of 7.85 mm; neither metric differed significantly from the single-phase reference. HD95 generally increased with smaller cohorts, and both metrics deteriorated markedly in the five-patient configurations. Full cine training showed flatter phase-wise Dice profiles and lower across-phase variability. Between-subject volume differences exceeded phase-related within-subject differences for all structures (all <i>p</i> < 0.05).</p><p><strong>Conclusion: </strong>In this cross-cohort comparison, full cine training showed higher overall Dice, lower overall HD95, and greater phase-wise consistency than single-phase training. The evaluated two-phase configuration may inform annotation-efficient cine CTA dataset design, but broader validation is needed.</p>","PeriodicalId":33315,"journal":{"name":"Frontiers in Artificial Intelligence","volume":"9 ","pages":"1914255"},"PeriodicalIF":6.7,"publicationDate":"2026-10-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13677920/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149224989","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Hybrid semantic-lexical framework for ambiguity detection in software requirements using TF-IDF, character n-grams, and SBERT.","authors":"Fariha Khalid, Muhammad Yaseen, Muhammad Awais Sattar, Abdelhamid Ghoul, Aida Mustapha","doi":"10.3389/frai.2026.1901190","DOIUrl":"10.3389/frai.2026.1901190","url":null,"abstract":"<p><p>Ambiguity in software requirements is a major cause of misinterpretation, design defects, rework, and project delays in software engineering. This study proposes a hybrid semantic-lexical ambiguity detection framework that integrates lexical, structural, and contextual representations for identifying ambiguous software requirement statements. Two benchmark datasets were employed, including a Software Requirements Specification (SRS) dataset and a Functional-Non-Functional Requirements (FR-NFR) dataset. Data pre-processing involved tokenisation, lemmatisation, and noise removal, while Synthetic Minority Oversampling Technique (SMOTE) was applied to address class imbalance within training partitions. Three modeling strategies were evaluated: baseline classifiers using TF-IDF and character n-gram features, semantic models based on Sentence-BERT (SBERT) embeddings, and a hybrid representation combining TF-IDF, character n-grams, and SBERT embeddings. Experimental results indicate that baseline models achieved high precision but comparatively lower recall, limiting their ability to identify ambiguous requirements under class imbalance. The proposed hybrid representation achieved the strongest and most balanced performance across both datasets in practical terms, particularly when combined with Support Vector Machine classification, resulting in improved recall and F1-score, although statistical testing indicated that these gains over the strongest lexical baseline were not significant at the 5% level. By jointly modeling lexical vagueness, structural ambiguity patterns, and contextual semantics, the framework captures both explicit and implicit forms of ambiguity that are frequently overlooked by single-representation approaches. To strengthen experimental validity, five-fold stratified cross-validation, transformer-based baseline comparison using fine-tuned BERT, statistical significance testing, and annotation reliability assessment using Cohen's Kappa coefficient were incorporated. The findings demonstrate that integrating lexical, structural, and semantic representations provides a robust and potentially scalable approach for ambiguity detection in software requirements engineering. The proposed framework contributes to improved requirements quality, reduced risk of misinterpretation, and more reliable automated support for requirements analysis.</p>","PeriodicalId":33315,"journal":{"name":"Frontiers in Artificial Intelligence","volume":"9 ","pages":"1901190"},"PeriodicalIF":6.7,"publicationDate":"2026-10-05","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC13677944/pdf/","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149224957","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}