Search Results (4,504)

Named entity recognition (NER) is a key technique for extracting entities such as equipment, components and defect types from GIS defect texts, providing a basis for subsequent knowledge graph construction. However, GIS defect texts contain many engineering structures, including engineering abbreviations, equipment numbers and phase identifiers, making it difficult for general-purpose models to stably recognize their semantic associations and entity boundaries. To address this problem, this paper proposes an engineering format-aware masking strategy. The strategy identifies candidate fragments using format rules for phase identifiers, measurement value-unit patterns and engineering abbreviations and preferentially selects them as perturbation targets to strengthen the model’s understanding of engineering structures and their contextual relationships. Bidirectional long short-term memory is used to extract bidirectional sequence features, and a conditional random field is used to model transition constraints between labels and obtain the globally optimal label sequence. The results show that the proposed model achieves precision, recall and F₁ scores of 0.89, 0.92 and 0.90, respectively. The analysis indicates that the proposed method improves entity recognition for phase-related structures, engineering abbreviations and equipment hierarchy fragments. Full article

Cloud-based large language model (LLM) services are increasingly used to process natural-language queries that may contain private or sensitive information. Conventional privacy-preserving approaches, such as cryptographic protection and text sanitization, often introduce substantial computational overhead or disrupt the semantic integrity of the original query, resulting in a trade-off between privacy protection and task utility. To address this limitation, we propose generative semantic chaffing (GS-Chaff), a training-free multi-agent framework for privacy-preserving LLM inference over natural-language text queries. Rather than explicitly masking sensitive content, GS-Chaff hides the user’s true intent among semantically plausible chaff queries. The framework is implemented through two small language model (SLM)-based agents: a privacy policy agent that adaptively determines the required semantic abstraction level and chaffing factor for each text query, and a generative semantic chaffing agent that produces semantically aligned dummy queries. After cloud-side inference, the response corresponding to the protected real query is recovered locally using a stateless index, without modifying the cloud-based LLM. Experimental results on text-based benchmarks demonstrate that GS-Chaff reduces the attacker’s real-query identification rate to 22.5%, close to random guessing, while maintaining inference utility on the evaluated benchmarks. In addition, GS-Chaff reduces local preprocessing time by $1.85\times$ compared with a fixed chaffing configuration using $\beta =5$. Full article

Detecting disinformation in morphologically rich and under-resourced languages remains a significant challenge in natural language processing. This paper examines the role of lemmatization as a preprocessing strategy for disinformation detection in Croatian. Building on previous work and extending the evaluation to a newly collected dataset of nearly 25,000 social media comments, we systematically compare traditional machine learning classifiers (SVM, Random Forest, and neural networks) and a transformer-based model (croBERT) on both original and lemmatized text. Our findings demonstrate that lemmatization does not produce uniform gains across architectures: while linear models and croBERT display small but measurable improvements from morphological normalization, non-linear models such as RBF SVM and neural networks experience substantial declines in performance. These results indicate that lemmatization interacts differently with model inductive biases and feature extraction mechanisms. Overall, the study provides a detailed empirical assessment of preprocessing choices for low-resource, morphologically complex languages and offers practical guidance for developing disinformation detection systems in Croatian and similar contexts. Full article

Background: Artificial intelligence (AI) is one of the most influential technological innovations in contemporary mental healthcare. Advances in machine learning, natural language processing, conversational agents, and large language models have accelerated the integration of AI into clinical practice, professional education, and healthcare. Despite its increasing adoption, important questions remain regarding its clinical effectiveness, implementation, safety, and ethical governance. Objective: This systematic review aimed to synthesize the current evidence on the application of artificial intelligence in mental health, with particular emphasis on clinical practice, educational transformation, and ethical governance. Methods: This systematic review was conducted in accordance with the PRISMA 2020 guidelines. PubMed/MEDLINE, Scopus, Web of Science, PsycINFO, and Google Scholar were systematically searched. The electronic database search was last conducted on 31 December 2025, and studies published between January 2019 and December 2025 were considered eligible. Eligible studies examined the application of artificial intelligence in mental health across clinical practice, educational contexts, and ethical governance. Study selection, data extraction, and methodological quality assessment were carried out independently by two reviewers using predefined eligibility criteria and standardized extraction forms. Owing to substantial methodological heterogeneity across the included studies, the findings were synthesized narratively. Results: A total of 88 studies met the eligibility criteria and were included in the final qualitative synthesis. The findings showed that AI demonstrated potential to improve diagnostic support, risk prediction, treatment planning, symptom monitoring, and access to psychological support. AI also supported educational innovation and workforce development while highlighting the importance of ethical governance for responsible implementation in mental healthcare. Conclusions: Future progress will depend on interdisciplinary collaboration to ensure that AI complements rather than replaces human expertise. Although AI demonstrates substantial potential, many systems remain experimental, with limited external validation. Prospective multicenter evaluation, transparent algorithm development, and robust ethical governance are therefore essential before widespread clinical implementation. Full article

With the rapid evolution of Large Language Models (LLMs) in natural language understanding and generation, Relation Extraction (RE) has achieved substantial milestones in low-resource and open-domain scenarios. However, prevailing LLM-based RE methodologies predominantly rely on one-shot prompting or static workflows, which lack autonomous evaluation and iterative optimization mechanisms. Consequently, these approaches are prone to issues such as missing relations, type confusion, and factual hallucinations when navigating complex relational contexts. To address these limitations, this paper proposes DyFIRER (Dynamic Feedback Iterative Relation Extraction Framework), a multi-agent framework characterized by dynamic feedback. By constructing three functionally complementary agents—Extraction, Verification, and Optimization—the framework models the RE task as a closed-loop iterative process consisting of “extraction-verification-feedback-optimization,” thereby enabling dynamic adjustment and continuous refinement of extraction strategies. Experimental results on the DuIE 2.0 open relation extraction extension subset demonstrate that DyFIRER achieves an F1-score of 80.5%, modestly but statistically significantly outperforms GPT-4 (p = 0.014), a result that holds on both the augmented and non-augmented test sets and mainstream static methods (yielding a 10.3% improvement over Qwen-7B). Ablation studies further substantiate the critical role of the dynamic feedback iterative mechanism and the strategy retrieval module in mitigating complex relation omissions and factual hallucinations. The framework requires no additional annotated data or fine-tuning, suggesting potential applicability to low-resource settings, though this was not directly evaluated in the current study. Full article

Against the backdrop of high-density Asian cities shifting from production- to consumption-oriented spaces, commercial pedestrian streets are key to urban public life and vitality. This study selects six major commercial pedestrian streets in Chongqing and employs natural language processing (NLP) and importance–performance analysis (IPA) methods to construct a four-dimensional evaluation framework (spatial, commercial, cultural, location/facility). It analyzes public perception and experience based on user-generated content (UGC). Findings show that: (1) Significant differences across dimensions form three development types: cultural identity, functional hub, and distinctive growth, reflecting structural bottlenecks in transitioning from single- to multi-functional spaces. (2) IPA identifies “business formats,” “cultural activities,” and “consumption experience” as priorities for improvement, while “commercial atmosphere” and “transportation conditions” are current strengths to maintain. (3) Sentiment analysis reveals that negative perceptions focus on basic functions and sense of place, whereas positive sentiments relate to cultural expression and spatial esthetics, highlighting the role of cultural soft power and visual design in street appeal. This study reveals public perception patterns via big data analysis, offering empirical support for the refined renewal, cultural preservation, and sustainable management of commercial pedestrian streets in high-density Asian cities. Full article

Large Language Models (LLMs) have been rapidly evolving lately, resulting in the need for strong, explainable models to detect the difference between human-generated and machine-generated articles. Existing approaches which are mostly based on fine-tuned transformers suffer from several drawbacks such as rapid obsolescence, paraphrasing attacks, and lack of interpretability. To improve their ability to detect, this paper proposes a novel paradigm called Human vs. LLM Identification (HLI) which introduces a Retrieval-Augmented Generation (RAG)-inspired evidence-based detection strategy alongside a fine-tuned transformer classifier. Our core model, DeBERTa-Sentinel, is built on top of a fine-tuned Microsoft DeBERTa-v3-small model, which uses a disentangled attention mechanism to better capture subtle syntactic and stylistic deviations characteristic of AI-generated text. We evaluate our framework on a balanced dataset of 43,456 text samples, curated from the OpenGPTText corpus and covering AI-generated and human-authored content across diverse domains including news, education, and creative text. The experimental results show improved performance over the selected baselines, with our framework achieving an accuracy of 97.53%, precision of 95.89%, recall of 99.34%, and ROC-AUC of 99.53%. In addition, explainability is integrated into our framework through Local Interpretable Model-agnostic Explanations (LIME) analysis, providing token-level insight into classification decisions. This study establishes a benchmark for scalable, explainable AI text detection, with implications for academic integrity, content moderation, and combating misinformation. Full article

This study aims to identify the key destination-related factors associated with tourists’ revisit behavior in peri-urban rural areas within Chinese metropolitan regions in the context of short-video marketing, and to reveal the structural relationships and hierarchical characteristics among these factors. First, the LDA topic model was employed to conduct text mining on authentic tourist-generated comments posted on Douyin, through which the core factors related to revisit behavior were identified and conceptually standardized based on tourists’ expressions. Building on this process, 153 experts in relevant fields were invited to evaluate the direction and strength of the relationships among these factors. An integrated DEMATEL–ISM–MICMAC approach was then applied to analyze their causal attributes, hierarchical structure, and systemic roles. The results indicate that the identified factors do not operate independently but instead form a multilayered structure with clear hierarchical characteristics. Among them, rural visual imagery, escape-oriented experience, rural lifestyle experience, and rural industry integration occupy deeper structural levels and exert relatively strong structural influences on factors located at intermediate and surface levels. The findings further suggest that the sustained attractiveness of rural tourism destinations in metropolitan regions cannot rely solely on short-video exposure or isolated “internet-famous” attractions; rather, it requires coordinated alignment among digital communication content, rural industries, lifestyle experiences, and tourism supply. By integrating tourist-generated content, natural language processing, and expert-based structural assessment, this study extends research on short-video tourism marketing from a systems perspective and provides practical insights for peri-urban rural destinations in Chinese metropolitan regions seeking to optimize the structural configuration of tourism resources, products and services, and marketing communication. Full article

Social media has become an important tool for understanding visitor experiences in museums and cultural heritage sites. This literature review identifies, organizes, and thematically synthesizes existing studies on museum visitor experience based on social media data. It examines 41 studies retrieved from Scopus and Web of Science and published between 2017 and 2026. Furthermore, the review examines the selected studies across six dimensions: social media platforms, types of user-generated data, the role of digital interactions, the museums and cultural heritage sites studied, the analytical methodologies applied, and the main findings on visitor experience. The findings indicate that TripAdvisor is the most frequently used platform for collecting textual reviews and star ratings, whereas Instagram and Flickr are mainly used for visual and spatial data. Most studies rely on computational methods, often combined with quantitative techniques, while qualitative approaches are used less frequently. The identified methods include content analysis, statistical analysis, sentiment analysis, topic modeling, machine learning, image analysis, and spatial analysis. Across the reviewed studies, visitor experience is examined as a multidimensional phenomenon encompassing emotions, service quality, authenticity, historical connection, aesthetics, education, and social participation. Finally, the review identifies recurring themes across the dimensions and synthesizes them into broader research streams. These are brought together in an integrative synthesis framework that organizes existing research, highlights research gaps, and outlines directions for future studies. Full article

Distribution drift can substantially degrade the performance of deployed machine learning models; for example, accuracy on SST-2 can fall from $88\%$ to $58\%$. Detecting such degradation is fundamentally challenging because deployment provides inputs but not labels, so the detection itself must be reference-free. We introduce DecayBench, the first reference-free, calibrated benchmark for evaluating drift detectors. DecayBench measures detector trustworthiness along five axes (calibrated, valid, timely, no-regret, adaptive), and compares ten existing detectors across ten NLP, vision, and multimodal datasets using paired-bootstrap significance testing. Evaluation on DecayBench shows that no existing detector is uniformly optimal. Motivated by this observation, we propose Alert, a label-free aggregation rule for drift detection. Unlike all competing combiners, it uses a label-free self-configuring selection rule with a no-regret guarantee. Alert has three contributions: (i) a dilution analysis yielding a self-configuring detector selection rule; (ii) a finite-sample conformal guarantee that controls the false-alarm probability on clean data at any prescribed level (e.g., $5\%$) for arbitrary score distributions; and (iii) a no-regret result: when no single detector dominates (constituents of comparable effect size, a condition checkable offline), Alert matches or beats the best constituent, being never significantly worse and sometimes better by a large margin; this holds across NLP, NLI, and vision (ResNet), with the largest gains under multimodal drift, and the proof identifies a dominant single detector (MMD on CLIP) as the only dilution exception. We prove the no-regret property and, across the benchmark, report its empirical counterpart, non-dominance under a paired bootstrap (Alert is never significantly worse than the best constituent), which at some operating points is statistically inconclusive rather than a strict win. Because Alert combines only embedding- and logit-based detector scores, it directly transfers across NLP, vision, and multimodal models. Empirically, Alert strictly improves over single-modality monitoring, increasing AUC by up to 25 points under mixed-modality drift and by approximately 50 points under cross-modal mismatch, where individual modality-specific detectors perform near chance. Alert also matches or outperforms the Fisher, Simes, Bonferroni, and median combiners, performs best under low-severity drift, and matches or surpasses early fusion (Concat-MMD) in both multimodal settings. Full article

The rapid growth of organisational textual data has increased the value of Natural Language Processing (NLP) and semantic analytics for strategic decision support. However, many employees still lack the knowledge and skills needed to evaluate AI-generated information critically. This study develops an explainable Management Information Systems (MIS) framework that integrates NLP-based semantic analytics with PLS-SEM to examine the relationship between AI illiteracy and strategic decision quality. A convergent mixed-methods design with sequential analytical integration was used with a valid sample of 200 knowledge workers from public organisations in Saudi Arabia across six industries. The sample included employees from Saudi Arabia, Egypt, Jordan, Sudan, Syria, India, and the Philippines. Quantitative data were analysed using PLS-SEM, while textual data were analysed using Sentence-BERT, BERTopic, semantic network analysis, and Aspect-Based Sentiment Analysis. The results showed that higher AI illiteracy was negatively associated with strategic decision quality and positively associated with automation bias, uncritical trust in AI, and cognitive offloading. Digital proficiency and AI governance awareness weakened the negative association between AI illiteracy and decision quality, while functional-background differences were examined through multigroup analysis. The semantic analysis identified six themes: AI competency, decision trust, AI governance, decision support, organisational learning, and risk awareness. Sentiment analysis showed positive views of productivity and decision support, together with concerns about algorithmic bias, explainability, transparency, and AI governance. The study contributes an integrated human–AI decision vulnerability framework in which semantic evidence complements structural modelling and provides a clearer understanding of AI-related competency, reliance, governance, and decision-support issues. Full article

Background/Objectives: Free-text veterinary pathology diagnoses contain essential information for cancer registration but are difficult to convert into standardized ontology-based codes because of linguistic variability, contextual modifiers, and large ontology search spaces. This study evaluated a hybrid lexical–semantic architecture for the automated assignment of Vet-ICD-O-Canine-1 morphology codes. Methods: A retrospective single-registry benchmark included 211 diagnoses from the São Paulo Animal Cancer Registry. Of these, 190 contained sufficient morphological information for expert-reviewed reference coding, whereas 21 generic or insufficiently specified descriptions were retained as an exploratory challenge subset. Fuzzy lexical matching retrieved Top-10, Top-20, or Top-30 candidates from the complete 971-entry morphology ontology, followed by semantic selection using Claude Haiku 4.5 and structured JSON output. Performance and computational efficiency were compared to direct full-ontology inference. Results: Among the evaluated fuzzy metrics, token_set_ratio achieved the highest Top-30 reference-code retrieval rate of 89.5%. End-to-end exact-match agreement increased from 73.7% with Top-10 to 79.5% with Top-20 and 85.8% with Top-30 (95% CI, 80.1–90.0%). Top-30 generated non-null codes for 93.2% of the 190 evaluable diagnoses and achieved a conditional exact-match agreement of 92.1%. By contrast, the direct full-ontology baseline achieved 71.2% conditional exact-match agreement (42/59) among non-null predictions and 22.1% end-to-end exact-match agreement (42/190) when incorrect predictions, null outputs, and technical failures were considered non-concordant outcomes. Compared to direct full-ontology inference, Top-30 reduced input-token consumption by 92.4%, total token consumption by 92.2%, and inference cost by 91.3%, while avoiding the 118 API rate-limit failures observed with the direct baseline. Among the 21 insufficiently specified diagnoses, Top-30 returned null codes in 38.1% and non-null codes in 61.9%. Conclusions: Ontology-guided candidate reduction improved coding agreement, computational efficiency, and operational robustness within this retrospective single-registry benchmark. However, the reported performance estimates require confirmation in larger independent datasets, and an upstream data-sufficiency or abstention mechanism is needed before prospective operational deployment. Full article

Gender bias in generative artificial intelligence (GenAI) is both a technical and a social phenomenon: it emerges from historically patterned data, model design, and interactions in institutional use, and it cannot be understood by engineering or by social critique alone. This critical integrative review develops a more differentiated account. It connects feminist epistemology, Science and Technology Studies, critical AI scholarship, natural language processing, and governance research to examine five levels: historical knowledge production, technical representation and generation, benchmark evaluation, institutional deployment, and accountability. The review explains tokenization, next-token prediction, transformers, and the transition from static embeddings to contemporary language models before assessing evidence from standard fairness tests—coreference tests (WinoBias), sentence-pair tests (CrowS-Pairs), and stereotype tests (StereoSet)—as well as open-ended generation, multilingual testing, and text-to-image systems. It shows that measured bias varies with task, prompt, language, model version, and metric. What a test records and what that record means are therefore distinct questions: measurements are situated and depend on the instrument, and their interpretation draws on theory rather than following from the numbers alone. Evidence from employment, education, healthcare, and translation further indicates that the relevant unit of analysis is the model-in-context—the model together with the institution and workflow in which its outputs are used. Technical mitigation can reduce specific harms but does not repair unequal criteria, incomplete evidence bases, or weak institutional accountability. The review proposes a multilevel governance approach combining technical evaluation, documentation, professional and community oversight, appeals, remedies, and public-interest knowledge infrastructure. Its distinctive contribution is to connect three observations usually kept apart—how bias is measured, how generative systems concentrate epistemic authority, and how statistical learning is oriented toward past data—and to show why democratic and feminist governance can keep alternative technological futures open. Full article

Earth Observation (EO) analysis increasingly relies on large and heterogeneous satellite datasets, yet developing EO workflows often requires specialized expertise in data selection, geospatial programming, and cloud-based processing. Recent advances in Large Language Models (LLMs) offer new opportunities for natural-language interaction with EO systems, although challenges related to transparency, reproducibility, and domain-specific reasoning remain. This study presents SeaScope, an explainable AI framework that integrates LLMs, Retrieval-Augmented Generation (RAG), scientific knowledge retrieval, and Google Earth Engine (GEE) to transform natural-language requests into transparent and executable EO workflows. The framework combines knowledge retrieval, code generation, cloud execution, provenance tracking, and interactive visualization within a unified environment. A pilot implementation is demonstrated through maritime and coastal monitoring applications, including oil spill detection, vessel monitoring, water quality assessment, floating debris detection, and air quality analysis. Multiple state-of-the-art LLMs are evaluated under both RAG and non-RAG configurations using representative EO case studies. The results indicate substantial differences among model families and show that retrieval augmentation can significantly improve workflow generation quality and reliability for capable models, while providing more limited benefits for smaller models. The proposed framework demonstrates the potential of explainable AI agents to support transparent, reproducible, and scalable EO analysis. Full article

Retrieval-augmented fine-tuning (RAFT) can improve small language models (sLMs) on retrieval-grounded question answering, but the synthetic training data produced by commercial large language models (LLMs) vary in quality. This paper contributes a quality-aware selection protocol—rather than a new RAFT or QLoRA method—that scores LLM-generated alternatives along four embedding-based dimensions (question relevance, answer faithfulness, QA coherence, and semantic similarity) and selects one alternative per task before parameter-efficient fine-tuning. Under pre-specified paired-bootstrap contrasts with Holm correction, the parameter-free faithfulness-based selector only-AF significantly exceeds random selection on Gemma-2-9B-IT (ΔF1 = +0.106, 95% CI [+0.043, +0.174], Holm-corrected p = 0.019), and its pre-specified weighted companion af-70 (w_AF = 0.70) shows the same confirmed pattern (Holm-corrected p = 0.002). Both effects persist under a Korean character-level F1 that removes particles and punctuation (Holm-corrected p = 0.004 and p = 0.042), indicating robustness to the choice of lexical metric. Relative to training on the full 150-row augmented pool, the quality-selected 50-row sets are statistically indistinguishable while using one third of the training data, which we interpret as data efficiency rather than superiority. Across six instruction-tuned models (2B–27B), a significant selector-by-model interaction indicates that the optimal quality axis is model-dependent, and the two smallest models show no benefit from selection. The study’s confirmatory contrasts use a small controlled Korean corpus under a transductive design; two pre-registered validation experiments probe external validity. On an independent five-fold larger corpus with a passage-level train/test split, fine-tuning transfers strongly and the selected one-third subsets show no significant difference from the full pool, while the advantage over random selection is directionally positive but small and not significant; under controlled corruption of 35% of the pool, the metrics detect the damaged rows, and for the score-sum selector the selection-versus-random benefit is significantly larger than on the clean pool (difference-in-differences p = 0.0014; directionally consistent but not significant for the faithfulness selectors). Within this scope, quality-aware selection is a promising, data-efficient safeguard for synthetic RAFT data—performing comparably to full-pool training at one third of the cost, with growing value as pool quality degrades—and larger-scale external validation remains future work. Full article