domain expertise

Addressing the longstanding challenge of jointly optimizing accuracy, interpretability, and computational efficiency in medical device regulatory classification, this study introduces the first multi-model AI assessment framework tailored for regulatory compliance decisions. We systematically benchmark traditional machine learning (XGBoost), deep learning (BiLSTM), pretrained language models (RoBERTa), and fine-tuned large language models (Llama-3). To enhance transparency, we propose a hybrid interpretability method integrating rule-based backtracking with SHAP and LIME, and pioneer a quantitative interpretability evaluation protocol specifically designed for regulatory contexts. Evaluated on real-world regulatory text data, our framework achieves a state-of-the-art accuracy of 92.3%. Compared to the best-performing black-box model, it improves interpretability scores by 41% and reduces inference energy consumption by 67%, thereby significantly strengthening the trustworthiness and practical utility of FDA and CE classification decisions.

This study addresses the challenge of limited interpretability in machine learning models within manufacturing contexts, where opaque predictions often hinder effective decision-making. To bridge this gap, the authors propose a novel paradigm that leverages large language models to dynamically retrieve relevant triples from domain-specific knowledge graphs, thereby structurally linking expert knowledge with model predictions and generating user-friendly natural language explanations. Integrating knowledge graphs, large language models, and explainable artificial intelligence (XAI), the approach was evaluated on 33 manufacturing-related tasks. Results demonstrate superior performance across both quantitative metrics—such as accuracy and consistency—and qualitative dimensions, including clarity and practical utility, significantly enhancing model interpretability and decision support capabilities in real-world industrial settings.

This study addresses critical challenges in multilingual NLP—model bias, insufficient robustness, and difficulty in ethical alignment—by proposing a fine-tuning and deployment framework for large language models (LLMs) targeting low bias and high robustness. Methodologically, it integrates the Hugging Face ecosystem with Transformer architectures, incorporating multilingual tokenization, domain-aware data cleaning and augmentation, and a progressive fine-tuning strategy that jointly optimizes fairness and task performance. Contributions include: (1) a lightweight, cross-lingual fine-tuning paradigm resilient to bias-induced interference; (2) empirical validation across high-stakes domains (e.g., healthcare and finance), demonstrating significant improvements in generalization and fairness for classification and named entity recognition; and (3) an interpretable, auditable, and production-ready LLM deployment pipeline that advances the practical implementation of ethically aligned AI.

Deep learning (DL) faces an “expectation–reality gap” in business analytics: it delivers no substantial performance gains over traditional machine learning (e.g., random forests, XGBoost) on structured data (fixed-length feature vectors), and its adoption is hindered by five key barriers—high computational complexity, poor model interpretability, lack of scalable big-data infrastructure, shortage of domain-specialized AI talent, and insufficient executive sponsorship. Method: This study conducts the first systematic, empirical evaluation of DL’s effectiveness and interpretability on standard business-oriented structured datasets, integrating content analysis with multi-dimensional comparative experiments against established ML baselines. Contribution/Results: We demonstrate that DL is not a universal replacement for traditional ML but rather a complementary enhancement tool. The work proposes a pragmatic, enterprise-focused AI adoption framework for model selection—bridging the cognitive gap between theoretical hype and industrial practice—and advances rational, sustainable AI deployment in commercial settings.

General-purpose large language models (LLMs) lack domain-specific expertise in composite materials processing and manufacturing equipment operation. Method: We developed two vertical-domain LLM systems integrating GPT-4 architecture with industry-specific knowledge bases, employing domain-adaptive fine-tuning and retrieval-augmented generation (RAG). Automated evaluation used ROUGE and BERTScore; human evaluation involved domain experts. Contribution/Results: Our models match or exceed GPT-4o on automated metrics; expert feedback confirms significantly improved answer specificity, depth, and technical query responsiveness. This work presents the first structured modeling of end-to-end operational knowledge in composite manufacturing and its specialized LLM deployment—establishing a reusable technical pathway and empirical benchmark for industrial AI adoption.

The deployment of large language models (LLMs) in high-stakes domains—such as law, healthcare, and finance—introduces underexplored compliance risks, including sensitive information leakage, intellectual property infringement, and uncontrolled outputs; existing NLP tools lack domain-specific compliance adaptation. Method: Through semi-structured interviews and qualitative analysis with frontline domain experts, we systematically identify real-world risk perceptions and emergent mitigation strategies, uncovering a structural misalignment between current tools and human-centered compliance requirements. Contribution/Results: We propose a human-centered RegTech compliance design framework for LLM-based NLP systems, centered on three core mechanisms: sensitive data protection, intellectual property security, and output quality controllability. This framework provides empirically grounded, actionable design principles to support the development of compliance-embedded NLP systems.

This paper addresses the lack of standardized evaluation criteria for machine learning algorithm selection in healthcare, telecommunications, and marketing. We propose a multidimensional, automated model selection framework that jointly optimizes predictive performance—measured by accuracy, precision, and recall—and model complexity, as quantified by the Akaike Information Criterion (AIC). The framework is designed to be domain-agnostic, supporting seamless adaptation across eager, lazy, and hybrid learning paradigms. Evaluated on eight real-world datasets—including cardiovascular disease prediction and fetal health classification—the method consistently identifies optimal models, achieving statistically significant improvements in both predictive accuracy and generalization performance. Crucially, it delivers interpretable and reusable model recommendations tailored to mission-critical applications, thereby bridging the gap between theoretical model selection and practical deployment.

Traditional machine learning metrics and generic benchmarks frequently fail in financial applications of generative AI, while reliance on subject-matter expert (SME) evaluation introduces subjective bias and systemic risks, leading to erroneous performance assessment. Method: This paper proposes the first LLM evaluation framework specifically designed for financial domains, systematically identifying four canonical risk categories—e.g., semantic mismatch and misalignment with business objectives—that arise when combining automated metrics with SME judgment. It integrates SME input, multi-granularity metric analysis, formal risk modeling, and validation on real-world industrial use cases. Contribution/Results: The resulting multidimensional evaluation framework significantly improves assessment reliability and business alignment. Deployed across multiple financial institutions, it demonstrably reduces metric misuse risks in LLM deployment, thereby enhancing model trustworthiness and decision-making robustness.

This work addresses the limited performance of large language models (LLMs) in high-dimensional software engineering optimization tasks, where they often fail to surpass Bayesian optimization. For the first time, it systematically compares human- and AI-generated domain knowledge injection strategies and introduces four novel architectures: Human-feedback-informed Domain Knowledge Prompting (H-DKP), Adaptive Multi-stage Prompting (AMP), Dimension-aware Progressive Refinement (DAPR), and a hybrid approach combining statistical scouting with RAG-enhanced knowledge integration (HKMA). By leveraging a multi-stage, dimension-aware, and hybrid knowledge fusion framework, the proposed methods effectively incorporate structured domain knowledge to significantly enhance LLMs’ ability to generate high-quality initial solutions. Evaluated on the MOOT high-dimensional benchmark, the approaches markedly reduce the Chebyshev distance to the optimal solution and, according to Scott-Knott clustering, outperform existing LLM warm-start baselines.