Distributed processing of unstructured text data is a challenge in the rapidly changing and evolving natural language processing (NLP) landscape. This landscape is characterized by heterogeneous systems, models, and formats, and especially by the increasing influence of AI systems. While many of these systems handle text data, there are also unified systems that process multiple input and output formats, while allowing for distributed corpus processing. However, there are hardly any user-friendly interfaces that allow existing NLP frameworks to be used flexibly and extended in a user-controlled manner. Due to this gap and the increasing importance of NLP for various scientific disciplines, there has been a demand for a web and API based flexible software solution for deploying, managing and monitoring NLP systems. Such a solution is provided by Docker Unified UIMA Interface-gateway. We introduce DUUIgateway and evaluate its API and user-driven approach to encapsulation. We also describe how these features improve the usability and accessibility of the NLP framework DUUI. We illustrate DUUIgateway in the field of process modeling in higher education and show how it closes the latter gap in NLP by making a variety of systems for processing text and multimodal data accessible to non-experts.

Maternal undernutrition and micronutrient deficiencies remain pervasive, contributing to adverse pregnancy outcomes and long-term health risks for mothers and offspring. Multiple micronutrient supplementation (MMS) during pregnancy has demonstrated benefits, including reduced risks of low birth weight, small-for-gestational-age births, and neonatal mortality, when compared with standard iron-folic acid supplementation. Current MMS strategies, however, often follow a standard MMS, overlooking variations in nutritional status, health profiles, and context. Advances in artificial intelligence (AI), particularly deep learning and natural language processing, provide opportunities to strengthen maternal nutrition programs by integrating diverse data sources. Rather than promising fully individualized recommendations, AI could help stratify women by risk of insufficiencies or deficiencies, highlight groups most likely to benefit from additional support, and inform the design of more responsive supplementation strategies during preconception and pregnancy. We outline a conceptual model in which multimodal health data-including electronic health records (EHRs), wearable sensor outputs, nutrition and fertility app logs, genomic markers, and sociodemographic information-are aggregated and analyzed by AI systems to inform personalized MMS plans. The framework introduces the concept of a "nutritional digital twin," a virtual profile of the patient's nutritional and metabolic state. This digital twin can simulate micronutrient needs and predict maternal-fetal outcomes under different supplementation scenarios, enabling clinicians to test scenario-based options (e.g. standard MMS ± targeted add-ons) for individuals. We describe how deep learning models can identify complex patterns (e.g. diet-genome interactions or behavioral trends) while natural language processing (NLP) algorithms extract clinically relevant insights from unstructured data (such as medical notes or patient queries). In addition, we discuss the role of digital maternal health tools, such as mobile apps and wearable trackers, in supplying real-time data to the AI models and in engaging women to improve adherence to supplementation regimens. Harnessing AI for MMS could transform maternal nutrition care in both high- and low-resource settings. In high-income contexts, rich data (comprehensive EHRs, genetic tests, continuous monitoring devices) could feed advanced predictive models to support risk-stratified care with protocolized supplementation options, under clinical oversight. In low- and middle-income countries, where maternal undernutrition and micronutrient gaps are most prevalent, AI-driven approaches can help stratify risk groups and optimize limited resources. Ubiquitous mobile phone access and digital health tools in many such settings provide avenues for data collection and intervention delivery. We highlight examples where machine learning on population data revealed "hidden hunger" patterns and key predictors of low supplement uptake (e.g. low education, minimal antenatal visits)-insights that policymakers can use to target nutrition programs. A nutritional digital twin could further allow scenario-testing (e.g. predicting the impact of adding a vitamin D supplement for a specific patient) before clinical decisions are made. To realize this vision, the key concerns are ethics, credibility, and fairness. Ethical frameworks must guide development so that sensitive reproductive health data are protected and clinician oversight remains central. The credibility of AI-generated recommendations depends on transparency about the assumptions used to translate nutritional and health data into supplement type and dose, and on prospective validation against maternal and neonatal outcomes. This requires a continuous feedback loop in which recommendations are tested in real-world settings and recalibrated using outcomes data, ensuring that the system learns from observed benefits and harms, rather than relying solely on theoretical modeling. Fairness demands that training data sets represent diverse populations and that solutions are tailored to local contexts to reduce bias and avoid widening disparities. Critically, the approach must be fed by data streams that extend beyond initial demographics and clinical baselines to include biomarkers, adherence patterns, and pregnancy outcomes, so that the models can be refined and dosing rules adjusted over time. If these safeguards are embedded, AI-enhanced personalized MMS can move beyond proof of concept towards a credible, equitable, and empirically grounded contribution to global maternal health. AI-driven personalized nutrition support represents a frontier in obstetric care. By combining clinical knowledge with data-driven intelligence, we can move beyond generalized prenatal supplements towards precision maternal nutrition. The integration of deep learning models and digital health innovations into antenatal care pathways has the potential to better nourish pregnancies, save lives, and ensure healthier futures for mothers and children worldwide.

A development of a new measure for renewable energy uncertainty in Vietnam by using natural language processing and textual analysis

From association to intervention: Semantic trajectories and knowledge frontiers in epilepsy-gut microbiota research revealed by bibliometrics and NLP.

Adverse drug reactions (ADRs) are a major cause of morbidity, hospital admissions, and in-hospital mortality, yet remain incompletely captured by post-marketing pharmacovigilance, which suffers from underreporting. Electronic health records (EHRs) contain clinical narratives that can reveal otherwise unreported ADRs. Natural language processing (NLP) offers a scalable means to extract structured information from clinical narratives, supporting ADR detection and assessment. We conducted a retrospective cross-sectional study within a multisite hospital network in Southern Switzerland to develop and evaluate NLP systems for ADR detection and information extraction from electronic discharge summaries. ADR classification models were trained on 400 discharge summaries and compared across multiple machine learning and vectorization strategies against a regular expression (regex) system. Drug and clinical event extraction were evaluated using 100 manually annotated summaries, benchmarking a dictionary-based approach against a two-step deep learning (DL) pipeline integrating transformer-based named entity recognition (NER) with a pharmacovigilance-oriented contextual relevance classifier. Performance was evaluated using standard metrics and a custom top-k ranking metric aligned with pharmacovigilance experts' daily capacity for reviewing positive cases to confirm the presence of ADRs. Logistic regression with Bag-of-Words achieved the best overall performance, combining high precision and effective case ranking. In a simulated deployment, this model identified nearly twice as many discharge summaries containing confirmed ADRs than as regex system. The two-step DL pipeline outperformed the dictionary-based approach for drug and clinical event recognition and accurately classified them according to pharmacovigilance purposes. These results demonstrate that NLP-based analysis of real-world clinical narratives can enhance pharmacovigilance while maintaining a manageable expert workload.

Natural language processing (NLP) has become an increasingly popular approach for analyzing textual responses in educational assessments. An important part of NLP involves cleaning and structuring examinees' written responses to create input data that conserves the syntax, semantics, and pragmatics of the words, thereby enabling the extraction of these features. This paper provides foundational knowledge on the steps needed for using NLP in educational measurement tasks, guiding researchers and practitioners through text preprocessing, feature extraction, and analyzing textual data from constructed response items. Additionally, an R-based example using Latent Dirichlet Allocation is provided, illustrating each step in the pipeline.

A Bengali-Hindi-Telugu parallel corpus for enhanced literary machine translation.

• K-means with GloVe embeddings yields the most semantically coherent clusters. • DBSCAN performs best for identifying outliers but underperforms in thematic clustering. • Embedding model selection has a stronger impact than algorithm choice on cluster quality. • GloVe is optimal for taxonomy development, while Wiki captures broader thematic patterns. This study conducts a comprehensive empirical evaluation of semantic clustering algorithms to identify the most effective approach for automatically organizing and extracting meaning from textual data. By systematically comparing the performance of K-means, K-medoids, and DBSCAN on word embeddings from GloVe and Wiki models, it provides data-driven insights for optimizing Natural Language Processing (NLP) pipelines in information management systems. The research suggests a practical framework for selecting clustering algorithms and embedding models based on specific operational objectives, such as document clustering, knowledge base construction, and content-based recommendation. The investigation employed a two-phase methodology. Initially, predefined word lists were transformed into numerical vectors using pre-trained GloVe and Wiki models. K-means, K-medoids, and DBSCAN algorithms were applied, with performance evaluated via Silhouette Score and Davies-Bouldin Index, complemented by Principal Component Analysis (PCA) for visualization. Results were benchmarked against manually curated semantic groupings. Subsequently, the findings were validated on a large-scale corpus of 303 research articles to assess scalability and real-world applicability. Analysis indicates that, under the evaluated configurations, K-means combined with GloVe embeddings produced comparatively higher semantic coherence and more interpretable cluster structures than the alternative methods considered. K-medoids demonstrated robustness against outliers but yielded less compact groupings. While DBSCAN indicated effective for outlier identification, it consistently underperformed in forming semantically meaningful clusters. The GloVe model significantly outperformed Wiki embeddings in generating precise and interpretable clusters, whereas Wiki produced broader, less distinct groupings. Large-scale validation confirmed these results, with K-means successfully identifying dominant research themes, including digital library adoption (43.2%), reference services (15.2%), and research data management (8.9%)—in a corpus of academic literature. Under the evaluated corpus characteristics and parameter settings, DBSCAN classified most documents as outliers, indicating limited suitability for this specific balanced document collection. K-means and K-medoids emerge as comparatively effective algorithms under the evaluated conditions. The study underscores the critical influence of vector representation models, with GloVe embeddings providing superior semantic distinction compared to Wiki. These findings offer clear, actionable guidance for selecting clustering methods in NLP applications, highlighting the necessity of aligning algorithmic choice with specific dataset characteristics and information management goals. This research moves beyond theoretical descriptions by delivering a rigorous, empirical comparison that elucidates the crucial interaction between algorithm selection and embedding models for semantic tasks. The findings provide practitioners with a context-dependent decision matrix: K-means with GloVe is effective under the studied conditions for taxonomy development and thematic categorization, whereas DBSCAN is preferable for outlier detection in noisy data. By demonstrating that GloVe's global statistical approach yields more distinct clusters than Wiki's contextual model for this purpose, the study contributes a practical, evidence-based framework for enhancing semantic analysis in real-world information systems.

Intelligent decision support for tunnel fire incidents: integrating dynamic knowledge graph with large language models

Automated detection of superior mesenteric artery occlusion on post-contrast CT Using a 3D deep learning model.

HeartBERT: A self-supervised ECG embedding model for efficient and effective medical signal analysis.

As a new trend in Natural Language Processing (NLP), prompt tuning has been explored to provide a reliable answer without requiring massive labeled samples and training learning in sentiment and emotion detection tasks. However, how to effectively encode the commonness and uniquesness across difference affections into prompts sets a limit to the potential of multi-affection joint detection. To fill this gap, we propose a multi-affection prompt (MAP) learning framework that takes both the commonness of multiple affections and the uniqueness of specific affection into consideration. More specifically, two different prompt encoders are first proposed to elaborate the multi-task shared prompt and the task-specific prompt, respectively. Second, a multi-task prompt interaction learning layer is proposed to capture the correlation between the multi-task and task-specific prompts. MAP adopts separate multi-task and task-specific prompts to learn different vectors for different affection tasks, thus mitigating the affection discrepancy of the [MASK] token in the masked language modeling task. Extensive experiments on two benchmark datasets show that our proposed method can significantly improve the multi-task generalization capability of PLMs, and yield better results than other state-of-the-art (SOTA) baselines, by the margin of 2.7% and 3.4%.

Keyword extraction is a foundational task in natural language processing, underpinning countless real-world applications. One of these is contextual advertising, where keywords help predict the topical congruence between ads and their surrounding media contexts to enhance advertising effectiveness. Recent advances in artificial intelligence have improved keyword extraction capabilities but also introduced concerns about computational cost. Moreover, although the end-user experience is of vital importance, human evaluation of keyword extraction performances remains under-explored. This study provides a comparative evaluation of prevalent keyword extraction algorithms with different levels of complexity represented by~TF-IDF, KeyBERT, and Llama~2. To evaluate their effectiveness, a mixed-methods approach is employed, combining quantitative benchmarking with qualitative assessments from 855 participants through four survey-based experiments. The findings demonstrate that KeyBERT achieves an effective balance between user preferences and computational efficiency, compared to the other algorithms. We observe a clear overall preference for gold-standard keywords, but there is a misalignment between algorithmic benchmark performance and user ratings. This reveals a long-overlooked gap between traditional precision-focused metrics and user-perceived algorithm efficiency. The study underscores the importance of human-in-the-loop evaluation methodologies and proposes analytical tools to facilitate their implementation.

Estimate renal cell carcinoma recurrence rates using electronic health records.

From linguistic analyses to large language models: A scoping review of methods used to investigate language features in depression research.

This paper aims to examine accounting students’ perceptions of learning Python through the lens of the Technology Acceptance Model (TAM). Reflective survey data were collected from 25 accounting students enrolled in a Python module at a UK university. A structured scaffolding approach was adopted to support students without prior coding experience, progressing from conceptual introduction to guided practice and independent tasks. Natural language processing techniques were used to analyse the data, including topic modelling and sentiment analysis. The findings show that students perceived Python as useful for automating tasks, handling data, and supporting employability, while reporting moderate ease and varied technical challenges. Positive attitudes persisted despite challenges. The paper contributes to accounting education by showing how TAM can explain accounting students’ experiences of coding. From a teaching excellence perspective, the paper shows that a structured scaffolding approach could support teaching by building confidence among non-technical learners.

Artificial intelligence in mental health: Knowledge mapping from traditional methods to large language models

Large language models (LLMs) have demonstrated remarkable capabilities in natural language processing. Their massive computational and memory requirements often necessitate cloud-based deployment, introducing challenges related to cost, latency, privacy, and network reliability. Deploying on-device LLMs alleviates these challenges, but is hindered by the severe resource constraints of edge hardware. This survey reviews efficient inference techniques for edge LLMs, with a focus on two key strategies of speculative decoding and model offloading. We categorize strategies into single-device and multi-device types, systematically analyzing the principles, recent advancements, implementations, and support within edge frameworks. Finally, we highlight the open challenges and future research directions that will advance the field of edge LLM inference.

Temporal local attention with adaptive decoding: Enhancing spiking neural networks for temporal computing applications.

The aim of this study is to conduct a linguistic analysis of English essays produced by intermediate-level second language (L2) learners compared to those generated by AI, across five linguistic dimensions: lexical richness, syntactic complexity, semantic similarity, discourse cohesion, and surface-level errors. A parallel corpus of 160 essays, 80 AI-generated and 80 learner-written, was collected and analyzed using Natural Language Processing (NLP) techniques. The results revealed that AI essays tend to be longer and syntactically more complex, with significantly higher lexical diversity and greater use of content words. While both types of essays share similar sentiment and cohesion patterns, the AI essays demonstrate more advanced sentence structures and deeper syntactic tree depths. Readability metrics show that the learners’ essays are simpler and more accessible. Error analysis revealed that the human essays contain four times more errors, particularly in spelling and stylistic choices. The study highlights how AI-generated language diverges from learner-produced writing and offers insights into how AI tools can be effectively leveraged to support language development at this proficiency level. • A comparative linguistic analysis of English essays produced by intermediate-level second language (L2) learners • A total of 160 essays, 80 AI-generated and 80 human-written, were analyzed using NLP tools • Focusing on five linguistic dimensions: lexical richness, syntactic complexity, semantic similarity, discourse cohesion, and surface-level errors • AI essays were longer, lexically denser, and syntactically more complex, with deeper parse trees and more frequent use of content words • The human-authored essays demonstrated simpler sentence structures and greater readability, containing 4 times more surface-level errors