Computación y Sistemas

Mathematical Model of Gastric Cancer with Immunotherapy: Global Dynamics and Tumor Clearance Conditions

Leonardo F. Martinez — 2025-09-25

Gastric cancer has positioned itself amongthe leading causes of cancer death worldwide. Mostof these tumors are gastric adenocarcinomas, which originate in the gastric mucos a from a chronic infection linked to H. Pylori bacterium. Traditional treat mentsare not entirely effective, however, there are high expectations of using the immunotherapies for gastric cancer treatment. Nevertheless, knowledge of the mechanisms of tumor evolution and their interactions with the immune system is limited. For this reason, we present a qualitative mathematical model of first-orderOrdinary Differential Equations (ODEs), which describes some survival mechanisms of intestinal-type gastricadenocarcinoma and its interaction with the immune system, assuming that H. Pylori and cellular cannibalism influence the tumor growth. We study the local and global dynamics of the model and propose sufficient conditions in an immunotherapy treatment parameter to eradicate gastric cancer. Finally, we perform numerical simulations and discuss the biological implications of our results.

Video Object Tracking by Feature Point Descriptor and Template Matching

Andrés Ely Pat-Chan — 2025-09-25

The present research focuses on developing a method based on feature point descriptors and template matching and comparing its performance with a method based on deep learning. These methods have particular aspects in how they were implemented; some stand out for the simplicity of their structure and others for the complexity they entail. The methods presented in this work range from developing a basic template matching algorithm, developing an algorithm based on feature point descriptors incorporating the template matching qualities to obtain better results, to implementing a method based on deep learning. Performance and precision tests are carried out to compare the methods on a selected dataset of video object tracking.

Generation of Feature Vectors for Identifying Medical Entities in Spanish

Gabriela A. García-Robledo — 2025-09-25

Natural Language Processing (NLP) encompasses a range of high impact techniques for enabling computers to interact with humans in a more natural manner. One such technique is the extraction of entities, which allows computers to identify relevant information within a text. This paper presents a methodology for the recognition of medical entities within a texts written in Spanish. The methodology combines syntactic, semantic, and contextual features at the word level. The principal objective of a feature-based approach is the identification of drug, anatomy, and disease entities. A training evaluation was conducted on two types of machine learning algorithms, with an accuracy of 98\% on an external set. Additionally, an accuracy check was performed for each medical class.

WOA-SVM: Whale Optimization Algorithm and Support Vector Machine for Hyperspectral Band Selection and 2D Images Feature Selection

Seyyid Ahmed Medjahed — 2025-09-25

This paper proposes a new optimization based framework for feature selection and parameters determination of support vector machine, called WOA-SVM and it is applied on band selection in hyperspectral images and feature selection in 2D images. The proposed approach WOA-SVM is based on Whale Optimization Algorithm (WOA), which is a meta-heuristic inspired from the social behaviors of humpback whale and never been benchmarked in the context of feature selection nor parameters determination. A new fitness function is designed. WOA-SVM is tested with three hyperspectral images widely used for band selection and classification. Note that one of the problems in hyperspectral image classification research is the identification of informative bands (band selection). In addition, we demonstrate the efficiency of the proposed approach on Mammographic Image dataset (MIAS). The experimental results prove that the proposed approach is high performance and very competitive approach. The WOA-SVM approach is useful for parameter determination and feature/band selection in SVM.

MultiLate Classifier: A Novel Ensemble of CNN-BiLSTM with ResNet-based Multimodal Classifier for AI-generated Hate Speech Detection

Advaitha Vetagiri — 2025-09-25

The rise of multimodal hate speech, which combines text and visual elements, poses significant challenges for online content moderation. Traditional detection models often focus on single modalities and struggle with AI-generated content that is contextually nuanced and semantically complex. These limitations lead to suboptimal performance, as existing frameworks are not robust enough to handle the evolving nature of hate speech across diverse contexts and datasets. An integrated approach that captures the interplay between text and images is needed for more accurate identification. This paper introduces a novel MultiLate classifier designed to synergistically integrate text and image modalities for robust hate speech detection to address these challenges. The textual component employs a CNN-BiLSTM architecture, augmented by a feature fusion pipeline incorporating Three W's Question Answering and sentiment analysis. For the image modality, the classifier utilizes a pre-trained ResNet50 architecture alongside Diffusion Attention Attribution Maps to generate pixel-level heatmaps, highlighting salient regions corresponding to contextually significant words. These heatmaps are selectively processed to enhance both classification accuracy and computational efficiency. The extracted features from both modalities are then fused to perform comprehensive multimodal classification. Extensive evaluations of the MULTILATE and MultiOFF datasets demonstrate the efficacy of the proposed approach. Comparative analysis against state-of-the-art models underscores the robustness and generalization capability of the MultiLate classifier. The proposed framework enhances detection accuracy and optimizes computational resource utilization, significantly advancing multimodal hate speech classification.

Cyberbullying Detection on Social Media Using Machine Learning Techniques

Abdullah Abdullah — 2025-09-25

With the advancement of social media, cyberbullying among social media users has grown knowingly, causing serious problems such as economic, health, social, and psychological problems. Nowadays, most of the social media users have been affected by cyberbullying, aggressive behavior, and suffering from digital harassment. Several machine-learning techniques are applied to detect cyberbullying from text data sets generated by Twitter and Formspring.com. It is found that cyber-bullies participate in fewer online groups and are not much more popular than normal users. However, one major issue in cyberbullying detection and sentiment analysis is the deficiency of labeled data. Consequently, researchers are usually forced to depend on survey-based data where committers and sufferers provide details about their imitations. Therefore, un-labeled data is gathered, and data annotation techniques are applied to label the data set. After that two more labeled datasets are gathered and machine learning classifiers (voting classifier, random forest, SVM, linear SVM, and Gaussian naïve Bayes) are applied to these labeled datasets to train the model. Thus, the trained model was applied to the primary dataset, but the results gained were not satisfactory. Therefore, the select-k-best (chi^2) feature selection algorithm was applied to reduce features. Hence, the primary dataset is verified by applying the trained model with an accuracy of 91%.

Harnessing Uncleaned Data for Stress Detection in Tamil and Telugu Code-Mixed Texts

Luis Ramos — 2025-09-25

Stress is a common experience in daily life, but it can significantly impact mental well-being in certain situations, making the development of robust detection models imperative. This proposal introduces a methodical approach to the stress detection in code-mixed texts for Dravidian languages. The challenge encompassed two datasets, targeting Tamil and Telugu languages respectively. This proposal underscores the importance of testing uncleaned text, such as deleting emojis, special characters, etc., in classification methodologies. In this proposal were evaluated Logistic Regression, Random Forest and Support Vector Machine algorithms featuring three textual representations: TF-IDF, word and character N-grams. This proposal demonstrated strong performance across both languages, achieving a Macro F1-score of 0.75 for Tamil and 0.74 for Telugu, surpassing the results obtained using other complex techniques involving LLMs. The results underscore the value of uncleaned text for mental state detection and the challenges of classifying code-mixed texts in Dravidian languages, indicating that there is potential to be explored, especially in Tamil and Telugu texts.

Cross-Platform Performance Evaluation of Matrix Multiplication: Insights from MKL, cuBLAS, and SYCL

Luis Alejandro Torres Niño — 2025-09-25

Matrix multiplication is a fundamental operationin deep neural network training and scientific computing, optimized through libraries such as Intel MKL and NVIDIA cuBLAS. MKL enhances CPU execution using multithreading and AVX-based vectorization, improving memory bandwidth utilizationand computational throughput. Conversely, cuBLAS leverages CUDA’s massive parallelism, employing thousands of GPU cores and Tensor Cores to accelerate matrix computations, though Tensor Core usage introduces numerical precision loss. SYCL extends heterogeneous computing capabilities, enabling efficient workload distribution across CPUs and GPUs. This study analyzes execution time, computational efficiency, and power consumption, utilizing PAPI and PERF to evaluate third- and fourth- generation Intel CPUs and selected NVIDIA GPUs. Results indicate that MKL delivers high CPU performance, while SYCL offers an alternative approach with distinct efficiency characteristics. GPU-based benchmarks show that cuBLAS with Tensor Cores achieves maximum throughput but at the cost of precision, whereas cuBLAS without Tensor Cores preserves accuracy with minimal performance trade-offs. These differences highlight the importance of optimization strategies in artificial intelligence and scientific computing, where scaling models and simulations demand efficient, high-performance, and sustainable computation.

Integration of Agile Approaches with Quantum High-Performance Computing in Healthcare System Designs

. Abdullah — 2025-09-25

This study explores the integration of Agile methodologies and quantum high-performance computing (HPC) in healthcare. Agile methodology offers flexibility, while quantum HPC has the potential to revolutionize healthcare. The paper presents an architectural design for HPC in healthcare, discussing tool alignment, software development, and integration challenges. It emphasizes user-centered design, cross-functional teams, continuous integration, and retrospectives for software development. Personalized medical tools leverage data analytics and machine learning (ML) with quantum HPC for improved speed and accuracy. Real-time patient monitoring systems utilize wearable devices and sensors, developed iteratively using Agile methodology. The integration of Agile methodologies with quantum HPC can transform healthcare, improving access, efficiency, and patient outcomes. The study focuses on developing personalized medicine, real-time monitoring, and telemedicine tools to ensure improved care, security, and privacy through quantum HPC advancements.

Protocolos criptográficos de consenso en Blockchain para el internet de las cosas

Luis Miguel Saldaña Trejo — 2025-09-25

El éxito de Blockchain en el ámbito de las criptomonedas ha motivado su exploración en otros sectores, como los servicios financieros, públicos, sociales y el Internet de las Cosas (IoT). Esta tecnología ofrece ventajas clave, como descentralización, integridad, pseudoanonimato y trazabilidad completa de las transacciones. Su aplicación en redes IoT genera expectativas por el potencial de cómputo distribuido, almacenamiento descentralizado y resistencia a la manipulación. Sin embargo, la mayoría de los protocolos de consenso actuales fueron diseñados para nodos con alta capacidad de procesamiento y almacenamiento, recursos con los que no cuentan los dispositivos típicos de IoT. Aunque han surgido propuestas que reducen los requisitos de cómputo o almacenamiento, aún no existen protocolos específicamente diseñados para nodos con recursos limitados. En este artículo se analizan distintos protocolos criptográficos de consenso y se identifican aquellos que podrían adaptarse a redes IoT.

Testing three different Speech Synthesizers to acknowledge the advantages of DNN systems against HMM Methods

Carlos Angel Franco-Galvan — 2025-09-25

Abstract. This document reports MOS results after testing naturalness and expressiveness in three different speech synthesis systems. A first system is based on HMM, the second one combines HMM and DNN and the third one is solely based on DNN. According to the results, DNN systems outperform HMM systems in synthetic speech quality.

Minería de datos y aprendizaje máquina aplicado en la predicción de salud mental en trabajadores de Tecnologías de la Información

Kori Xiomara Antúnez Palomino — 2025-09-25

En la actualidad la salud mental es un problema cada vez más frecuente en las personas. Trastornos mentales tales como los de ansiedad y depresión tienden a contribuir a los cambios de comportamiento relacionados con su trabajo, como la reducción del nivel de actividad y el mal rendimiento. El objetivo de esta investigación es medir las actitudes respecto la salud mental en el área de TI. Como técnicas de selección de atributos se usó el filtro ANOVA F-test y el filtro Chi-Cuadrado, para el modelado se aplicaron los algoritmos de K-Nearest Neighbor (KNN), Decision Tree, Support Vector Machine (SVM), Naive Bayes y Neural Network. Se trabajó con un conjunto de datos obtenidos de OSMH/OSMI Mental Health in Tech Survey, el cual contó con alrededor de 1400 respuestas, estos resultados fueron obtenidos de una encuesta llevada a cabo en el 2016. El algoritmo que obtuvo los mejores resultados en el conjunto de datos analizados fue Neural Network. Como métricas del modelo se obtuvo un F1-Score del 85.92%, un área bajo la curva ROC de 0.903, un menor valor de falsos negativos con 23 y mayor valor de verdaderos positivos con 119 en la matriz de confusión.

An Evaluation of FRQI and NEQR Encoding Using QCNN for Forecasting Tropical Cyclone Intensity

S. P. Rajamohana — 2025-09-25

Effective weather forecasting for cyclones is crucial for minimizing harm to both people and the environment. Accurate estimation of tropical cyclone (TC) intensity is essential for disaster prevention. Although convolutional neural networks (CNNs) have improved this process, they often struggle to capture global spatial relationships in images. Quantum Image Processing (QIP) leverages quantum computing advantages but faces challenges such as noise and hardware limitations. This study represents the first effort to estimate tropical cyclone intensity prediction using two popular quantum image representation formats: Flexible Representation of Quantum Images (FRQI) and a Novel Enhanced Quantum Representation (NEQR), as data encoders in Quantum Convolutional Neural Networks (QCNN) utilizing INSAT 3D satellite images. By employing TCs from 2012 to 2021 as training data, the model achieved an overall mean square error (MSE) of 0.0384 for FRQI and 0.0002 for NEQR. The findings indicate that NEQR significantly outperforms FRQI in cyclone image prediction.

Detección de enfermedades en hojas de tomate usando aprendizaje profundo

Edgar R. Arredondo-Basurto — 2025-09-25

En 2023, el sector agrícola de México representó el 2.2% del Producto Interno Bruto (PIB). Sin embargo, este sector se encuentra continuamente amenazado por bacterias, virus o enfermedades que podrían tener un impacto financiero significativo en los agricultores. Por esta razón, la identificación y detección temprana de enfermedades adquiere una gran relevancia. Para lograrlo, el agricultor debe contar con una formación integral que abarque diversas disciplinas, experiencia en el reconocimiento de síntomas y un conocimiento profundo de la amplia gama de causas de estas enfermedades. En la presente investigación, centrada en las plantas de tomate, se desarrolló un sistema de reconocimiento de imágenes basado en técnicas de aprendizaje profundo (deep learning) para identificar enfermedades en los cultivos de tomate y ponerlo a disposición de los usuarios finales a través de un sistema web. Las pruebas realizadas muestran hasta un 98% de precisión en la clasificación, lo que subraya su potencial en aplicaciones agrícolas.

Prerequisites of in-Context Learning for Transformers on Queries

Bulat Shkanov — 2025-09-25

Pre-trained generative transformers (GPT) with their 200+ billion parameters have already demonstrated their ability to successfully solve a wide range of text related problems without the need for additional taskspecific training. However, it has been observed that solution quality can be significantly improved for certain queries that reflect task formulation and conditions. It indicates that the transformer is further trained based on the query context, and the aim of this study is to show why GPT transformers enable to do it. To this end, the article jointly considers: elements of transformer architecture (data compressors and sentiment neurons), elements of the user interface with transformers (zero-shot and few-shot prompts), and text processing procedures (arithmetic coding and minimum description length). The authors attempt to provide a theoretical justification for the convergence of the sequential fine-tuning process using Hoeffding's inequality. The study presents experimental results demonstrating GPT transformers' capabilities for in-context learning. This confirms their potential for further development in natural language processing technologies.

Financial Fraud Detection in the Banking Sector Using Machine Learning: An Exhaustive Systematic Review

Orlando Jeri-Alvarado — 2025-09-26

In recent years, the application of machine learning techniques for detecting financial fraud within the banking sector has experienced a remarkable increase. This paper seeks to highlight this progress and emphasize its impact on enhancing fraud prevention and control systems. The objective of this paper is to explore, determine, and identify the current state of knowledge regarding the use of machine learning in financial fraud detection in the banking sector. This study was based on 61 papers selected from six major digital libraries: IEEE Xplore, Scopus, ScienceDirect, ProQuest, ARDI, and Web of Science. Only peer-reviewed journal papers were included. The systematic review covered publications between 2019 and 2025, available in openaccess databases, focusing on the use of machine learning in detecting financial fraud in the banking sector. The findings from the 61 reviewed papers indicate that the most widely used programming language for machine learning solutions is Scala. Additionally, tools implemented in fraud detection and gaps in model comparison were identified. It is recommended to explore more recent approaches and banking contexts that have not yet been addressed.

A Review on the Role of Fuzzy Logic in Hybrid Intelligent Systems

Oscar Castillo — 2025-09-25

In this review the role of fuzzy logic (FL) in hybrid intelligent systems is discussed. We first review the papers in which fuzzy logic is used in conjunction with neural networks, as well as their application areas. Then, we review the papers in which fuzzy logic has been used in combination with evolutionary algorithms, and the corresponding application areas. We also review the papers in which fuzzy logic has been used in a hybrid way with optimization algorithms, as well as the application areas. Regarding FL, we consider the evolution that has been undergoing, where initially type-1 fuzzy logic was proposed and used, later type-2 was proposed and now more recently type-3 has been put forward. The evolution of FL has occurred due to the need of handling the higher uncertainty levels that real world problems have. In this regard, we analyze the impact of this evolution on different types of hybrid intelligent systems.

Impact of the Intelligent Assistants with RAG on Information Access and Consultations in Local Governments: A Systematic Review

Jhair Collado — 2025-09-28

The need to ensure clear, timely, and equitable citizen access to public information has driven local governments to modernize their service processes. Within this framework, Intelligent Assistants with Retrieval-Augmented Generation (RAG) are emerging as a promising solution, although scientific evidence regarding their impact remains scattered and requires critical systematization. This paper aims to determine the impact of RAG-based intelligent assistants on query resolution and information access in local governments. A systematic literature review was conducted following the PRISMA methodology, analyzing 80 open-access papers published between 2020 and 2025 in IEEE Xplore, Scopus, ScienceDirect, ACM Digital Library, Wiley Online Library, and Taylor & Francis Online. The findings indicate that the most frequently used algorithms in e-commerce are Random Forest, SVM, and neural networks; that Python predominates as the development language, followed by Scala and Matlab; that most studies are published in Q1 journals with high academic rigor; and that the most recurrent keywords emphasize classification, prediction, and user experience. This paper provides a solid foundation for future research, guiding the development of more diverse and methodological approaches in the use of intelligent assistants with RAG in local governments.

Design of Ensemble Neural Networks with Type-3 Fuzzy Aggregation using Particle Swarm Optimization and Genetic Algorithms for Ethereum Prediction

Martha Pulido — 2025-09-26

In this study, an ensemble neural network (ENN) for Ethereum time series prediction was optimized using particle swarm optimization and genetic algorithms. Additionally, Type-1, Type-2, and Type-3 fuzzy inference systems, of both Mamdani and Sugeno types, were designed for achieving the prediction. The integration performed with these fuzzy systems is achieved by utilizing the results from optimizing the ENN with each optimization algorithm. In this case, the Ethereum data is the series being used for testing the proposal. This approach aims to minimize prediction error by combining the responses of the ENN with Type-1, Type-2, and Type-3 fuzzy systems, each consisting of five inputs and consequently 32 fuzzy rules are utilized. The results show that the Type-1, Type-2, and Type-3 fuzzy system approach yields an accurate prediction of the Ethereum series, as further validated by statistical tests on the results of the fuzzy systems.

Representation of Best Practices in IoT Systems by Using the SEMAT Essence Kernel

Carlos M. Medina O — 2025-09-28

The Internet of Things (IoT) facilitates coordinated interaction among machines, devices, and users. Best practices in IoT encompass processes designed to enhance the efficiency of IoT systems implementation. While state-of-the-art reviews reveal diverse methods for modeling such practices, existing models in the literature remain fragmented: they often address isolated development phases and lack replicability due to insufficiently structured methodologies. This study addresses this gap by modeling IoT best practices found in scientific literature on IoT systems using the SEMAT Essence Kernel language (Software Engineering Method and Theory). From an analysis of 97 scientific papers, four best practices were selected and processed through a terminological extractor, generating a dictionary of 123,566 terms to standardize their nomenclature. Each practice’s components were systematically mapped to SEMAT Essence Kernel elements. The resulting models represent best practices in power consumption, data security, cloud computing resource utilization, and Big Data integration for IoT systems. The proposed approach demonstrates the SEMAT Essence Kernel’s efficacy in formalizing IoT best-practice knowledge. Validation by a panel of IoT experts yielded promising results, confirming the models’ robustness~and~applicability.

PSO and Random Forest Techniques to Improve IDS Performance for Multi-class classification

Benaissa Safa — 2025-09-28

With the increasing digitization of the world, the risk of attacks also increases, creating a need to develop effective network intrusion detection techniques. In this research, a two-phase approach was proposed by the authors to improve IDS performance in multi-class classification case. In the first phase, only the relevant features are identified and conserved using an evaluator based on Particle Swarm Optimization. In the second phase, network attacks are classified using the Random Forest classifier. Furthermore, a comparative study is conducted, involving other classifiers such as Naïve Bayes, Stochastic Gradient Descent, Deep Learning, etc. For multi-class classification, the NSL-KDD dataset was used to conduct experiments, and the obtained results showed an accuracy of 99.40%. The performance results of our technique are presented and compared with other competing techniques. The obtained results clearly indicate that our technique outperforms the others.

Optimization of Access Control in Private Educational Institutions: An Approach with Geolocation and Biometrics Using Mobile-D

Alexis Prado-Feliciano — 2025-09-28

This paper addresses the application of the Mobile-D methodology in an educational setting, motivated by the need for efficient software to comprehensively manage faculty access registration. The lack of adequate technological solutions has limited the optimization of this process, generating inefficiencies in both time and cost. The main objective was to evaluate the effectiveness of Mobile-D in improving faculty access procedures, with particular emphasis on reducing registration times and decreasing associated administrative costs. To address this problem, a pure experimental and quantitative approach was employed, analyzing the impact of Mobile-D. The study population included all access processes carried out by faculty members, with a sample of 30 cases used for testing. Technologies such as Firebase, a NoSQL database manager, and the OpenCV library, used to develop a facial detection algorithm, were integrated. The results demonstrated significant improvements in operational efficiency, confirming the potential of Mobile-D to transform educational processes. A considerable reduction in both time and costs was observed. However, practical challenges also emerged, such as the need to optimize facial recognition through pretrained models and to explore automatic geolocation to enhance the effectiveness of technology.

Detection of Silent Water Leaks in Household Using Artificial Intelligence Methods

Uriel Amado Ramírez-Hernández — 2025-09-28

Water losses in distribution systems constitute a significant global challenge, undermining water resource sustainability, increasing operational costs, and threatening the water security of millions. In Latin America, up to 40% of treated water is reportedly lost due to leaks, ruptures, and defective connections (Xylem, 2025). At the household level, silent leaks—particularly in toilet flushing systems—can waste over 37,850 litres annually per dwelling (US EPA, 2024). Various international studies have addressed early leak detection using intelligent systems. In Europe, wireless sensor networks and machine learning models such as Random Forest, Support Vector Machines, and neural networks have been deployed for anomaly detection in urban networks. Asian research has demonstrated detection accuracies exceeding 97% through convolutional neural networks trained on acoustic and vibrational signals, enhanced by contrastive learning to address data scarcity. Hybrid approaches combining hydraulic modelling with AI have been applied in the Middle East and China, whereas logic-based and anomaly detection algorithms have been integrated into real-time platforms in Australia and Canada. Sensor placement optimisation via graph partitioning has further improved coverage efficiency. Despite their effectiveness, these solutions often require substantial investment and advanced infrastructure, limiting their applicability in resource-constrained environments. This study proposes a cost-effective, perceptron-based model for detecting silent leaks in household toilets, integrated within an Internet of Things (IoT) framework. The system employs a Hall-effect flow sensor to capture high-resolution filling-time and pulse-count data, processed through supervised learning to discriminate between normal consumption and leakage. Experimental results under real-use conditions achieved 98% classification accuracy, demonstrating both technical feasibility and operational suitability. This approach offers a practical, computationally efficient solution for domestic contexts in Latin America, enabling real-time monitoring and immediate user alerts, thus supporting water conservation efforts through accessible intelligent detection.

EEG-Based Classification of Consumer Preferences Using PCA

Mauro Daniel Castillo Pérez — 2025-09-28

This study explores the neural correlates of consumer preferences for functional foods using EEG signals from 83 participants. Using Principal Component Analysis (PCA) for dimensionality reduction and visualization, we identified distinctive brain wave patterns associated with liked and disliked food products. PCA revealed dominant activity in Delta (0.97) and Theta (0.92) waves for preferred foods, indicating strong sensoriemotional interaction, while disliked foods showed reduced Alpha (0.23) and Beta (0.14) activity, reflecting decreased cognitive processing. Statistical validation (70\% explained variance using PCA, p < 0.05 in permutation tests) confirmed the robustness. The approach demonstrates how integrating PCA can decode consumer behavior, providing useful insights for neuromarketing and product development, such as optimizing sensory attributes or adapting formulations based on neural profiles. Future work could integrate machine learning for predictive modeling.

Application of Large Language Models to the Diagnosis of Respiratory Diseases

Anastasiia A. Shamrikova — 2025-09-28

The implementation of large language models (LLM) using artificial intelligence can currently become extremely popular for solving various medical problems. Eight publicly available AI systems were prompted to make an otolaryngological diagnosis based on known symptoms obtained using the standard SNOT-22 medical questionnaire. The aim of the study was to find out to what extent modern AI systems can make a diagnosis without prior training. The results showed that most systems, with one exception, performed satisfactorily, achieving an accuracy of 70-80% compared to an accuracy of 84% achieved by a human specialist using various machine learning methods. The advantages and disadvantages of AI systems for medical diagnostics are discussed in the paper.

Machine Learning approaches for Predicting Medical Costs in Oncology Patients: A Systematic Literature Review

Jesus Saenz-Chang — 2025-09-28

Although the use of machine learning (ML) in healthcare has increased significantly, a critical systematization of its application to medical cost prediction is still lacking. This paper aims to rigorously examine recent literature to identify methodological approaches, knowledge gaps, and emerging trends related to the economic use of ML in health. To this end, a systematic review of 71 papers was conducted, complemented by bibliometric analysis, journal quartile assessment, and thematic categorization. These strategies were applied across highly recognized academic databases, including Scopus, IEEE Xplore, ACM Digital Library, PubMed, and Springer Nature Link. The main findings indicate that: (1) most studies are concentrated in highly digitalized countries, which restricts their applicability in less developed contexts; (2) although a significant number of publications appear in Q1 journals, they do not always achieve high levels of scientific objectivity; and (3) the predominant topics focus on image-based diagnosis, while the prediction of medical costs remains an emerging and underexplored field. Overall, the results highlight a substantial gap between the technical development of ML and its integration into financial decision-making in healthcare. It is recommended to promote research with greater geographical diversity, grounded in more robust theoretical frameworks and guided by ethical principles that ensure implementation.

HAMP: A Hardware Accelerator Multi-Platform Benchmark for High Performance Heterogeneous Computing

Rogelio Valdez — 2025-09-26

This paper introduces a Hardware Accelerator Multi-Platform Benchmark for high performance heterogeneous computing called HAMP, a multi-platform tool designed to analyze the performance of hardware accelerators such as CPUs, GPUs and FPGAs. HAMP benchmark stands out for its ability to evaluate different hardware architectures within a unified environment using a common programming language, C++ with the OpenCL framework. The multi-platform tool integrates a selection of state-of-the-art kernels for the evaluation of hardware characteristics and arithmetic operations, employing two performance metrics (speed and bandwidth). Developed in Python, a graphical user interface (GUI) simplifies interaction with the tool, allowing users to configure and execute kernels on the hardware accelerator without needing technical knowledge. Experiments were performed on three hardware accelerators with the five kernels comprising the HAMP benchmark. The results obtained indicate that HAMP is an efficient and reliable multi-platform tool, facilitating the comparison of diverse hardware architectures within a unified design environment, a capability not previously available in the state-of-the-art.

Evolución de la estimación del esfuerzo en el desarrollo de software: De los modelos tradicionales a la automatización basada en IA y DevOps

Iliana Lizbeth Alvarado Lara — 2025-09-26

La estimación del esfuerzo y los recursos es crucial para el éxito en el desarrollo de software, un campo que enfrenta desafíos constantes debido a su complejidad y dinamismo. Este estudio analiza la evolución de las técnicas de estimación del esfuerzo en el desarrollo de software desde la adopción de metodologías ágiles hasta la consolidación del paradigma DevOps, mediante un análisis bibliométrico de 4,196 documentos de Scopus (2008-2024) utilizando SciMAT y VOSviewer. Los resultados evidencian una transición de modelos tradicionales, como COCOMO y puntos de función, hacia enfoques basados en inteligencia artificial, redes neuronales y modelos predictivos, destacando tres áreas clave: gestión de proyectos, técnicas automatizadas y tecnologías avanzadas. Los mapas estratégicos muestran que la estimación del esfuerzo en el desarrollo de software es el tema más influyente, mientras que fiabilidad del software y optimización mantienen su relevancia, y emergen tendencias como algoritmos de aprendizaje automático y árboles de decisión. A nivel global, Estados Unidos, China y Alemania lideran la producción científica. Se identificaron instituciones influyentes dentro de este campo, como la Universidad del Sur de California y la Universidad del Oeste de Canadá, aunque estas no necesariamente pertenecen a los países con mayor volumen de publicaciones, lo que sugiere que su impacto se debe a la calidad y relevancia de sus contribuciones científicas. Finalmente, se destaca la necesidad de desarrollar herramientas híbridas que integren modelos clásicos y técnicas avanzadas, promoviendo enfoques de estimación del esfuerzo en el desarrollo de software en entornos DevOps y ágiles para mejorar la precisión y eficiencia en la gestión de proyectos de software.

Arduino Devices as a Platform for Execution of Machine Learning Algorithms: A Brief Review and Experimentation

Juan José Flores Sedano — 2025-09-26

At present, the Internet of Things and Artificial Intelligence are among the most relevant transformative technologies for making a smart world a reality. In this context, this paper explores the transformative synergy between the Internet of Things (IoT) and Artificial Intelligence (AI) by integrating AI algorithms into Arduino devices. The literature review has demonstrated a current need for optimization in implementing AI algorithms on Arduino platforms. Through a empirical literature review and practical experimentation, this paper provides a comprehensive analysis of several Arduino boards, including the Portenta H7 Lite, Arduino Uno, Wemos D1 ESP8266, and Arduino Nano 33 BLE, comparing their performance for AI projects. The selection of an IoT board is emphasized based on project-specific needs and budget considerations. The research presented in this paper reveals the impact of combining IoT, AI, and Arduino on reshaping interactions with the connected world, paving the way for intelligent systems enabled for decision-making and to execute complex tasks.

Marco de evaluación para la óptima selección de componentes de hardware para sistemas del Internet de las cosas

Julio Víctor Sánchez-Hernández — 2025-09-26

The selection of hardware components is crucial in the design of Internet of Things solutions, as it directly impacts their performance, cost, and scalability. The wide variety of available development boards, sensors, and peripherals complicates the selection process, given that each component has distinct characteristics. This paper proposes a systematic framework for hardware selection in Internet of Things systems, utilizing an approach based on the objective identification of key criteria such as energy efficiency, cost, reliability, and ease of integration. The framework was applied in a case study to design a sensor node for precision agriculture aimed at monitoring soil temperature and humidity. The results demonstrate that the framework streamlines decision-making, ensuring the selection of components aligned with project requirements while enhancing the quality of Internet of Things system designs.

Classification of pH Status in Stevia Plants Using Supervised Neural Networks and Computer Vision

Jesús Emmanuel Brizuela-Ramírez — 2025-09-26

Artificial Intelligence and Computer Vision have revolutionized Precision Agriculture, enabling automated crop monitoring. This study proposes a model based on Neural Networks to classify soil pH in Stevia rebaudiana Bertoni crops, optimizing agricultural management and crop sustainability. To achieve this, Stevia images were processed using data augmentation techniques, extracting color features in RGB and hexadecimal formats. A supervised Artificial Neural Network was then trained to classify soil pH into acidic, optimal, and alkaline categories. The proposed model, StePHVIA, achieved 99% accuracy, outperforming pretrained architectures such as MobileNetV2 (97.38%) and ResNet-50 (76.38%). The evaluation was conducted using metrics such as Matthews Correlation Coefficient, accuracy, recall, and F1-score. These results confirm the effectiveness of Computer Vision and Deep Learning in Precision Agriculture, providing a real time and low cost alternative for soil monitoring. StePHVIA facilitates the early detection of soil imbalances, optimizing fertilizer application and improving Stevia crop productivity.

Evolutive Layers in M3GP Basis for Symbolic Deep Learning Models

Luis Muñoz Delgado — 2025-09-26

This paper introduces a novel approach toconstructive induction in genetic programming throughthe Multidimensional Multiclass Genetic Programmingwith Multidimensional populations also known as M3GPalgorithm. M3GP is leveraged to create new featuresthat either augment the original dataset or transform itinto a refined version, resulting in improved performancefor learning algorithms. With this premise the primarycontribution of this work is the integration of an evolutivelayer structure within M3GP, where the n best-performingfeatures generated in the previous iterations are reusedto continuously enhance the algorithm’s performance.This approach parallels the concept of layers inneural networks, establishing a pathway for symbolicconstruction methods, such as genetic programming, toincorporate layered learning. The second contribution isdefining the structure of operation that can be appliedin any constructive induction method to connect theimprovement of symbolic models, and sampling keypoints of improvement for configuration options. Thefindings underscore the potential of evolutionary layeringto improve feature generation and model accuracy,marking an advancement in the constructive inductionfield into a deep learning process. The result showsa higher tendency of fitness improvement againstnon-layered networks for regression problems and alower improvement in classification problems, openingthe possibilities for a new niche for deep evolutivenetworks.

Importance of the Parameterization Schemes in the WRF Model for Wind Speed Forecasting: Case Study of Tepuxtepec, Michoacán

Itzagueri García-Rodríguez — 2025-09-26

The planning of wind energy dispatchconstantly faces the challenges of wind speed intermittency and variability. Therefore, it is crucialto have models that generate reliable forecasts tosupport the development of this renewable energysource. This study evaluates the performance of threeconfigurations of the Weather Research and Forecasting(WRF) model for hourly wind speed forecastingin Tepuxtepec, Michoacán, Mexico, with a 24-hourforecast horizon. Three parameterization schemeswere compared: WRF_WMK, WRF_TBG, andWRF_MYB. These schemes were selected based on thegeographic, climatic, and meteorological characteristicsof the region, as well as the need to assess theWRF model’s performance under different physicalconfigurations. Simulations were conducted for fourrepresentative dates—one per season—considering theannual temperature cycle that influences wind behavior.The simulations used MERRA-2 reanalysis data asinput and were evaluated against measurements fromNASA’s POWER project. The comparison betweensimulated and observed wind speeds was performedusing four error metrics: Root Mean Square Error(RMSE), Mean Absolute Error (MAE), Bias, and theCorrelation Coefficient (r). Additionally, PredictionIntervals (PIs) at 80% and 95% confidence levels werecalculated to assess the reliability of the forecasts.The results showed that the WRF_TBG configurationoutperformed the others, reducing RMSE by up to 60%compared to WRF_WMK. The forecasted values werewithin the 80% PI for up to 80% of the total values, andwithin the 95% PI for up to 100%. Seasonal evaluationrevealed that the model performed best in winter andworst in summer, likely due to the influence of intenseconvective processes during the latter season.

Modelo híbrido de programación genética y redes neuronales para el reconocimiento de emociones

Alvaro A. Colunga-Rodriguez — 2025-09-26

El reconocimiento automático de emociones es un área clave de la computación afectiva, con aplicaciones en interacción humano-computadora, salud mental y personalización de sistemas inteligentes. En este contexto, las señales fisiológicas, como el ritmo cardíaco, han cobrado relevancia debido a su relación con el estado emocional y la posibilidad de ser registradas de manera no invasiva mediante dispositivos portátiles. Este estudio propone un modelo híbrido basado en Programación Genética y Redes Neuronales para el reconocimiento de emociones a partir del ritmo cardíaco. Se llevó a cabo un experimento controlado para recopilar datos de cuatro emociones: calma, enojo, felicidad y tristeza. Tras el preprocesamiento con un filtro de media móvil y la extracción de características, la Programación Genética transformó los datos en un espacio más adecuado para la clasificación mediante una red neuronal multicapa. El modelo alcanzó una precisión del 95% en entrenamiento y 94% en prueba, aunque se observaron dificultades en la diferenciación entre enojo y tristeza. Estos resultados demuestran la viabilidad del uso de dispositivos portátiles y metodologías híbridas para la detección de emociones, contribuyendo al desarrollo de la computación afectiva.

Evaluation of Geographic Accessibility in Areas without Road Infrastructure through GIS Simulation

Odette Alejandra Pliego-Martínez — 2025-09-26

Spatial accessibility to services allows for the assessment of people's access to opportunities, especially in disadvantaged areas. This study employs the simulation of hypothetical scenarios using Geographic Information Systems (GIS) to analyze future accessibility to healthcare services, aiming to evaluate their potential impact and contribute to long-term strategic planning that reduces spatial inequalities. The proposed methodology models hypothetical hospital locations in areas with high levels of marginalization and no road infrastructure. In this context, Euclidean distance was used as an alternative to estimate proximity since, although road-based distance provides more realistic data, its application is limited in these regions. The scenario results show that the strategic location of new healthcare facilities in disadvantaged areas can significantly improve accessibility levels in the future, thereby contributing to a better quality of life for the population.

AstraMT: Instruction-Tuned Few-Shot Assamese–English Translation with Context-Aware Prompting and Reranking

Basab Nath — 2025-09-25

Developing machine translation (MT) systems for low-resource languages such as Assamese remains challenging due to limited parallel corpora and morphological complexity. Recent instruction-tuned large language models (LLMs) offer few-shot translation capabilities, but static prompt-based methods often yield suboptimal performance in real world scenarios. This paper introduces AstraMT, a modular pipeline for Assamese–English few-shot translation using LLMs. AstraMT incorporates a context-aware prompt selector (CAPS), syntactic prompt templates, multi-output reranking based on BLEU and COMET scores, and a lightweight post-editing module that corrects named entity errors and auxiliary omissions. The framework was evaluated on two datasets: the FLORES-200 devtest set and a manually aligned subset of the Samanantar corpus. AstraMT achieved BLEU improvements of up to +3.2 and COMET gains of +0.07 over static few-shot prompting. The AstraMT-Mixtral variant reached a BLEU of 23.0 on FLORES-200 and 21.3 on Samanantar, outperforming the supervised IndicTrans2 baseline. Qualitative and error analyses further highlighted AstraMT’s ability to generate fluent and semantically accurate translations. These results demonstrate that AstraMT provides an effective and extensible framework for LLM based translation in low-resource settings and can generalize across different LLMs without additional fine-tuning.

A Graph-based Word Segmentation Algorithm for Dialectal Arabic: Libyan Dialect as a Case Study

Husien Alhammi — 2025-09-25

Arabic language and its dialects both have a very rich and complex morphology, and they face the same challenge, which is called agglutination, where the words might be attached to one or more affixes. However, word segmentation has become a very important preprocessing procedure for many natural language processing tasks that deal with agglutinative languages to improve their performance. Besides Arabic, Arabic dialects are known for their complex agglutination system, which makes word segmentation challenging. To address this challenge, this paper presents an out-of-context full word segmentation algorithm that is based on weighted directed graph theory. The main purpose of this algorithm is to tackle the agglutination phenomena observed in dialectal Arabic. To illustrate the efficacy of the algorithm, the Libyan dialect is selected as a case study for testing its feasibility. A test dataset of 1,200 Libyan dialect words was used to manually evaluate the algorithm for accuracy. The experimental results show that the proposed algorithm achieves good out comes on the test dataset.

An Optimized Workflow for Odia Handwritten Character Recognition

Pragnya Ranjan Dash — 2025-09-25

Classifying handwritten Odia scripts is challenging because of the script’s complex character shapes and the lack of large annotated datasets. Odia is a low-resource language with only limited digital materials, making the development of effective recognition systems important for improving access and ensuring fair digital representation. The present study addresses the classification of handwritten Odia data, including basic characters, digits, and a set of frequently used compound characters. The proposed method combines several preprocessing steps with a lightweight Convolutional Neural Network (CNN), and data augmentation is applied to enrich the training samples and reduce overfitting. To evaluate the approach, four benchmark datasets were used: NITROHCS V1.0 (basic characters), ISI Kolkata (digits), IIT Bhubaneswar (digits and characters), and IIITBOdiaV2 (digits and characters). The model was trained on one dataset and tested on the others to assess adaptability. Additional evaluation was performed on real handwritten data consisting of both characters and digits. The experimental results demonstrate the effectiveness of the CNN model, showing an accuracy that either surpasses or closely matches earlier proposed systems using the same dataset.

Edu Vault: An Interactive, Multilingual, and Intelligent Topic-Conscious Video Discovery System for Enhanced Conceptual Learning Using Advanced NLP Techniques

Lopa Mandal — 2025-09-25

The present work developed an intelligent topic-conscious video discovery system to retrieve videos from YouTube to enhance e-learning. Speech recognition and machine translation techniques have been used to transform educational videos into easy-to-understand, organized content. The platform supports multilingual content and can transcribe, translate, summarize, and illustrate concepts in an effective manner. It also calculates the readability score of the extracted documents to ensure learners’ understanding. The platform uses live data from YouTube, is 93\% accurate, and responds quickly to search queries, in less than a second. The effective management of large data is handled by the four tier Command Query Responsibility Segregation (CQRS) architecture. Using their API simplifies the link up to YouTube and Google translate. This innovative approach provides solutions towards earing language barrier, saves learners time by helping them discover their needs quickly, and simplifies understanding of difficult subjects.

Detection of Tendency to Depression through Text Analysis

Lauro Reyes-Cocoletzi — 2025-09-25

A project is proposed with the objectiveof detecting tendencies toward depression throughtext analysis, using Natural Language Processingtechnologies and Large Language Models (LLM). Thedevelopment included several phases, such as theselection and preprocessing of English transcripts fromthe low-resource Distress Analysis Interview Corpus Wizard of Oz (DAIC-WOZ) dataset [18, 19], as well asthe training of models based on Transformer architectures, specifically Bidirectional Encoder Representationsfrom Transformers (BERT), Robustly Optimized BERTApproach (RoBERTa), and Decoding-enhanced BERTwith Disentangled Attention (DeBERTa). The resultshighlight the performance of the BERT fine-tuningmodel, which achieved better metrics compared to theother architectures evaluated (RoBERTa and DeBERTafine-tuning models), with an average F1 score of0.76 and a consistently high Receiver OperatingCharacteristic – Area Under the Curve (ROC-AUC) value> 0.82. This demonstrates its ability to balance precisionand sensitivity, as well as identify linguistic patternsassociated with depressive symptoms.

Named Entity Recognition (NER) for Sino-Tibetan Languages: A Comprehensive Review and Status

Jinia Angeline Gayary — 2025-09-25

As technology continues to advance at a rapid pace, there is a growing interest in Natural Language Processing (NLP) tools and applications. However, creating NLP tools that can effectively process natural languages presents numerous difficulties. One crucial aspect of NLP is Named Entity Recognition (NER), which involves identifying and classifying named entities in a text based on their surrounding context. Although there has been extensive research, NER tagging still struggles to accurately tag unfamiliar named entities. NER for Sino-Tibetan languages, such as Bodo and Myanmar, poses various challenges, including word segmentation, lack of resources, and ambiguity. In this paper, we review the state-of-the-art in NER for Sino-Tibetan Languages, focusing on the methods, datasets, and performances achieved. We also highlight underlying issues and future directions for NER research in this domain. Although there are not many works on NER related to Sino-Tibetan languages available, we tried to cover a good number of papers with a wide spectrum of languages, so that this review could be best utilised by researchers interested in NER studies and development for language technologies for languages from this group. As many as different works on Sino-Tibetan NER studies have been covered. We also tried to cover NER works with a variety of approaches and techniques ranging from rule-based to machine learning, deep learning, hybrid, and cross-lingual methods and highlighting their relevance towards the specific linguistic demands of Sino-Tibetan languages. Apart from these, we also reviewed a brief status on the NLP tasks for low-resourced languages, Bodo and Assamese. We have analyzed and presented in a structured way all the approaches, methods used, along with datasets, performances and challenges encountered. We hope that this paper can provide a comprehensive overview and a useful resource for the research community interested in NER for Sino-Tibetan languages.

Yuhmu Database: A Corpus of Tonal Speech Lacking Conventional Writing

Eric Ramos-Aguilar — 2025-09-28

This paper presents the development and analysis of a digital audio database of words pronounced in Yuhmu, a tonal and endangered variant of the Otomi language spoken in Ixtenco, Tlaxcala, Mexico. The database is composed of over 8,000 word recordings, including both correct and incorrect pronunciations, which were evaluated by native speakers through perceptual judgments. Statistical analyses reveal linguistic diversity in the phonetic components. Additionally, three experimental methodologies were implemented to evaluate the database: automatic segmentation of Mel-scale spectrograms using cosine distance, pronunciation classification via a multilayer perceptron, and implicit segmentation based on cosine distance thresholds. The results demonstrate good accuracy and successful detection of phonetic boundaries, which is comparable to methods applied to languages with a strong digital presence. This database constitutes a fundamental resource for the analysis of under documented tonal indigenous languages, highlighting the importance of preserving linguistic diversity. The controlled acoustic conditions and phonetic variability present in the database provide a solid foundation for future interdisciplinary studies in computational linguistics, machine learning, and language preservation.

Towards Inclusive Fact-Checking: Claim Verification in English, Hindi, Bengali, and Code-Mixed Languages

Pritam Pal — 2025-09-28

Automated claim verification has gained significant attention in recent years due to the widespread dissemination of misinformation across various digital platforms. While substantial progress has been made for high-resource languages like English, claim verification for low-resource languages and specifically for Code-Mixed texts remains largely unexplored in a multilingual country like India. In the present work, we introduce a novel multilingual dataset for claim verification, covering English, Hindi, Bengali, and Hindi-English Code-Mixed languages. The dataset is developed by engaging large language models (LLMs) as well as human annotators. The dataset contains claims, evidence passages, and veracity labels (\textit{SUPPORTS} or \textit{REFUTES}) on news headlines collected from three important domains (Politics, Healthcare, Law and Order). We proposed a rule-based baseline algorithm and a dual-encoder framework based on transformer models to effectively verify claims across diverse languages. Our results show that XLM-RoBERTa achieves the best performance for English and Code-Mix texts, while IndicBERTv2 outperforms for Hindi and Bengali, respectively. This study highlights the challenges and opportunities in multilingual and Code-Mixed claim verification, offering a step towards building inclusive, language-diverse fact-checking systems even for low resource setup.

Sentiment Analysis with Khasi Low-Resource Language through Generation of Sentiment Words using Machine Learning

Banteilang Mukhim — 2025-09-28

Sentiment Analysis is a Natural Language Processing (NLP) technique to find out the opinion and classify the opinion expressed in a text data with polarity (e.g., positive, negative and neutral). Khasi NLP is just starting to take shape, and ways back as compared to some Indian languages. Sentiment analysis with low resource language is a challenging task as the input data has limited annotated data. The proposed method suggests employing machine translation for the Khasi-English language pair to extract emotion-carrying words from Khasi text using an English emotion word dictionary. Despite the lack of specific sentiment analysis resources for Khasi, this approach enables the identification of sentiment-bearing phrases. After generation of Khasi sentiment words, a transformer-based model is considered for sentiment analysis as a validation tool.

PromptMaster: Engineering Essentials and Basic NLP Tools

Shillpi Mishrra — 2025-09-28

Prompt Engineering (PE) and Large Language Models (LLM) are important developments in Natural Language Processing (NLP) research. This technique is key for crafting effective prompts that shape how the language model behaves. It is widely applied across various NLP tasks. Research has mainly focused on creating efficient prompts to boost performance in different applications, including chatbots, sentiment analysis, and text summarization. However, some fundamental questions remain unanswered, such as whether this work supports basic NLP or linguistic research in the context of PE and LLM. In traditional NLP, basic tools and techniques, such as part-of-speech taggers, named entity recognition, and morphological analyzers, are crucial for understanding language. Developing such tools remains a challenging issue, particularly for resource-sparce languages. In this paper, we are try to address this question. We have chosen Bengali as the language and have employed language models such as ChatGPT to tackle these challenges. For our experiments, we used publicly available datasets and the results were surprising when compared with the latest state-of-the-art models. We also identified the need to develop new prompts to fulfill basic requirements.

WavLM-Based Automatic Pronunciation Assessment for Yuhmu Speech: A Low-Resource Language

Eric Ramos-Aguilar — 2025-09-28

This paper presents an approach to classify correct and incorrect pronunciation in Yuhmu, an endangered Indigenous Minority Language, using acoustic embeddings combined with SVM and MLP models. Unlike typical low-resource language tasks focused on automatic speech recognition (ASR) or machine translation, this work employs deep acoustic representations to detect phonetic quality, achieving high accuracy and consistency across different embedding sizes. The results highlight the potential of leveraging labeled audio data and advanced speech models like WavLM to provide phonetic feedback and support language revitalization. This research establishes a foundation for deeper computational phonetic analysis in Yuhmu and opens avenues for future exploration in direct audio-to-audio translation, automatic phonetic segmentation, and detailed phoneme-level evaluation, contributing to the documentation and preservation of underrepresented languages.

Prompt-Based Anaphora Resolution in Large Language Models

Mridusmita Das — 2025-09-28

With the advancement of Large Language Models (LLMs), the scope of research in the Natural Language Processing (NLP) domain has significantly shifted. The LLM has context-based advanced language understanding that is suitable for various types of discourse analysis. Creating suitable prompts can effectively guide the model’s responses toward the desired outcome. Anaphora resolution is a complex problem that is highly context-dependent. This paper attempted to explore a prompt-based LLM technique for the resolution of anaphora. Our experiment used a \textbf{text-based question} prompt within the OpenAI LLM framework. The experiment is conducted in the Assamese language, initially using a rule-based system. The results are then compared with those obtained from a prompt-based approach. The main contribution of this paper is the exploration of prompt engineering techniques for anaphora resolution. The results indicate that the prompt-based approach is significantly superior to the rule-based approach.

Pre-trained Model Sentiment Analysis of Tunisian Telecommunications Operators’ Comments on Social Media

Abir Masmoudi — 2025-09-28

Sentiment analysis (SA) has emerged as a crucial computational method for extracting subjective information from text, facilitating organizations to transform unstructured opinions through actionable insights that drive strategic decision-making across domains covering from business intelligence to public policy formation [46]. Pre-training models for SA have gained significant attention for improving opinion extraction from text. In recent years, social media has become a crucial platform for customer engagement, with SA playing a key role in maintaining client loyalty. Extracting sentiments from comments and reviews is particularly challenging for under-resourced languages like the Tunisian Dialect (TD), which is written in both Arabizi and Arabic scripts. Despite advancements in SA, processing TD remains complex. In this study, BERT and CNN-Bidirectional LSTM models are employed to perform SA on unstructured data collected from Facebook. The dataset, TUNisian TElecom Sentiment Analysis (TUNTESA), consists of 27,080 Arabizi and 17,816 Arabic comments sourced from official telecommunications operators’ Facebook pages. The comments are labeled as positive, negative, or neutral. The results demonstrate high accuracy (Acc), with the BERT Arabic model achieving 0.99 and the BERT Arabizi model reaching 0.94-outperforming existing studies. These findings highlight the practical applications of SA for businesses leveraging social media interactions. By effectively analyzing sentiments, telecom operators can enhance customer satisfaction, manage relationships, and extract valuable feedback, ultimately maintaining a competitive edge.

Predicting the Future of Text: A Hybrid Approach to Next-Word Prediction

Sanjit Kumar Dash — 2025-09-28

Text input has become an integral part of modern communication, spanning from everyday conversations to formal content creation. However, manual typing is often slow and prone to errors, which has driven the need for efficient text prediction models to improve user experience and productivity. By anticipating and generating the next likely word in a sequence, next-word prediction systems contribute significantly to faster and more accurate text composition. Early approaches like N-grams established the foundational concepts but were limited in their ability to grasp complex, wide-reaching context. In the recent years, this field has been dominated by large-scale Transformer architectures, which have set new benchmarks in language understanding. However, their significant computational demands often create a barrier to deployment in resource-constrained environments such smartphones or embedded systems . This paper addresses this challenge by introducing a hybrid deep learning model that offers a predictive accuracy with computational efficiency. Our proposed architecture merges CNNs with Bi-LSTM networks. CNNs are highly effective at extracting local, spatial features from text, while Bi-LSTMs excel at learning long-range sequential dependencies. By training this model on the classic Sherlock Holmes dataset, we demonstrate its ability to achieve nearly 76\% contextual accuracy, proving it is a powerful and viable alternative for real-world applications. This work validates the effectiveness of hybrid models in creating intelligent text generation systems for tools like smart keyboards and assistive writing technologies.

Little Wins: Collecting, Preparing and Publishing Resources for Assamese Word Sense Disambiguation

Jumi Sarmah — 2025-09-28

This paper presents the creation of the Assamese Ambiguous Sense Inventory (ASI) and Sense Annotated Data Set (SeAnDa) for the Assamese Word Sense Disambiguation (WSD) task. WSD is a computational process that identifies the appropriate sense of an ambiguous term relevant to the context. In this paper, we describe the process of creating ASI and SeAnDa for the implementation of the Assamese Supervised WSD task. The ASI consists of a database of ambiguous terms with their multiple senses, and based on the ASI, a sense-annotated dataset was prepared from the Assamese raw Corpus. The ambiguous terms are extracted from the Assamese WordNet and Corpus. Currently, we have an inventory of 100 ambiguous terms with their various glosses in both Assamese and English, and a sense annotated dataset of minimal size 2K sentences. The authors have analyzed the ambiguous words considering the parameters- Parts of speech and the number of senses. It is reported that most of the ambiguous terms in the inventory are nouns, and most of the terms have binary senses. The ASI and SeAnDa acts as the preliminary resources for implementing the Assamese Supervised WSD task with Iterative learning and Hold-out evaluation strategy. We here adopted and applied the Naïve Bayes Classifier achieving an accuracy of 71\%. As Assamese is a computationally low-resourced language, these resources will assist researchers and developers in their future research purpose.

Indian Sign Language Recognition using MobileNetV2 Fine-Tuned by Transfer Learning

Sanjit Kumar Dash — 2025-09-25

Sign Language is the language used for communication involving hearing impaired and hearing disabled people that involves the movement of hands to exchange information. But even with the existence of such language, people find it difficult to communicate using the same due to its vast diversity across different regions and geographical areas of the world. For instance, ISL (Indian Sign Language) and ASL are the respective sign languages used in USA and India but they are completely different from one another from the perspective of hand signs as well as understanding. This arises the requirement for a model which provides people a basis to translate and understand ISL.The model that has been used in this work involves a pretrained model, MobileNetV2, which is further aided by fine-tuning and Transfer Learning techniques so that the model's components are reapplied to the new model thereby reducing time and computational resources. The Indian Sign Language (ISLRTC referred) dataset is employed using signs demonstrated on the ISLRTC website taken as images under different lighting conditions and backgrounds and is preprocessed and augmented thereby undergoing operations like Rescaling, Normalization, Standardization of pixels. It consists of 36 labeled classes(26 Alphabets + 10 digits) each containing a set of 1000 sample images that represent a certain gesture. The preprocessed dataset is then splitted into training and evaluation sets and the model is evaluated based on evaluation metrices that include metrices like accuracy, precision, recall and f1-scores. For better visualization purposes, confusion matrix along with graphs between accuracy and loss with epochs were plotted. An accuracy of 95.06\%, precision, recall, f1-scores of 0.9438, 0.9411, 0.9410 respectively and training time of 40 minutes concluded that transfer learning balances the performance and computational cost of the model unlike other deep learning models.