https://journal.iberamia.org/index.php/intartif/issue/feed Inteligencia Artificial 2025-08-25T11:26:22+02:00 Editor editor@iberamia.org Open Journal Systems <p style="text-align: justify;"><span style="color: #000000;"><strong><em><a style="color: #003366; text-decoration: underline;" href="http://journal.iberamia.org/" target="_blank" rel="noopener">Inteligencia Artificial</a></em></strong><span id="result_box" class="" lang="en"> is an international open access journal promoted by <span class="">the Iberoamerican Society of</span> Artificial Intelligence (<a href="http://www.iberamia.org">IBERAMIA</a>). </span></span>Since 1997, the journal publishes high-quality original papers reporting theoretical or applied advances in all areas of Artificial Intelligence. <span style="color: rgba(0, 0, 0, 0.87); font-family: 'Noto Sans', -apple-system, BlinkMacSystemFont, 'Segoe UI', Roboto, Oxygen-Sans, Ubuntu, Cantarell, 'Helvetica Neue', sans-serif; font-size: 14px; font-style: normal; font-variant-ligatures: normal; font-variant-caps: normal; font-weight: 400; letter-spacing: normal; orphans: 2; text-align: justify; text-indent: 0px; text-transform: none; white-space: normal; widows: 2; word-spacing: 0px; -webkit-text-stroke-width: 0px; background-color: #ffffff; text-decoration-thickness: initial; text-decoration-style: initial; text-decoration-color: initial; display: inline !important; float: none;">There are no fees for subscription, publication nor editing tasks<span class="VIiyi" lang="en"><span class="JLqJ4b ChMk0b" data-language-for-alternatives="en" data-language-to-translate-into="es" data-phrase-index="0">.</span></span> <span class="VIiyi" lang="en"><span class="JLqJ4b ChMk0b" data-language-for-alternatives="en" data-language-to-translate-into="es" data-phrase-index="0">Articles can be written in English, Spanish or Portuguese and <a href="https://journal.iberamia.org/index.php/intartif/about/submissions">will be subjected</a> to a double-blind peer review process.</span></span> <span class="VIiyi" lang="en"><span class="JLqJ4b ChMk0b" data-language-for-alternatives="en" data-language-to-translate-into="es" data-phrase-index="0">The journal is abstracted and indexed in several <a href="http://journal.iberamia.org/index.php/intartif/metrics">data bases</a>. </span></span><br /></span></p> https://journal.iberamia.org/index.php/intartif/article/view/1443 Analysis and Processing of Driver Behavior for Emotion Recognition 2025-04-03T21:32:42+02:00 Carlos H. Espino-Salinas carlosespino@uaz.edu.mx Huizilopoztli Luna-García hlugar@uaz.edu.mx José M. Celaya-Padilla jose.celaya@uaz.edu.mx Jorge A. Morgan-Benita alejandro.morgan@uaz.edu.mx Ana G. Sánchez-Reyna ing.agsreyna19@gmail.com Jorge I. Galván-Tejada gatejo@uaz.edu.mx Hamurabi Gamboa-Rosales hamurabigr@uaz.edu.mx <p>Road traffic injuries cause considerable economic losses to individuals, families and nations. Knowing the driver’s condition means continuously recognizing whether the driver is physically, emotionally and physiologically fit to drive the vehicle, as well as effectively communicating these situations to the driver. This research aims to collect, analyze and process behavioral signals in drivers through the interaction of the driver with the basic elements of driving to recognize different types of emotions established in the continuous model of emotional characterization proposed by Russell using emotion induction through augmented autobiographical recall and<br />machine learning algorithms, in order to generate models capable of recognizing the emotional state of drivers through a minimally invasive, objective and efficient process. With this methodology of signal analysis of driver behavior, 4 types of emotions could be recognized within the two-dimensional excitation-valence plane with an accuracy of 73% using the Random Forest algorithm. In conclusion, a first scientific perspective on the relationship between driver behavior and emotions is offered, and the most significant information signal windows for emotion identification in a simulated driving experimentation environment are successfully identified.</p> 2025-06-17T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/1674 Deepfake Detection in Manipulated Images/ Audio/ Videos: A Three-Stage Multi-Modal Deep Learning Framework 2024-08-20T14:18:40+02:00 Leema Nelson leema.nelson@gmail.com Harshita Batra batraharshita12@gmail.com Radha P. pradha@mepcoeng.ac.in <p>The proliferation of deepfake content presents a significant threat to digital integrity and necessitates the development of efficient detection techniques. This study aims to establish a three-stage framework utilizing advanced deep learning models for multimedia datasets encompassing audio, video, and image data. The initial stage comprises an XceptionNet-based image deepfake detection model developed by providing its capacity to capture subtle artifacts and inconsistencies through depth-wise separable convolutions. This model, developed using the CelebA dataset, achieved an accuracy of 95.56% for the image data. The second stage, focusing on<br />audio deepfakes, employs a novel approach combining Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, selected for their capacity to process both the spatial and temporal aspects of audio data. The hybrid CNN and LSTM achieved an accuracy of 98.5% on the DEEP-VOICE dataset. The third stage, addressing video-based deepfake detection, integrates the XceptionNet and LSTM networks, harnessing the strengths of both spatial and temporal analyses. This integrated approach yields an accuracy of 97.574% across the Forensic++, DFDC, and Celeb-DF datasets. To address class imbalances in the datasets, class weighting is employed, assigning greater weights to the minority class during training, thereby enhancing the robustness of themodel. This framework is used to develop an app for detecting deepfakes across images, audio, and video data. This study underscores the significance of deep learning architectures and comprehensive datasets for accurate deepfake detection across various media forms. By advancing detection methodologies, this research contributes to combating misinformation and safeguarding the authenticity of digital content, thus supporting the preservation of online ecosystems.</p> 2025-06-17T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/1781 Deep Learning-Based Intrusion Detection System: Embracing Long Short-Term Memory (LSTM) and Roughly Balanced Bagging Synergies 2024-11-05T20:17:17+01:00 Onuorah Martins Onyekwelu martins.onuorah@ulk.ac.rw Sun Yanxia ysun@uj.ac.za Daniel Mashao dmashao@uj.ac.za <p>This study introduces a novel approach to address class imbalance issues in network traffic datasets within a deep learning framework. We propose the implementation of roughly balanced bagging (RBB) in a long short-term memory (LSTM) architecture, using information gain (IG) to identify optimal features from an intrusion detection system (IDS) dataset exhibiting class imbalance. The approach begins with feature selection via information gain, applies RBB to create balanced subsets of the data, and then trains multiple LSTM models on these subsets to form an ensemble for improved classification of imbalanced network traffic data. Specifically, experimentation is conducted on subsets of features categorized into quartiles on the basis of their information gain, utilizing the CIC-IDS 2017 dataset. The minority class within each quartile is upsampled via the synthetic minority oversampling technique (SMOTE). Then, 10 roughly balanced bags are created from the upsampled data for classification by 10 long short-term memory (LSTM) models. This process is repeated across the first, second, and third quartiles, enabling a comprehensive analysis of feature importance and model performance across the different dataset subsets. Additionally, the dataset's 15 class labels were grouped into 7 classes on the basis of their characteristics, facilitating multiclassification tasks. Our methodology achieved an accuracy of 91.04%, precision of 91.04%, recall of 96.73%, AUC of 96.73%, and F1 score of 91.04% on binary classification using the first quartile (19) features. The performance of our methodology for multiclassification is measured by three metrics: recall, precision, and the F1 score. Class 2 has the highest recall of 98.00%, the F1 score of 92.00%, and class 3 has the highest precision of 97.00%.</p> 2025-06-17T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/1788 Evaluating the Performance of Deep Convolutional Neural Networks and Support Vector Regression for Creditworthiness Prediction in the Financial Sector 2024-11-28T14:53:25+01:00 Kianeh Kandi kianeh.kandi@alumnos.upm.es Antonio Garcia Dopico agdopico@gmail.com <p>Creditworthiness prediction plays a crucial role in the financial sector, where accurate assessments of individuals’ credit risk are essential for making informed lending decisions. In recent years, the use of advanced machine learning algorithms, such as Deep Convolutional Neural Networks (DCNNs) and Support Vector Regression (SVR), has gained traction for creditworthiness prediction tasks. These algorithms offer unique capabilities for analyzing complex financial data and extracting valuable insights to effectively assess credit risk. This study develops and compares credit risk prediction models using DCNNs and SVR, leveraging two real-world financial datasets: the Bank Churners Dataset (10,127 records, 23 features) from Kaggle and a Personal Loan Dataset (5,000 records, 14 features) with a significant class imbalance. The datasets include variables such as income, credit limit, transaction history, and loan acceptance, which are critical for assessing financial behavior. Given the imbalance in both datasets (e.g., only 16.1% of customers churned in the Bank Churners Dataset and 10% accepted loans in the Personal Loan Dataset), we apply the Synthetic Minority Over-sampling Technique (SMOTE) to balance the classes and improve model performance. Evaluation metrics, including accuracy, precision, recall, and F1-score, demonstrate that SVR outperforms DCNN across key parameters, achieving an accuracy of 0.92, F1-score of 0.95,<br />precision of 0.93, and recall of 0.97 on Dataset 1. In comparison, DCNN achieved an accuracy of 0.88, F1-score of 0.89, precision of 0.86, and recall of 0.91. On Dataset 2, while DCNN’s accuracy improved to 0.93, SVR excelled with 0.98. These findings underscore the superiority of SVR in scenarios demanding high accuracy and precision, while DCNN offers a more balanced trade-off between precision and recall. This study provides actionable insights into selecting optimal models for credit risk evaluation, contributing to the development of reliable, data-driven financial systems.</p> 2025-06-17T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/1809 Explainable Artificial Intelligence Techniques for Speech Emotion Recognition: A Focus on XAI Models 2025-01-13T17:33:44+01:00 Michael Norval mnorval@gmail.com Zenghui Wang wangz@unisa.ac.za <p>This study employs Explainable Artificial Intelligence (XAI) techniques, including SHAP, LIME, and XGBoost, to interpret speech-emotion recognition (SER) models. Unlike previous work focusing on generic datasets, this research integrates these tools to explore the unique emotional nuances within an Afrikaans speech corpus. The complexity of architectures poses significant challenges regarding model interpretability. This paper explicitly aims to bridge the gaps in existing Speech Emotion Recognition (SER) systems by integrating advanced Explainable Artificial Intelligence (XAI) techniques. The objective is to develop an Ensemble stacking model that combines CNN, CLSTM, and XGBoost, augmented by SHAP and LIME, to enhance the interpretability, accuracy, and adaptability of SER systems, particularly for underrepresented languages like Afrikaans. Our research methodology involves utilising XAI methods to explain the decision-making processes of CNN and CLSTM models in speech emotion recognition (SER) to enhance trust, diagnostic insight, and theoretical understanding. We train the models for SER using a comprehensive dataset of emotional speech samples. Post-training, we apply SHAP and LIME to these models to generate explanations for their predictions, focusing on the importance of features<br />and the models’ decision logic. By comparing the explanations generated by SHAP and LIME, we assess the efficacy of each method in providing meaningful insights into the models’ operations. The comparative study of various models in SER demonstrates their capability to discern complex emotional states through diverse analytical approaches, from spatial feature extraction to temporal dynamics. Our research reveals that XAI techniques improve the interpretability of complex SER models. This enhanced transparency builds end-user trust and provides valuable insights. This study contributes to the importance of explainability in deploying AI technologies in emotionally sensitive applications, paving the way for more accountable and user-centric SER systems.</p> 2025-06-17T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/1753 Distributed two phase intrusion detection system using machine learning techniques and underlying big data storage and processing architecture- HDFS 2024-08-30T21:23:45+02:00 Abhijit Dnyaneshwar Jadhav abhijit.jadhav29@gmail.com Vidyullatha Pellakuri pvidyullatha@kluniversity.in Ahire Prashant G. prashant.ahire@sitpune.edu.in Archana Chaugule archana.chaugule@pccoer.in Harish U. Tiwari harish.tiwari@pccoepune.org <p>It is crucial for organizations to secure their data in the internet era. The use of Intrusion Detection Systems (IDS) implies this security. Several researchers used various tools and methods to implement various IDS models. However, a few performance concerns that must be resolved are crucial from a security standpoint. The problems pertain to the IDS time efficiency referred as timeliness, accuracy as well as the fault tolerance. The proposed model of intrusion detection has two phases of detection. Every phase uses a different set of machine learning algorithms. Phase I employs Support Vector Machine (SVM) and k nearest neighbor (kNN), whereas Phase II uses Decision Tree and Naïve Bayes. This two phase detection takes care of reducing false positives and false negatives. To compensate the execution time of these four techniques, the big data environment—Hadoop Distributed File System (HDFS)—is utilized as the underlying storage and processing structure. With such arrangement of two phases, the model gives accuracy of 97.29% overall for known and unknown attacks. For known attacks it gives 99.49% and for unknown attacks it gives 96.28% accuracy in detecting intrusion. Also, the time efficiency is measured for training and testing of the model, for training with 10,000 records, it took 0.7 seconds which is very efficient as considered to existing systems. The detailed performance achievements are discussed in results section. Also, because of HDFS, it becomes distributed and fault tolerant intrusion detection system.</p> 2025-07-10T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/1794 Breast Cancer Prediction using Stacking Models & Hyperparameter Tuning 2025-01-31T10:40:59+01:00 Rahul Karmakar rkarmakar@cs.buruniv.ac.in Akhil Kumar Das dasakhi@gmail.com Debapriya Sarkar debapriyasarkar.bardhaman@gmail.com Saroj Kumar Biswas bissarojkum@yahoo.com Ardhendu Mandal am.csa.nbu@gmail.com Arijit Bhattacharya barijit@hotmail.com <p>This paper explores the application of stacking models for breast cancer detection, integrating key techniques such as data balancing, hyperparameter tuning, and feature selection. We implemented five different stacking configurations. Initially, Logistic Regression (LR) was used as the meta-classifier, while the base estimators included Decision Tree (DT), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Random Forest (RF) classifiers. In the second configuration, we reversed the roles: DT acted as the meta-classifier, with SVM, KNN, RF, and LR serving as the base estimators. In a third setup, SVM was used as the meta-classifier, with DT, LR, KNN, and RF as the base learners. Fourth, we implemented KNN as the stacking classifier, with LR, DT, SVM, and RF as the base estimators. Finally, in the fifth configuration, RF was the meta-classifier, supported by LR, DT, KNN, and SVM as base learners. The evaluation of stacking models was conducted in five phases, starting with a baseline with no adjustments, followed by applying data balancing alone, then adding hyperparameter tuning, applying Chi-square feature selection with data balancing, and finally using correlation-based feature selection with data balancing, all systematically excluding certain elements to analyze their individual impact. Among all cases, the stacking model with LR delivers the best performance, achieving an accuracy of 97.63%, precision of 97.68%, recall of 97.63%, and an F-measure of 97.63%, showcasing its exceptional reliability and balanced effectiveness. All models were evaluated using 10-fold cross-validation.</p> 2025-07-22T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/1904 Multi-Robot Exploration and Semantic Map Building: Heterogeneous Terrestrial Robots and a Drone 2025-01-04T21:21:00+01:00 Gabriel Aguilar gabriel.agauilar@cimat.mx Israel Becerra israelb@cimat.mx Rafael Murrieta-Cid murrieta@cimat.mx <p>In this work, we propose motion strategies for exploration and mapping of an unknown environment with a team of heterogeneous robots composed of two ground robots with different sensing and motion capabilities and a drone. Our proposal is based on stochastic dynamic programming with incomplete information; we improve this technique, generating a new approach that requires significantly fewer computations. We also propose new observation and motion models that generate a map that combines geometric and semantic information. For dealing with real data obtained with a video camera and a laser, we use machine learning techniques for building such a geometric-semantic map. Experiments are presented and analyzed in simulation and on real robots. We compare the approach with other strategies reported in the literature, showing that our approach requires shorter paths and fewer sensing locations to explore the environment, thus demonstrating the effectiveness of this approach.</p> 2025-07-22T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/1820 M2ANET: Mobile Malaria Attention Network for efficient classification of plasmodium parasites in blood cells 2024-10-25T22:52:23+02:00 Salam Ahmed salamah@garmian.edu.krd Peshraw Salam Abdalqadir pesharw@spu.edu.iq Shan Ali Abdullah shan@spu.edu.iq Yunusa Haruna yunusa2k2@buaa.edu.cn <p>Malaria is a life-threatening infectious disease caused by Plasmodium parasites, which poses a significant public health challenge worldwide, particularly in tropical and subtropical regions. Timely and accurate detection of malaria parasites in blood cells is crucial for effective treatment and control of the disease. In recent years, deep learning techniques have demonstrated remarkable success in medical image analysis tasks, offering promising avenues for improving diagnostic accuracy. However, limited studies focus on hybrid mobile models due to the complexity of combining two distinct architectures and the significant memory demand of self-attention mechanisms, especially for edge devices. In this study, we introduce (Mobile Malaria Attention Network), a hybrid model integrating MBConv3 (MobileNetV3 blocks) for efficient local feature extraction and a modified global-MHSA (multi-head self-attention) mechanism for capturing global context in blood cell images. Experimental results on the Malaria Cell Images Dataset show that achieves a top-1 accuracy of 95.45% and a Cohen Kappa score of 0.91, outperforming some state-of-the-art lightweight and mobile networks. These results highlight its effectiveness and efficiency for malaria diagnosis. The development of demonstrates the potential of hybrid mobile models for improving malaria diagnosis in resource-constrained settings.</p> 2025-07-24T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/1833 Estimating the severity of coffee leaf rust using deep learning and image processing 2025-06-10T10:51:20+02:00 Juan José Zuñiga Cajas juanjosezuniga@unicauca.edu.co Oscar Daniel Peña Ramos poscar@unicauca.edu.co Emmanuel Lasso eglasso@unicauca.edu.co Jacques Avelino jacques.avelino@cirad.fr Juan Carlos Corrales jcorral@unicauca.edu.co Cristhian Figueroa cfigmart@unicauca.edu.co <p>The global coffee industry faces significant challenges from crop diseases, of which coffee leaf rust (CLR)<br />caused by the fungus Hemileia vastatrix, stands out as one of the most damaging. Accurate assessment of disease severity is essential for applying effective control strategies. In response to this need, this study introduces a modern approach using deep learning and image processing techniques to identify and quantify CLR injury automatically. We developed thirteen models using convolutional neural networks, to classify lesions into different degrees of severity. It offers a promising alternative to conventional methods, especially under data-limited conditions, although some limitations remain in robustness across datasets. Manual rust detection requires close visual inspection of leaves, a laborious and error-prone process, especially in large cultivation areas. This challenge makes it harder to apply timely and effective disease management strategies.</p> 2025-09-02T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/1956 Firefly-Based Segmentation and Residual Deep Learning for Multi- Class Diabetic Retinopathy Detection 2025-08-25T11:26:22+02:00 Prathibha S prathijournal@gmail.com Madhusudhan K.N. madhusudhankn.ece@bmsce.ac.in Victor Ikechukwu Agughasi victor.agughasi@gmail.com <p>In recent years, the rise in diabetic retinopathy cases has posed significant challenges to existing computeraided diagnosis (CAD) solutions. These systems often focus on detecting specific abnormalities, such as microaneurysms, exudates, or hemorrhages, rather than providing a comprehensive diagnostic approach. Moreover, state-of-the-art deep learning-based methods face critical limitations, including a lack of contextual understanding, gradient vanishing/explosion issues, and failure to address class imbalance at the instance level, which impacts multi-class classification accuracy. To overcome these challenges, a novel diabetic retinopathy prediction model is proposed, leveraging firefly heuristic segmentation and residual deep spatio-textural feature learning. Instead of processing entire fundus images, the model applies Firefly heuristic-driven Fuzzy C-Means (FFCM) clustering to segment regions of interest (ROIs) corresponding to microaneurysms, exudates, and hemorrhages. Residual deepspatio-textural features are then extracted using Gray-Level Co-occurrence Matrix (GLCM), ResNet50, and AlexNet. These complementary features enhance diversity and heterogeneity, which are further processed using random forest learning. The proposed model achieves outstanding performance, with an average accuracy of<br />99.77%, precision of 99.88%, recall of 99.64%, F-measure of 99.75%, sensitivity of 99.64%, and specificity of 99.86%, surpassing existing approaches. FFCM mitigates the class imbalance problem, ResNet50 addresses gradient challenges, and AlexNet contributes high-dimensional features, ensuring robust and scalable diagnostics. This innovative solution demonstrates exceptional generalizability and runtime efficiency, offering a cost-effective, comprehensive CAD tool for diabetic retinopathy detection. </p> 2025-09-10T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/2080 NAIDS4IoT: A Novel Artificial Intelligence-Based Intrusion Detection Architecture for the Internet of Things 2025-07-21T13:51:02+02:00 Haythem Hayouni hayouni.haythem.isi@gmail.com Leila Nasraoui leila.nasraoui@supcom.tn <p>Internet of Things (IoT) has brought unprecedented opportunities across various sectors, including healthcare, transportation, industrial automation, and smart cities. However, this expansion has also introduced significant security vulnerabilities due to the heterogeneous nature, limited computational capabilities, and large-scale deployment of IoT devices. Detecting anomalies, which often signify security breaches or system malfunctions, is crucial to maintaining the integrity and reliability of IoT systems. Traditional anomaly detection methods, typically rule based or signature driven, struggle to adapt to evolving threats and diverse data patterns in IoT networks. This paper proposes a novel architecture named NAIIDS4IoT (Novel Artificial Intelligence-based Intrusion Detection System architecture for IoT), designed to provide efficient, accurate, and scalable anomaly detection using Artificial Intelligence. The core of NAIIDS4IoT lies in the integration of federated learning with deep autoencoders, enabling decentralized model training across edge devices without sharing raw data, thereby preserving user privacy and reducing communication overhead. Each edge node independently learns patterns of normal behavior and identifies anomalies based on reconstruction errors. A global model is continuously refined through collaborative learning across nodes. Furthermore, NAIIDS4IoT incorporates lightweight encryption and blockchain based model integrity verification to enhance security and trust in the detection process. Experimental validation using real-world IoT datasets demonstrates that NAIIDS4IoT achieves high detection accuracy, low false positive rates, and strong adaptability to dynamic environments, significantly outperforming conventional centralized and shallow learning based solutions. This architecture represents a significant step toward intelligent, autonomous, and privacy-preserving anomaly detection in next generation IoT ecosystems.</p> 2025-09-10T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors https://journal.iberamia.org/index.php/intartif/article/view/2119 Generating a Culturally and Linguistically Adapted Word Similarity Benchmark for Yucatec Maya 2025-05-13T17:08:59+02:00 Alejandro Molina-Villegas amolina@centrogeo.edu.mx Joel Suro-Villalobos jsv2858@gmail.com Jorge Reyes-Magaña jorge.reyes@correo.uady.mx Silvia Fernandez-Sabido sfernandez@centrogeo.edu.mx <p>In the field of AI, word embedding models have proven to be one of the most effective methods for capturing semantic and syntactic relationships between words, enabling significant advancements in natural language processing. However, producing word embeddings for low-resource indigenous languages—such as Yucatec Maya—often suffers from poor reliability due to limited data availability and unsuitable evaluation benchmarks.<br />In this work, we propose a novel methodology for constructing reliable word embeddings by adapting the Swadesh List for semantic similarity evaluation. Our approach involves translating the Swadesh List from a high-resource pivot language into the target language, applying linguistic and cultural filtering, and correlating similarity scores between pivot-language embeddings from large language models and target-language embeddings. Our results demonstrate that this method produces reliable and interpretable embeddings for Yucatec Maya. Furthermore, our analysis provides compelling evidence that the choice of evaluation benchmark has a far greater impact on reported performance than hyperparameter optimization.<br />This approach establishes a robust new framework with the potential to be adapted for improving word embedding generation in other low-resource languages.</p> 2025-09-25T00:00:00+02:00 Copyright (c) 2025 Iberamia & The Authors