IAES International Journal of Artificial Intelligence (IJ-AI)
Not a member yet
1769 research outputs found
Sort by
Fine-tuning multilingual transformers for Hinglish sentiment analysis: a comparative evaluation with BiLSTM
Get PDFGrowing trend of code-mixing in languages, in the form of Hinglish, greatly tests the skills of conventional sentiment analysis tools. The research contributes a fine-tuned multilingual transformer model built exclusively for classifying sentiment of Hinglish customer reviews. Drawing from pre trained BERT-base-multilingual-case architecture, the model gets transformed with the process of fine-tuning the same on synthetically prepared and balanced dataset simulating positive, negative, and neutral sentiments. Sophisticated methods like focal loss for addressing the class imbalance and mixed precision training for maximization of computational effectiveness are embedded within the training process. Experimental results suggest that the proposed method significantly captures the fine-grained linguistic patterns of code-mixed text, improving sentiment classification accuracy. The results show promising potential for enhancing customer feedback analysis in e-commerce, social media monitoring, and customer support use cases, where it is crucial to comprehend the sentiment behind code-mixed reviews
Exploring the influence of soft information from economic news on exchange rate and gold price movements
Get PDFInformation on business conditions is an important concern for market players and regulators. Hard information relates to easily validated characteristics such as production levels and employment conditions. In contrast, soft information such as consumer and public perceptions—is subjective and difficult to verify. Although previous studies on hard and soft information mainly focus on microeconomics and banking, current developments in big data and machine learning enable broader applications in financial market analysis. This study combined VADER sentiment analysis and support vector machine (SVM) classification (accuracy=85%) to analyze economic news, followed by Granger causality and multiple linear regression to examine causal effects and predictive relationships. The findings reveal that negative news sentiment and the Indonesian Rupiah (IDR) exchange rate influence each other, while positive sentiment has no causal impact on the exchange rate. Both negative and positive sentiments affect gold prices, whereas gold price movements do not influence sentiment. Regression analysis shows that negative sentiment has a stronger effect in decreasing the IDR exchange rate than positive sentiment, with the model explaining approximately 20% of the variance. Integrating sentiment and exchange rate data enhances the predictive model for gold price forecasting and highlights the asymmetric roles of positive and negative news in financial dynamics
Fine-tuning bidirectional encoder representations from transformers for the X social media personality detection
Get PDFUnderstanding personality traits can help individuals reach their full potential and has applications in various fields such as recruitment, advertising, and marketing. A widely used tool for assessing personality is Myers-Briggs type indicator (MBTI). Recent advancements in technology have allowed for research on how personalities can change based on social media use. Previous research used machine learning methods, deep learning methods, until transformers-based method. However, these previous approaches must be revised to require extensive data and a high computational load. Although transformer-based methods like bidirectional encoder representations from transformers (BERT) excel at understanding context, it still has limitations in capturing word order and stylistic variations. Therefore, this study proposed integrating fine-tuning BERT with recurrent neural networks (RNNs) consisting of vanilla RNN, long short-term memory (LSTM), and gated recurrent unit (GRU). This study also uses a BERT base fully connected layer as a comparison. The results show that the BERT base fully connected layer approach strategy has the best evaluation results in class extraversion/introversion (E/I) of 0.562 and class feeling/thinking (F/T) of 0.538. then, the BERT+LSTM approach strategy has the highest accuracy for the intuition/sensing (N/S) class of 0.543 and judging/perceiving (J/P) of 0.532.
Recognition of Indonesian sign language using deep learning: convolutional neural network-based approach
Get PDFThis study focuses on developing an automatic Indonesian sign language (SIBI) recognition system using a convolutional neural network (CNN). Sign language is essential for communication among deaf and hard-of hearing individuals, and automatic recognition helps improve accessibility and inclusivity. CNNs are chosen for their ability to learn image features automatically, eliminating manual extraction and improving classification accuracy. The SIBI dataset used contains 5,280 images of 26 letters, divided into training and validation sets. In early training, the model achieved low accuracy (3.63% training, 3.33% validation), but after five epochs, it significantly improved to 97.58% for training and 100% for validation
Semantic search-enhanced healthcare chatbot for hospital information management system using vector database and transformer models
Get PDFHealthcare chatbots are increasingly used to assist hospital staff, yet most existing systems rely on rule-based or generic machine learning (ML) approaches that lack the ability to comprehend natural language queries, while proprietary deep learning systems often incur high licensing costs. This work addresses this gap by proposing a cost-effective and scalable semantic vector retrieval solution for user intent recognition in a hospital information management system (HIMS) helpdesk chatbot. The MPNet based transformer model is employed to convert user inquiries and predefined intents into feature vectors, enabling highly accurate natural language understanding through cosine similarity retrieval within a dedicated vector database. The proposed vector search method was validated via an ablation study, achieving an accuracy of 0.70 for intent recognition, which demonstrates a significant performance gain of 28.0 percentage points over a traditional keyword-based search baseline. Usability testing across developer and doctor groups yielded an average score of 7.78 on a 10-point Likert scale. This study concludes that integrating semantic vector retrieval with a vector database is highly effective for recognizing specialized clinical intents, offering a more accurate solution that significantly reduces the manual helpdesk workload and enhances 24-hour assistance in healthcare
Multi-phase feature selection for detection of epithelial ovarian cancer using ensemble machine learning techniques
Get PDFEpithelial ovarian carcinoma is one of the most prevalent causes of death. Timely ovarian cancer diagnosis is significant for bettering patient outcomes and rates of survival. For prognostic and diagnostic evaluation of malignancies, AI-based machine learning algorithms are used. This novel technique is undoubtedly an effective tool that may aid in selecting the best course of action. The collection of data comprising 150 patients contained an extensive selection of clinical characteristics and markers of tumors. The recursive feature elimination (RFE) and correlation coefficient feature selection techniques were assimilated to pick the features for the machine learning model, such as age, CA-125, tumor laterality, size, tumor type, grade of tumor, and International Federation of Gynecology and Obstetrics (FIGO) stage. The study’s findings indicate that the base model accuracy was around 96%, sensitivity 93%, and specificity 100%. Using ensemble classification, accuracy was around 96%, sensitivity 98%, and specificity 94% for the RFE technique. By obtaining a deeper understanding of their decision-making process, explainable artificial intelligence makes sophisticated machine learning methods easier to explain. Before beginning treatment, this research offers crucial data for the diagnosis and prognosis assessment of individuals with epithelial ovarian cancer (EOC)
Kannada handwritten numeral recognition through deep learning and optimized hyperparameter tuning
Get PDFThe classification of handwritten numerals is a vital and challenging task in developing automated systems, including postal address sorting and license plate recognition. The present study elucidates a new methodology for recognizing Kannada handwritten numerals using deep learning ResNet and VGG architecture with transfer learning. The challenge in Kannada handwritten recognition is complicated structural hierarchy and large vocabulary. The major problem in deep neural networks is vanishing gradient, which can lead to degradation in character recognition, and was addressed using our new methodology using ResNet architecture. We apply the proposed ResNet method in various real-world applications and compare it with convolutional neural networks (CNN) architecture, VGG. The experiment was implemented with the Google Colab software version on a self-created dataset, with handwritten Kannada numerals fed as the input to the recognition process. Our proposed method achieved a high accuracy of 99.20% on training samples and a generalization accuracy of 97.5% on test samples, indicating our method's effectiveness in recognizing handwritten Kannada numerals
Optimizing sparse ternary compression with thresholds for communication-efficient federated learning
Get PDFFederated learning (FL) enables decentralized model training while preserving client data privacy, yet suffers from significant communication overhead due to frequent parameter exchanges. This study investigates how varying sparse ternary compression (STC) thresholds impact communication efficiency and model accuracy across the CIFAR-10 and MedMNIST datasets. Experiments tested thresholds ranging from 1.0 to 1.9 and batch sizes of 10, 15, and 20. Results demonstrated that selecting thresholds between 1.2 and 1.5 reduced total communication costs by approximately 10–15%, while maintaining acceptable accuracy levels. These findings suggest that careful threshold tuning can achieve substantial communication savings with minimal compromise in model performance, offering practical guidance for improving the efficiency and scalability of FL systems
Sentiment classification using gradient modulation and layered attention
Get PDFSentiment analysis is a technique for evaluating text to ascertain whether a statement is positive, negative, or neutral. Currently, transformer-based models capture the contextual relationships among words in a phrase and accomplish sentiment analysis in a nuanced manner via multi-head attention. This approach, with a fixed number of layers and heads, struggles to find the complex relationships between phrases and their semantic structures. To mitigate this issue, the suggested technique incorporates the graded multi head attention model (GMHA) at the base of the distilled bidirectional encoder representations from transformers (DistilBERT) model. It is employed to augment the layers and heads progressively, capturing the relationships between sentences in a sophisticated manner. By increasing the layers and heads the proposed model extracts long-term and hierarchical relationships from the sentence. Additionally, the attention sentient optimization technique is introduced, which improves model learning by giving more weight to important words in a sentence. During training, the process checks to see which words (“amazing" or "worst") get more attention and gives them more weight in the model update. This makes it easier for the model to understand important emotions. Our suggested model enhances performance in sentiment exploration, with an accuracy of 96.53%. This interpretation includes a comparison analysis with another contemporary framework
Bring your own device readiness and productivity framework: a structured partial least square approach
Get PDFBring your own device (BYOD) is defined as the practice that allow users to bring their private owned devices to organizations or institutions. BYOD bring benefits to both organizations and education sector in terms of cost efficient and productivity enhancement. However, there is a dearth of research on the determinant and impact of BYOD adoption in the context of educational sector. Therefore, the purpose of this study is to develop a BYOD readiness and impact framework based on four dimensions, namely technological readiness, individual readiness, contextual readiness, and organisational readiness. This study employed a quantitative research approach, utilizing an online survey questionnaire as the primary research instrument. Findings were analysed based on descriptive and inferential statistics using statistical package for social sciences (SPSS) version 26 and SmartPLS version 4.0. The findings shows that individual readiness, contextual readiness, and organisational readiness have a positive and significant relationship with BYOD adoption, while technological readiness proof to be an insignificant predictor. Subsequently, BYOD adoption also proven as a positive and significant predictor capable to improve user’s productivity