Abba Suganda Girsang

Nowadays, technology is advancing with various applications, especially in text processing, such as news recommendations, sentiment analysis, automatic scoring, and language translation. In some cases, spelling errors often occur when inputting text for the translation process, necessitating text correction methods to display suggestions as a result. Therefore, the problem statement raised is about how to improve the accuracy of text correction and evaluate the translation quality at the word level after correcting input text in the context of translation from Indonesian to English. This research aims to develop and evaluate the combination of Levenshtein distance algorithm and Trie to correct input text and evaluate the translation quality at the word level after correcting text. There are various text correction methods, such as Hamming distance, Levenshtein distance, Damerau-Levenshtein distance, and N-Gram. Among several text correction methods. Levenshtein distance algorithm is commonly used to calculate the distance between texts and can be enhanced by using a Trie for more efficient computation in evaluating text correction. This research method resulted in an accuracy of 82.25% with an F1 score of 84.39%, where the developed text correction model produced a good translation using BLEU score with an increase of 1.55% after text correction.

Retrieval-augmented generation systems integrate external information to mitigate hallucinations in large language models, yet existing multimodal retrieval-augmented generation implementations struggle with heterogeneous embedding spaces from diverse sources. This study introduces an adaptive multimodal fusion architecture that addresses the critical challenge of integrating embeddings from research models and commercial application programming interfaces with varying dimensionalities (ranging from 768 to 1,536 dimensions) and training paradigms. We propose a source-agnostic framework comprising lightweight adapters, cross-modal attention mechanisms, and learned confidence calibration. The system was evaluated on 20,000 multimodal samples across image, video, audio, and text modalities, demonstrating significant performance improvements. The architecture enables end-to-end training without separate pretraining steps while maintaining flexibility for pretrained encoders. Experimental results show 7.2% improvement in Recall@10 compared to single-best source approaches (82.4% versus 89.6%), with expected calibration error of 0.048, indicating well-calibrated confidence scores. The framework achieves graceful degradation with only 3.8% performance reduction when one modality becomes unavailable and maintains 81.3% of in-domain performance in zero-shot domain transfer tasks. This work advances multimodal retrieval-augmented generation systems by enabling plug-and-play integration of heterogeneous embedding sources while preserving computational efficiency and interpretability.

Accurate classification of diplomatic cables is crucial for Mission’s evaluation and policy formulation. However, these documents often cover multiple topics, hence a multi-label classification approach is necessary. This research explores the application of pre-trained language models (CahyaBERT, IndoBERT, and MBERT) for multi-label classification of diplomatic cable executive summaries, which align with the diplomatic representation index. The study compares full-parameter fine-tuning and low-rank adaptation (LoRA) techniques using cables from 2022-2023. Results demonstrate that Indonesian-specific models, particularly the IndoBERT, outperform multilingual models in classification accuracy. While LoRA showed slightly lower performance than full fine-tuning, it significantly reduced GPU memory usage by 48% and training time by 69.7%. These findings highlight LoRA’s potential for resource-constrained diplomatic institutions, advancing natural language processing in diplomacy and offering pathways for efficient, real-time multi-label classification to enhance diplomatic mission evaluation.

Accurate classification of diplomatic cables is crucial for Mission’s evaluation and policy formulation. However, these documents often cover multiple topics, hence a multi-label classification approach is necessary. This research explores the application of pre-trained language models (CahyaBERT, IndoBERT, and MBERT) for multi-label classification of diplomatic cable executive summaries, which align with the diplomatic representation index. The study compares full-parameter fine-tuning and low-rank adaptation (LoRA) techniques using cables from 2022-2023. Results demonstrate that Indonesian-specific models, particularly the IndoBERT, outperform multilingual models in classification accuracy. While LoRA showed slightly lower performance than full fine-tuning, it significantly reduced GPU memory usage by 48% and training time by 69.7%. These findings highlight LoRA’s potential for resource-constrained diplomatic institutions, advancing natural language processing in diplomacy and offering pathways for efficient, real-time multi-label classification to enhance diplomatic mission evaluation.

Microservices architecture offers theoretical benefits for software maintainability, yet empirical validation of design patterns' effectiveness for enhancing modifiability remains limited, with practitioners lacking evidence-based guidance for pattern selection and quantified improvement expectations. This study addresses this gap by presenting a quantitative assessment of microservices design patterns' impact on modifiability through service-level analysis of 110 individual microservices across 15 open-source applications. Following Systematic Literature Review and Grey Literature analysis, ten modifiability-enhancing patterns were identified and evaluated using a novel context-sensitive selection framework derived from ISO 25002:2024 standards. The framework introduces three key theoretical contributions: the "Worst First" optimization principle for architectural refactoring prioritization, a service-level analysis methodology enabling granular pattern effectiveness assessment, and an evidence-based pattern selection model linking architectural characteristics to improvement potential. Quantitative evaluation using Change Impact Factor (CIF), Service Independence Metric (SIM), and Modifiability Index (MI) demonstrated significant improvements: CIF decreased by 49.7% (from 0.303 to 0.152), SIM increased by 14.2% (from 0.764 to 0.873), and MI improved by 17.6% (from 0.732 to 0.861). Statistical analysis confirmed significance for all metrics (CIF and MI: p<0.0001; SIM: p<0.05) with effect sizes ranging from small to large (d=0.26-0.81). Strong correlations between initial architectural characteristics and improvement magnitude (r=-0.812) validated the "Worst First" principle, where services with poorest metrics benefit most from pattern application. The resulting evidence-based framework enables context-sensitive pattern selection and provides quantified improvement expectations, moving beyond universal approaches to establish scientific foundations for architectural decision-making in microservices development.

The Incorporation of Question Answering system holds immense potential for addressing Indonesia’s educational disparities between the abundance of high school students and the limited number of teachers in Indonesia. These studies aim to enhance the Question Answering System model tailored for the Indonesian language dataset through enhancements to the Indonesian IndoBART model. Improvement was done by incorporating Longformer’s sliding windows attention mechanism into the IndoBART model, it would increase model proficiency in managing extended sequence tasks such as question answering. The dataset used in this research was TyDiQA multilingual dataset and translated the SQuADv2 dataset. The evaluation indicates that the Longformer-IndoBART model outperforms its predecessor on the TyDiQA dataset, showcasing an average 26% enhancement across F1, Exact Match, BLEU, and ROUGE metrics. Nevertheless, it experienced a minor setback on the SQuAD v2 dataset, leading to an average decrease of 0.6% across all metrics.

Closed-circuit television (CCTV) systems have generated a vast amount of visual data crucial for security and surveillance purposes. Effectively categorizing security level types is vital for maintaining asset security effectively. This study proposes a practical approach for classifying CCTV screenshot images using visual geometry group (VGG-19) transfer learning, a convolutional neural network (CNN) classification model that works really well in image classification. The task in classification compromise of categorizing screenshots into two classes: “humans present” and “no humans present.” Fine-tuning VGG-19 model attained 98% training accuracy, 98% validation accuracy, and 85% test accuracy for this classification. To evaluate its performance, we compared fine-tuning VGG-19 model with another method. The VGG-19-based fine-tuning model demonstrates effectiveness in handling image screenshots, presenting a valuable tool for CCTV image classification and contributing to the enhancement of asset security strategies.

The twitter social network is widely used to discuss all kinds of topics, including those related to politics. Analyzing online conversations on Twitter to map the popularity of political figures as candidates for the Indonesian presidential election is a popular and challenging research area. In the Twitter network, citizens can express themselves and communicate with political figures. The conversational data in Twitter is very complex, so Business Intelligence is needed to transform raw data into meaningful and useful information to see the popularity of Indonesian presidential election candidates. The analysis used is Social Network Analysis (SNA) by measuring Degree Centrality, Eigenvector Centrality, Betweenness Centrality, Closeness Centrality. The presidential candidates in this study, Ganjar Pranowo with a twitter account “ganjarpranowo”, Puan Maharani with a twitter account “puanmaharani_ri”, and Anies Baswedan with a twitter account “aniesbaswedan”. The actor "aniesbaswedan" excels in the value of degree centrality and betweenness centrality. The “aniesbaswedan” account is the actor who has the most influence on social network interactions based on the total number of interactions generated, then the account also becomes a bridge or liaison in the interactions of other actors in the network.