EAR: Entity-Aware Partitioning Approach for Retrieval-Augmented Generation Development
EAR方法通过实体感知分区解决知识密集型问题回答中源语料库如何划分的问题，减少检索词数量并提高准确性。
0 citationsRead paper
Selected work
Representative Papers
Mutation Testing of Simulink Cyber-Physical System Models: Challenges and Solutions in Practice
本文探讨了在Simulink模型中应用变异测试遇到的挑战及解决方案，特别是在需求追溯和Stateflow变异测试方面。
0 citationsRead paper
Analyzing Network Robustness via Residual Closeness
This study investigates the structural robustness of networks under node failures, with a focus on closeness centrality and its residual counterpart. By conducting the first systematic analysis of closeness and residual closeness in intermediate graphs—leveraging graph theory, line graph theory, and algorithmic design—it establishes precise relationships among the original graph, its line graph, and associated intermediate graphs. The main contributions include deriving exact closed-form expressions for closeness in intermediate graphs of several special graph classes, establishing general upper and lower bounds for residual closeness across broader graph families, and proposing an efficient algorithm for computing these measures. Experimental results demonstrate the superior performance of the proposed method, validating both its theoretical soundness and practical efficacy.
0 citationsRead paper
OkanNet: A Lightweight Deep Learning Architecture for Classification of Brain Tumor from MRI Images
This study addresses the time-consuming and error-prone nature of manual analysis of brain tumor MRI images by proposing OkanNet, a lightweight convolutional neural network (CNN) architecture for the automated classification of four categories: glioma, meningioma, pituitary tumor, and no tumor. Compared to a ResNet-50–based transfer learning approach, OkanNet achieves a competitive accuracy of 88.10% while reducing computational overhead significantly—training 3.2 times faster—whereas ResNet-50 attains a higher accuracy of 96.49% at substantially greater resource cost. This work effectively balances model efficiency and diagnostic accuracy, offering a practical solution for medical image analysis in resource-constrained settings.
0 citationsRead paper
Relation Extraction Capabilities of LLMs on Clinical Text: A Bilingual Evaluation for English and Turkish
This study addresses the challenge of evaluating large language models—primarily developed and optimized for English—on relation extraction from non-English clinical texts, where annotated resources are scarce. To bridge this gap, the authors construct the first English–Turkish parallel clinical relation extraction dataset and propose Relation-Aware Retrieval (RAR), a novel in-context example selection method that leverages contrastive learning to capture semantic correspondences at both sentence and relation levels. The work systematically evaluates various in-context learning and chain-of-thought prompting strategies against fine-tuned baselines such as PURE. Results demonstrate that prompting approaches consistently outperform fine-tuned models, with RAR achieving micro F1 scores of 0.906 and 0.888 on English and Turkish subsets, respectively, using Gemini 1.5 Flash; further gains are realized through structured reasoning with DeepSeek-V3, pushing performance to 0.918.
0 citationsRead paper
Recent publications
Latest Papers
EAR: Entity-Aware Partitioning Approach for Retrieval-Augmented Generation Development
EAR方法通过实体感知分区解决知识密集型问题回答中源语料库如何划分的问题，减少检索词数量并提高准确性。
0 citationsRead paper
Mutation Testing of Simulink Cyber-Physical System Models: Challenges and Solutions in Practice
本文探讨了在Simulink模型中应用变异测试遇到的挑战及解决方案，特别是在需求追溯和Stateflow变异测试方面。
0 citationsRead paper
Analyzing Network Robustness via Residual Closeness
This study investigates the structural robustness of networks under node failures, with a focus on closeness centrality and its residual counterpart. By conducting the first systematic analysis of closeness and residual closeness in intermediate graphs—leveraging graph theory, line graph theory, and algorithmic design—it establishes precise relationships among the original graph, its line graph, and associated intermediate graphs. The main contributions include deriving exact closed-form expressions for closeness in intermediate graphs of several special graph classes, establishing general upper and lower bounds for residual closeness across broader graph families, and proposing an efficient algorithm for computing these measures. Experimental results demonstrate the superior performance of the proposed method, validating both its theoretical soundness and practical efficacy.
0 citationsRead paper
OkanNet: A Lightweight Deep Learning Architecture for Classification of Brain Tumor from MRI Images
This study addresses the time-consuming and error-prone nature of manual analysis of brain tumor MRI images by proposing OkanNet, a lightweight convolutional neural network (CNN) architecture for the automated classification of four categories: glioma, meningioma, pituitary tumor, and no tumor. Compared to a ResNet-50–based transfer learning approach, OkanNet achieves a competitive accuracy of 88.10% while reducing computational overhead significantly—training 3.2 times faster—whereas ResNet-50 attains a higher accuracy of 96.49% at substantially greater resource cost. This work effectively balances model efficiency and diagnostic accuracy, offering a practical solution for medical image analysis in resource-constrained settings.
0 citationsRead paper
Relation Extraction Capabilities of LLMs on Clinical Text: A Bilingual Evaluation for English and Turkish
This study addresses the challenge of evaluating large language models—primarily developed and optimized for English—on relation extraction from non-English clinical texts, where annotated resources are scarce. To bridge this gap, the authors construct the first English–Turkish parallel clinical relation extraction dataset and propose Relation-Aware Retrieval (RAR), a novel in-context example selection method that leverages contrastive learning to capture semantic correspondences at both sentence and relation levels. The work systematically evaluates various in-context learning and chain-of-thought prompting strategies against fine-tuned baselines such as PURE. Results demonstrate that prompting approaches consistently outperform fine-tuned models, with RAR achieving micro F1 scores of 0.906 and 0.888 on English and Turkish subsets, respectively, using Gemini 1.5 Flash; further gains are realized through structured reasoning with DeepSeek-V3, pushing performance to 0.918.
0 citationsRead paper