Selected work

Representative Papers

Adaptive Unequal Error Protection for Semantic Split Learning over Wireless Channels

Aug 17, 2026

This study addresses the challenge of adapting transmission reliability to task relevance in semantic segmentation for wireless edge cloud systems. To this end, we propose a task-aware learning framework that leverages mutual information gradients to establish a task prioritization mechanism. This approach enables fully learning-driven adaptive unequal error protection and end-to-end optimization of the communication interface. Experimental evaluations on real-world IoT datasets across varying signal-to-noise ratios demonstrate that the proposed framework significantly outperforms existing baselines. Furthermore, it exhibits superior generalization capability and robustness, effectively enhancing both semantic segmentation accuracy and transmission efficiency in resource-constrained scenarios.

0 citationsRead paper

A Semantic Communication Approach to Fiducial Marker Processing in 5G-Enabled Edge SLAM

Aug 10, 2026

This work addresses the challenge of achieving efficient communication–perception co-design in 5G-enabled edge SLAM, where conventional fiducial marker detection struggles to balance accuracy and resource constraints. To this end, the paper introduces a semantic segmentation inference framework that, for the first time, integrates semantic communication principles into fiducial processing for edge SLAM. Built upon a DeepTag-inspired CNN, the framework dynamically partitions the model between the robot and an edge server, transmitting task-oriented intermediate semantic features over wireless links to unify communication and perception. Evaluated on a 5G testbed with a ROS2-based robotic platform, the approach demonstrates high-precision keypoint estimation and its positive impact on pose estimation, while also quantifying the communication–computation trade-offs across different model split points, offering practical guidance for deploying visual perception in connected robotic systems.

0 citationsRead paper

Improving customer service with automatic topic detection in user emails

Feb 26, 2025

To address low efficiency in email subject identification within customer service, this paper proposes an unsupervised Serbian-language email topic detection method based on BERTopic. It represents the first adaptation of BERTopic to a low-resource, highly inflected language setting—Serbian—by introducing a lightweight preprocessing pipeline (including lemmatization and stopword filtering) and a rule-based post-processing engine, thereby establishing a transferable end-to-end email understanding framework. The model automatically clusters incoming emails into 12 business-relevant topics and enriches each cluster with multidimensional semantic labels, enabling real-time filtering and routing. Evaluated on a test set of 100 emails, the approach achieves a topic classification accuracy of 92% with an average processing time of under 1.2 seconds per email. Deployed in production, it has improved customer service response efficiency by 40% and supports daily processing of over 20,000 emails.

0 citationsRead paper

Named entity recognition for Serbian legal documents: Design, methodology and dataset development

Feb 14, 2025

To address the scarcity of high-quality resources and domain-adapted models for Named Entity Recognition (NER) in Serbian legal texts (e.g., court judgments), this paper introduces the first expert-annotated, domain-specific NER dataset for Serbian legal language. We propose a BERT-based domain-adaptive fine-tuning approach: leveraging continued pretraining on legal corpora to enhance domain specificity, employing a bidirectional encoder architecture, and rigorously evaluating performance via stratified cross-validation. Experimental results show that our model achieves an F1 score of 0.96 on a held-out test set—substantially outperforming off-the-shelf general-purpose models—and demonstrates strong robustness against lexical perturbations and formatting variations. Our key contributions are: (1) the first open-source, expert-annotated NER benchmark dataset for Serbian legal texts; and (2) a reproducible, high-accuracy, and robust end-to-end NER solution tailored to the Serbian legal domain.

0 citationsRead paper

Towards New Benchmark for AI Alignment&Sentiment Analysis in Socially Important Issues: A Comparative Study of Human and LLMs in the Context of AGI

Jan 05, 2025

This study investigates the affective understanding and dynamic evolution mechanisms of large language models (LLMs) regarding societally critical topics such as artificial general intelligence (AGI), aiming to advance scientifically grounded affective evaluation in AI alignment. Method: Leveraging Likert-scale–based human–AI comparative experiments, we systematically assess affective tendencies and three-day temporal dynamics across seven mainstream LLMs (e.g., GPT-4, Bard) and three human cohorts. Contribution/Results: We first reveal significant heterogeneity in LLM affective distributions on AGI—alongside quantifiable temporal evolution (evolution rate differences: 1.03%–8.21%)—and find that LLMs’ mean affective scores (3.32–4.12/5) significantly exceed the human average (2.97/5), exposing latent biases and conflict-of-interest risks. We propose the “human-like but non-uniform” hypothesis for LLM affect formation and introduce the first AI affective alignment benchmark tailored to societally salient issues.

0 citationsRead paper

Recent publications

Latest Papers

Adaptive Unequal Error Protection for Semantic Split Learning over Wireless Channels

Aug 17, 2026

This study addresses the challenge of adapting transmission reliability to task relevance in semantic segmentation for wireless edge cloud systems. To this end, we propose a task-aware learning framework that leverages mutual information gradients to establish a task prioritization mechanism. This approach enables fully learning-driven adaptive unequal error protection and end-to-end optimization of the communication interface. Experimental evaluations on real-world IoT datasets across varying signal-to-noise ratios demonstrate that the proposed framework significantly outperforms existing baselines. Furthermore, it exhibits superior generalization capability and robustness, effectively enhancing both semantic segmentation accuracy and transmission efficiency in resource-constrained scenarios.

0 citationsRead paper

A Semantic Communication Approach to Fiducial Marker Processing in 5G-Enabled Edge SLAM

Aug 10, 2026

This work addresses the challenge of achieving efficient communication–perception co-design in 5G-enabled edge SLAM, where conventional fiducial marker detection struggles to balance accuracy and resource constraints. To this end, the paper introduces a semantic segmentation inference framework that, for the first time, integrates semantic communication principles into fiducial processing for edge SLAM. Built upon a DeepTag-inspired CNN, the framework dynamically partitions the model between the robot and an edge server, transmitting task-oriented intermediate semantic features over wireless links to unify communication and perception. Evaluated on a 5G testbed with a ROS2-based robotic platform, the approach demonstrates high-precision keypoint estimation and its positive impact on pose estimation, while also quantifying the communication–computation trade-offs across different model split points, offering practical guidance for deploying visual perception in connected robotic systems.

0 citationsRead paper

Improving customer service with automatic topic detection in user emails

Feb 26, 2025

To address low efficiency in email subject identification within customer service, this paper proposes an unsupervised Serbian-language email topic detection method based on BERTopic. It represents the first adaptation of BERTopic to a low-resource, highly inflected language setting—Serbian—by introducing a lightweight preprocessing pipeline (including lemmatization and stopword filtering) and a rule-based post-processing engine, thereby establishing a transferable end-to-end email understanding framework. The model automatically clusters incoming emails into 12 business-relevant topics and enriches each cluster with multidimensional semantic labels, enabling real-time filtering and routing. Evaluated on a test set of 100 emails, the approach achieves a topic classification accuracy of 92% with an average processing time of under 1.2 seconds per email. Deployed in production, it has improved customer service response efficiency by 40% and supports daily processing of over 20,000 emails.

0 citationsRead paper

Named entity recognition for Serbian legal documents: Design, methodology and dataset development

Feb 14, 2025

To address the scarcity of high-quality resources and domain-adapted models for Named Entity Recognition (NER) in Serbian legal texts (e.g., court judgments), this paper introduces the first expert-annotated, domain-specific NER dataset for Serbian legal language. We propose a BERT-based domain-adaptive fine-tuning approach: leveraging continued pretraining on legal corpora to enhance domain specificity, employing a bidirectional encoder architecture, and rigorously evaluating performance via stratified cross-validation. Experimental results show that our model achieves an F1 score of 0.96 on a held-out test set—substantially outperforming off-the-shelf general-purpose models—and demonstrates strong robustness against lexical perturbations and formatting variations. Our key contributions are: (1) the first open-source, expert-annotated NER benchmark dataset for Serbian legal texts; and (2) a reproducible, high-accuracy, and robust end-to-end NER solution tailored to the Serbian legal domain.

0 citationsRead paper

Towards New Benchmark for AI Alignment&Sentiment Analysis in Socially Important Issues: A Comparative Study of Human and LLMs in the Context of AGI

Jan 05, 2025

This study investigates the affective understanding and dynamic evolution mechanisms of large language models (LLMs) regarding societally critical topics such as artificial general intelligence (AGI), aiming to advance scientifically grounded affective evaluation in AI alignment. Method: Leveraging Likert-scale–based human–AI comparative experiments, we systematically assess affective tendencies and three-day temporal dynamics across seven mainstream LLMs (e.g., GPT-4, Bard) and three human cohorts. Contribution/Results: We first reveal significant heterogeneity in LLM affective distributions on AGI—alongside quantifiable temporal evolution (evolution rate differences: 1.03%–8.21%)—and find that LLMs’ mean affective scores (3.32–4.12/5) significantly exceed the human average (2.97/5), exposing latent biases and conflict-of-interest risks. We propose the “human-like but non-uniform” hypothesis for LLM affect formation and introduce the first AI affective alignment benchmark tailored to societally salient issues.

0 citationsRead paper