Selected work

Representative Papers

Extreme-value forest fire prediction A study of the Loss Function in an Ordinality Scheme

Jan 06, 2026arXiv.org

This study addresses the challenge of predicting extreme wildfires, whose spatial distribution and intensity exhibit severe imbalance, by proposing the first ordinal classification framework tailored for operational wildfire risk decision-making in France. The approach integrates intensity ordering, data imbalance, and seasonal risk through ordinal-aware loss functions—specifically WKLoss and a newly introduced Truncated Discrete Exponential Generalized Pareto Distribution (TDeGPD)—aligned with fire intensity levels. Large-scale benchmarking on real-world data across multiple neural architectures demonstrates that WKLoss achieves an IoU improvement of over 0.1 on the most extreme intensity class while maintaining good calibration, confirming that ordinal supervision significantly outperforms conventional cross-entropy. Nevertheless, prediction of extremely rare events remains constrained by severe sample scarcity.

1 citationsRead paper

Annotating Scientific Uncertainty: A comprehensive model using linguistic patterns and comparison with existing approaches

Mar 14, 2025

This study addresses the automatic detection of epistemic uncertainty expressed by authors in scientific texts, supporting information retrieval and scientific text mining. We propose UnScientify, a multi-stage weakly supervised pipeline that integrates span-based pattern matching, dependency parsing, authorial reference verification, and a rule-driven uncertainty pattern lexicon—designed for resource-constrained settings where interpretability is paramount. Empirical evaluation demonstrates that this lightweight, domain-adaptive approach achieves an accuracy of 0.808, significantly outperforming mainstream large language models. Its core contribution lies in the principled integration of structured linguistic knowledge with weak supervision, effectively balancing lexical and syntactic variability in uncertainty expression while ensuring annotation transparency. UnScientify establishes a novel paradigm for scientific uncertainty modeling that jointly optimizes predictive performance and model interpretability.

0 citationsRead paper

Unlocking the Potential of Generative AI through Neuro-Symbolic Architectures: Benefits and Limitations

Feb 16, 2025

This study addresses key bottlenecks in generative AI—limited generalization, weak reasoning capability, poor interpretability, and low data efficiency. To this end, we propose a novel bidirectional coupled neuro-symbolic architecture (Neuro>Symbolic<Neuro), which seamlessly integrates deep learning with symbolic reasoning. Methodologically, we design a unified collaborative framework that synergistically incorporates retrieval-augmented generation, graph neural networks, reinforcement learning, and multi-agent systems, enabling dynamic, bidirectional interaction and mutual enhancement between neural and symbolic modules. Experimental results demonstrate substantial improvements in generalization, structured reasoning, interpretability, and data efficiency on complex tasks such as logical reasoning and trustworthy content generation. The architecture achieves state-of-the-art performance across multiple benchmarks, establishing a new paradigm for trustworthy, efficient, and scalable generative AI.

0 citationsRead paper

Recent publications

Latest Papers

Extreme-value forest fire prediction A study of the Loss Function in an Ordinality Scheme

Jan 06, 2026arXiv.org

This study addresses the challenge of predicting extreme wildfires, whose spatial distribution and intensity exhibit severe imbalance, by proposing the first ordinal classification framework tailored for operational wildfire risk decision-making in France. The approach integrates intensity ordering, data imbalance, and seasonal risk through ordinal-aware loss functions—specifically WKLoss and a newly introduced Truncated Discrete Exponential Generalized Pareto Distribution (TDeGPD)—aligned with fire intensity levels. Large-scale benchmarking on real-world data across multiple neural architectures demonstrates that WKLoss achieves an IoU improvement of over 0.1 on the most extreme intensity class while maintaining good calibration, confirming that ordinal supervision significantly outperforms conventional cross-entropy. Nevertheless, prediction of extremely rare events remains constrained by severe sample scarcity.

1 citationsRead paper

Annotating Scientific Uncertainty: A comprehensive model using linguistic patterns and comparison with existing approaches

Mar 14, 2025

This study addresses the automatic detection of epistemic uncertainty expressed by authors in scientific texts, supporting information retrieval and scientific text mining. We propose UnScientify, a multi-stage weakly supervised pipeline that integrates span-based pattern matching, dependency parsing, authorial reference verification, and a rule-driven uncertainty pattern lexicon—designed for resource-constrained settings where interpretability is paramount. Empirical evaluation demonstrates that this lightweight, domain-adaptive approach achieves an accuracy of 0.808, significantly outperforming mainstream large language models. Its core contribution lies in the principled integration of structured linguistic knowledge with weak supervision, effectively balancing lexical and syntactic variability in uncertainty expression while ensuring annotation transparency. UnScientify establishes a novel paradigm for scientific uncertainty modeling that jointly optimizes predictive performance and model interpretability.

0 citationsRead paper

Unlocking the Potential of Generative AI through Neuro-Symbolic Architectures: Benefits and Limitations

Feb 16, 2025

This study addresses key bottlenecks in generative AI—limited generalization, weak reasoning capability, poor interpretability, and low data efficiency. To this end, we propose a novel bidirectional coupled neuro-symbolic architecture (Neuro>Symbolic<Neuro), which seamlessly integrates deep learning with symbolic reasoning. Methodologically, we design a unified collaborative framework that synergistically incorporates retrieval-augmented generation, graph neural networks, reinforcement learning, and multi-agent systems, enabling dynamic, bidirectional interaction and mutual enhancement between neural and symbolic modules. Experimental results demonstrate substantial improvements in generalization, structured reasoning, interpretability, and data efficiency on complex tasks such as logical reasoning and trustworthy content generation. The architecture achieves state-of-the-art performance across multiple benchmarks, establishing a new paradigm for trustworthy, efficient, and scalable generative AI.

0 citationsRead paper