Selected work

Representative Papers

Construction of Historical Knowledge Graphs Based on BERT and Graph Neural Networks

Jun 01, 2026

This study addresses the challenges of entity and relation extraction from historical texts, where linguistic ambiguity, vague anaphora, and non-standard syntax impede accurate information retrieval. To tackle these issues, the authors propose a joint model that integrates BERT with graph neural networks (GNNs), leveraging context-sensitive semantic representations alongside relational graph learning. This approach effectively handles nested structures and implicit coreference, enabling the automatic construction of knowledge graphs from diverse unstructured historical sources such as municipal archives, parliamentary records, and historical correspondence. Experimental results demonstrate that the proposed system significantly outperforms both traditional rule-based methods and existing deep learning baselines in terms of Precision, Recall, and F1-score, thereby substantially improving the accuracy and completeness of historical knowledge graphs.

0 citationsRead paper

Research on Vision-Language Question Answering Models for Industrial Robots

May 02, 2026

This work addresses the challenges of visual-language question answering (VLQA) in human–robot interaction scenarios involving industrial robots, where semantic ambiguity, complex scene layouts, and domain-specific terminology hinder accurate understanding. To tackle these issues, the authors propose a hierarchical cross-modal fusion model that integrates region-level visual feature extraction, multi-scale encoding, syntactic parsing, and task-aware semantic attention. The model further incorporates an adaptive fusion strategy and a context-driven gating mechanism to enable fine-grained alignment between visual and linguistic signals within a unified joint reasoning space. Experimental results on the IVQA and RIF benchmarks demonstrate that the proposed approach significantly outperforms existing VLQA methods in Top-1 accuracy, semantic alignment capability, and robustness in interpreting ambiguous or procedural instructions.

0 citationsRead paper

Applications of the Transformer Architecture in AI-Assisted English Reading Comprehension

Apr 26, 2026

This study addresses the limitations of AI-assisted English reading comprehension systems—specifically, insufficient interpretability, algorithmic bias, and performance instability—by proposing an interpretable Transformer architecture that integrates advanced attention mechanisms with gradient-based feature attribution. The approach establishes a unified technical pipeline incorporating adversarial bias correction, token-level attribution analysis, and multi-head attention heatmap visualization, thereby significantly enhancing model fairness and pedagogical applicability without compromising high accuracy. Experimental results demonstrate that the proposed method outperforms state-of-the-art models in both accuracy and macro-averaged F1 score, with certain metrics approaching human-level performance. Furthermore, multi-week user studies confirm its practical effectiveness and foster strong trust among educators.

0 citationsRead paper

Recent publications

Latest Papers

Construction of Historical Knowledge Graphs Based on BERT and Graph Neural Networks

Jun 01, 2026

This study addresses the challenges of entity and relation extraction from historical texts, where linguistic ambiguity, vague anaphora, and non-standard syntax impede accurate information retrieval. To tackle these issues, the authors propose a joint model that integrates BERT with graph neural networks (GNNs), leveraging context-sensitive semantic representations alongside relational graph learning. This approach effectively handles nested structures and implicit coreference, enabling the automatic construction of knowledge graphs from diverse unstructured historical sources such as municipal archives, parliamentary records, and historical correspondence. Experimental results demonstrate that the proposed system significantly outperforms both traditional rule-based methods and existing deep learning baselines in terms of Precision, Recall, and F1-score, thereby substantially improving the accuracy and completeness of historical knowledge graphs.

0 citationsRead paper

Research on Vision-Language Question Answering Models for Industrial Robots

May 02, 2026

This work addresses the challenges of visual-language question answering (VLQA) in human–robot interaction scenarios involving industrial robots, where semantic ambiguity, complex scene layouts, and domain-specific terminology hinder accurate understanding. To tackle these issues, the authors propose a hierarchical cross-modal fusion model that integrates region-level visual feature extraction, multi-scale encoding, syntactic parsing, and task-aware semantic attention. The model further incorporates an adaptive fusion strategy and a context-driven gating mechanism to enable fine-grained alignment between visual and linguistic signals within a unified joint reasoning space. Experimental results on the IVQA and RIF benchmarks demonstrate that the proposed approach significantly outperforms existing VLQA methods in Top-1 accuracy, semantic alignment capability, and robustness in interpreting ambiguous or procedural instructions.

0 citationsRead paper

Applications of the Transformer Architecture in AI-Assisted English Reading Comprehension

Apr 26, 2026

This study addresses the limitations of AI-assisted English reading comprehension systems—specifically, insufficient interpretability, algorithmic bias, and performance instability—by proposing an interpretable Transformer architecture that integrates advanced attention mechanisms with gradient-based feature attribution. The approach establishes a unified technical pipeline incorporating adversarial bias correction, token-level attribution analysis, and multi-head attention heatmap visualization, thereby significantly enhancing model fairness and pedagogical applicability without compromising high accuracy. Experimental results demonstrate that the proposed method outperforms state-of-the-art models in both accuracy and macro-averaged F1 score, with certain metrics approaching human-level performance. Furthermore, multi-week user studies confirm its practical effectiveness and foster strong trust among educators.

0 citationsRead paper