DynaThresh-DualSem：面向人机混合跨层级指令的轻量化语义偏差检测框架

DynaThresh-DualSem: A Lightweight Semantic Bias Detection Framework for Human-Machine Hybrid Cross-Hierarchical Command Systems

摘要: 在中心化无人机集群的人机协同系统中，跨层级指令传播的语义保真度是影响任务成败的关键因素。针对人机认知差异与自然语言的语义歧义性引发的指令转译偏差问题，本文提出层次化语义偏差检测框架，用于精确验证指挥官原始指令和无人机转译指令的一致性。该框架运用三级递进式检测模式。第一步是构建自适应分块算法，通过正则规则、动态评估和窗口参数自适应策略，对长文本段落的语义连贯性进行划分。第二步设计匹配指标，融合语义特征和结构特征，实现跨文本段落级有效对齐。第三步提出基于共享编码器的双塔模型，结合动态阈值学习机制，有效解决无人机转移指令的跨域语义偏差检测。研究在自建的DroneCMD-4006数据集上开展实验，该数据集包含4006对面向人机混合仿真场景指令。实验结果显示，本方法在长文本指令对比中准确率达到94.76%，优于现有基线方法。本研究创建“要素解耦–段落匹配–语义校验”分层检测范式，成功克服长文本对比中的信息稀释问题，还开源核心代码与基准数据集。

Abstract: In human-machine collaborative systems for centralized unmanned aerial vehicle (UAV) swarms, the semantic fidelity of cross-level command propagation is a critical factor influencing mission success or failure. Aiming at the problem of command translation bias caused by cognitive disparities between humans and machines and the semantic ambiguity of natural language, this paper proposes a hierarchical semantic bias detection framework to accurately verify the consistency between commanders’ original commands and UAVs’ translated commands. The framework employs a three-level progressive detection model. First, an adaptive block segmentation algorithm is constructed. Through regular rules, dynamic evaluation, and window parameter adaptation strategies, it divides the semantic coherence of long text paragraphs. Second, matching indicators are designed by fusing semantic features and structural features to achieve effective cross-text paragraph-level alignment. Third, a Siamese network model based on a shared ALBERT encoder is proposed. By integrating a dynamic threshold learning mechanism, it effectively addresses cross-domain semantic bias detection for UAV-translated commands. Experiments were conducted on the self-constructed DroneCMD-4006 dataset, which contains 4006 pairs of commands for human-machine hybrid simulation scenarios. The results show that this method achieves an accuracy of 94.76% in long-text command comparison, outperforming existing baseline methods. This study establishes a hierarchical detection paradigm of “element decoupling-paragraph matching-semantic verification,” successfully overcoming the information dilution problem in long-text comparison. The core code and benchmark dataset are also open-sourced.