计算机科学
组分(热力学)
可扩展性
可信赖性
可靠性(半导体)
干预(咨询)
领域(数学分析)
软件部署
风险分析(工程)
压力(语言学)
人工智能
人机交互
知识管理
构造(python库)
服务(商务)
操作员(生物学)
图形
任务分析
领域知识
机器学习
软件工程
服务提供商
过程管理
主题专家
背景(考古学)
作者
Shuhui Lyu,Chun-Hsien Chen,Ziqing Xia,Kendrik Yan Hong Lim
标识
DOI:10.1080/09544828.2026.2758062
摘要
In safety-critical work, stress-related impairments in attention and judgment can compromise operational safety, making the design of stress-management support particularly important. A central design challenge is to connect stress-state estimates with contextual information and domain knowledge to support interpretation and inform candidate intervention recommendations. This study develops an In-Context Learning (ICL)–GraphRAG framework guided by requirements for input representation, knowledge organisation, and evaluation. GraphRAG organizes domain knowledge through competency questions and retrieves relationships with their source passages for explanation and recommendation. ICL supplies stress estimates from physiological, individual, and operational information using labeled demonstrations without task-specific parameter updates. The framework is instantiated in a Vessel Traffic Service (VTS) case study using data collected over five months. Under the defined structural path criterion, two-hop graph retrieval increased complete evidence-path coverage from 43.8% to 86.6% relative to vector retrieval. Retrieval-depth comparisons reveal a trade-off between coverage and the proportion of contributing relations. With 256 demonstrations, ICL achieved an F1-score of 88.37% under record-level evaluation, exceeding conventional classifiers using the same support data. The findings distinguish knowledge access from relational recovery and show how input representation and demonstration budget condition estimator selection within the detection-to-recommendation framework.
科研通智能强力驱动
Strongly Powered by AbleSci AI