1 Rong Chunming’s Academician Workstation, University of Sanya, Sanya 572022, China
2 Shijiazhuang Campus, Army Engineering University of PLA, Shijiazhuang 050003, China
3 Key Laboratory of Big Data and Digital-Intelligent Governance for Ecological Environment of Hainan Province, University of Sanya, Sanya 572022, China
In the process of urbanization, urban risks have become increasingly complex. The monitoring of urban coupled-risk and the situational awareness face challenges such as scattered data, uneven data quality, and difficulties in correlation analysis. The urban coupled-risk monitoring and situation-awareness technology based on large language models were focused. The Baichuan2-7B model was selected and optimized through fine-tuning and human-feedback-based reinforcement learning (COOL RLHF). The experiment employed datasets such as COIG, BELLE 1.5M and other multi-source data, which were trained after being preprocessed. The results show that the accuracy rate for urban planning and construction risk judgment reaches 74.84%, while the accuracy rate for public opinion attitude judgment is 70.19%. Various urban planning risks are successfully monitored, public opinion hotspots are captured, and emotional tendencies are analyzed. The accuracy rate for predicting the risk development trend is 83.25%, and the matching degree between situational assessment and reality is 85.07%, receiving high recognition from professionals. The research results provide effective tools and references for urban safety management.
YU Ying, WANG Chun-ping, YANG Bo-xiong, FU Qiang.
Urban Coupled-risk Monitoring and Situation Awareness Technology Based on Large Language Models[J].
Science Technology and Engineering, 2026, 26(8): 3397-3405 DOI:10.12404/j.issn.1671-1815.2501943
由于生成内容的质量评价具有主观性,需结合上下文以及具体问题进行衡量,所以有必要引入人类偏好和主观意见。基于人类反馈的强化学习(reinforcement learning from human feedback, RLHF)[25]是大型语言模型领域内的一种创新方法。通过融入人类反馈,RLHF创建了代理人类偏好的奖励模型,从而通过使用近端策略优化(proximal policy optimization, PPO)为LLMs提供用于学习的奖励信号。这种方法使得模型能更好地理解和执行难以通过传统方法定义的任务。
ZhangHaoran, ChenGuoping, ZhaoJunsan. Function identification of urban “Living Production Ecology Space” based on multi-source heterogeneous data[J]. Science Technology and Engineering, 2024, 24(27): 11529-11536.
LanYuexin, ZhangLiwei, WangHuawei, et al. Abnormal perception of network public opinion and empirical research orienting risk monitoring[J]. Journal of Modern Information, 2022, 42 (3): 102-108.
WangShuai, HeWenchun, WangFudi, et al. Question answering system based on large language model integrating knowledge graph and vector retrieval[J]. Science Technology and Engineering, 2024, 24(32): 13902-13910.
[7]
OuyangM, ChengZ, MaJ, et al. Coupledurban risks: a complex systems perspective with a people-centric focus[J]. Engineering, 2025, 44: 44-50.
WuLianquan. Evaluation of emergency preparedness for terrorist attacks in China under the new situation[J]. Science Technology and Engineering, 2022, 22(12): 5089-5094.
WangZhiying, ChenXiao, LiuHanjie. An integrated method for risk monitoring and early warning of disaster information based on text data[J]. Journal of Catastrophology, 2024, 39(4): 81-87.
ZhangPu, ZhangHao, KongFeng, et al. A study on public opinion characteristics of rainstorm flooding disasters based on Sina Weibo data: take the three rainstorm flooding disasters in China in 2021 as an example[J]. Water Resources and Hydropower Engineering, 2023, 54(2): 47-59.
CuiXuelian, DongHaitao. Research on social network public opinion control strategy based on key node identification[J]. Computer Engineering and Applications, 2024, 60(20): 262-273.
[18]
KaramouzasD, MademlisI, PitasI. Public opinion monitoring through collective semantic analysis of tweets[J]. Social Network Analysis and Mining, 2022, 12: 91.
[19]
AlamM H, RyuW J, LeeS K. Joint multi-grain topic sentiment: modeling semantic aspects for online reviews[J]. Information Sciences, 2016, 339: 206-223.
TianShihai, YuYue, DengShuyu. Research on the early warning of network public opinion crisis under multiple scenarios of public emergencies[J]. Information Studies: Theory & Application, 2023, 46(1): 132-140.
HaoYu, ZhangFan, ZhangLijun. Research on the whole process model of the multiple-subject risk information perception of sudden natural disaster events: a case study of “7·20” heavy rain in Zhengzhou[J]. Journal of Institute of Disaster Prevention, 2023, 25(3): 60-72.
WangMingcheng, WangGaokai, LiYongnan. Construction of a security risk situational awareness framework based on LLM-based agents[J]. Information Studies: Theory & Application, 2024, 47(7): 190-198.
ZhangAijun, JiaLu. Risk patterns and regulations of online public opinion security in the context of ChatGPT[J]. Journal of Intelligence, 2023, 42(12): 180-187.
WangChong, ZhangYajun, WangJuan. Public perspective of generative AI in education—opinion topic and sentiment analysis of Bilibili’s bullet comments on ChatGPT[J]. Library Tribune, 2024, 44(10): 61-71.
SunYifeng, LiaoShufan, WuJiang. Research on situation awareness agent based on large models[J]. Command Control and Simulation, 2024, 46(2): 1-7.
[32]
YangA, XiaoB, WangB, et al. Baichuan 2:open large-scale language models[EB/OL]. (2025-04-17)[2025-05-04].https://arxiv.org/abs/2309.10305.
[33]
HuangY, BaiY, ZhuZ, et al. C-eval: a multi-level multi-discipline Chinese evaluation suite for foundation models[C]// Proceedings of the 37th International Conference on Neural Information Processing Systems (NIPS'23). New York, NY: Association for Computing Machinery (ACM), 2023: 62991-63010.
[34]
HendrycksD, BurnsC, BasartS, et al. Measuring massive multitask language understanding[C]// Proceedings of the 9th International Conference on Learning Representations (ICLR’21). Vienna: OpenReview.net, 2021: 155-180.
[35]
LiH, ZhangY, KotoF, et al. CMMLU: measuring massive multitask language understanding in Chinese[C]//Proceedings of the Association for Computational Linguistics: ACL 2024. Stroudsburg, PA: Association for Computational Linguistics(ACL), 2024: 11260-11285.
[36]
ZhangX, LiC, ZongY, et al. Evaluating the performance of large language models on gaokaobenchmark[EB/OL]. (2024-02-24)[2025-05-04]. https://arxiv.org/abs/2305.12474.
[37]
ZhongW, CuiR, GuoY, et al. AGIEval: a human-centric benchmark for evaluating foundation models[C]//Proceedings of the Association for Computational Linguistics: NAACL 2024. Stroudsburg, PA: Association for Computational Linguistics(ACL), 2024: 2299-2314.
[38]
SuzgunM, ScalesN, SchärliN, et al. Challenging bIG-bench tasks and whether chain-of-thought can solve them[C]//Proceedings of the Association for Computational Linguistics: ACL 2023. Stroudsburg, PA: Association for Computational Linguistics(ACL), 2023: 13003-13051.
[39]
OuyangL, WuJ, JiangX, et al. Training language models to follow instructions with human feedback[C]// Proceedings of the 36th International Conference on Neural Information Processing Systems (NIPS'22). New York, NY: Curran Associates Inc., 2022: 27730-27744.
[40]
CaiZ, CaoM, ChenH, et al. InternLM2technical report[EB/OL]. (2024-03-26)[2025-05-04]. https://arxiv.org/abs/2403.17297.
[41]
WangY, KordiY, MishraS, et al. Self-Instruct: aligning language models with self-generated instructions[C]// Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics. Stroudsburg, PA: Association for Computational Linguistics (ACL), 2023: 13484-13508.
[42]
XTunerContributors. XTuner: a toolkit for efficiently fine-tuning LLM[EB/OL]. (2025-02-21)[2025-05-04].https://github.com/InternLM/xtuner.
YuYing, WangChunping, FuQiang, et al. Survey of evaluation metrics and methods for semantic segmentation[J]. Computer Engineering and Applications, 2023, 59(6): 57-69.