Research on Generative Artificial Intelligence-based Few-Shot Sentiment Identification Method for Emergency Enterprise Public Opinion
|
Li Kangjie, Master's Candidate |
Received date: 2026-03-24
Online published: 2026-09-29
Supported by
National Social Science Fund of China General Project titled "Research on the Propagation Mechanism and Governance Strategies of Artificial Intelligence-Generated Online Rumors"(25BGL298)
[Purpose/Significance] Targeting the core limitations of traditional sentiment analysis methods in emergency corporate public opinion scenarios—heavy reliance on annotated data, weak implicit sentiment recognition, and insufficient generalization—this paper examines the feasibility of Generative Artificial Intelligence (GAI) in achieving sentiment recognition and reputation risk segmentation under few-shot learning conditions, aiming to provide methodological support for emergency public opinion monitoring and early warning. [Method/Process] Using Weibo comments from the "Arc'teryx Himalayan Fireworks Incident" as empirical data, this paper constructed a single-inference dual-task output GAI sentiment analysis model integrating prompt engineering and in-context learning. A dual-process few-shot example construction scheme was adopted, and strictly controlled experiments were conducted between structured prompts and general prompts. Traditional machine learning and deep learning baseline models were incorporated for horizontal comparison, and reliability was further validated through dual example-set robustness testing. [Result/Conclusion] All three GAI models outperform the traditional baseline models in this experiment. Doubao-Seed-1.6 achieves an overall accuracy of 84.22% with a negative sentiment recall of 90.71% under structured prompts. Structured prompts outperform general prompts in both overall accuracy and negative sentiment recognition, while simultaneously outputting four core and other reputation risk labels alongside the three-class sentiment classification. Experimental results demonstrate good repeatability and robustness. This method enables sentiment classification and risk segmentation with low reliance on labeled data, providing structured data support for tiered early warning and targeted crisis response in emergency corporate public opinion management.
Li Kangjie Xie Yanxi . Research on Generative Artificial Intelligence-based Few-Shot Sentiment Identification Method for Emergency Enterprise Public Opinion[J]. Knowledge Management Forum, 2026 , 11(5) . DOI: 10.13266/j.issn.2095-5472.2026.039
表1 企业声誉风险类别Table 1 Corporate reputation risk categories |
| 风险类别 | 风险识别规范 |
|---|---|
| 生态伦理谴责 | 针对企业行为破坏自然生态、违背伦理规范的批评 |
| 品牌信誉损害 | 对品牌价值观、诚信形象的质疑与否定 |
| 责任追问诉求 | 要求企业道歉、追责或官方介入的诉求 |
| 消费抵制倾向 | 明确表达抵制购买、取消关注、退货退款等行为意向 |
| 其他风险 | 不明确指向上述4类的负向情感评论,包括风险类型混合或指向模糊的表达 |
表2 负面情感标注示例Table 2 Negative sentiment annotation examples |
| 评论内容 | 分词结果 | 情感标注依据 | 风险类别 |
|---|---|---|---|
| 这是致敬自然吗,是震慑自然吧,好傲慢的人类 | 致敬、自然、震慑、自然、傲慢的、人类 | 包含“震慑”和“傲慢”等负面性词汇,表达了对人类试图凌驾于自然之上的行为的强烈不满 | 生态伦理谴责 |
| 始祖鸟国内道歉和国外道歉说的完全不一样,双标品牌不值得信赖 | 道歉、双标、品牌、不值得、信赖 | 直接质疑品牌诚信形象,否定品牌价值观,对品牌公信力进行否定性评价 | 品牌信誉损害 |
| 看完这个视频我直接把始祖鸟旗舰店取关了,再也不会买 | 取关、旗舰店、再也不会买 | 使用“取关”和“再也不会买”,明确表达取消关注与拒绝购买的行为意向,具有典型的消费抵制特征 | 消费抵制倾向 |
| 到底有没有经过审批?相关部门应该介入彻查 | 审批、相关部门、介入、彻查 | 核心诉求为追问审批合规性并呼吁官方介入调查,属于责任追问类表达 | 责任追问诉求 |
表3 深度学习与传统模型结果Table 3 Results of deep learning and traditional models |
| 模型 | 情感倾向 | 精确率 | 召回率 | F1值 | 整体准确率 |
|---|---|---|---|---|---|
| BiLSTM-CNN-Attention | 0 | 0.883 7 | 0.870 8 | 0.877 2 | 0.773 7 |
| 1 | 0.768 2 | 0.780 8 | 0.774 5 | ||
| 2 | 0.758 3 | 0.666 7 | 0.709 6 | ||
| TF-IDF-SVM | 0 | 0.736 8 | 0.872 4 | 0.798 9 | 0.769 3 |
| 1 | 0.808 3 | 0.769 8 | 0.788 6 | ||
| 2 | 0.787 1 | 0.631 0 | 0.700 4 |
表4 结构化提示词生成式人工智能模型结果(SnowNLP示例)Table 4 Results of GAI models with structured prompts (SnowNLP examples) |
| 模型 | 情感倾向 | 精确率 | 召回率 | F1值 | 整体准确率 | 整体准确率95% CI |
|---|---|---|---|---|---|---|
| Moonshot-Kimi-K2-Instruct | 0 | 0.762 7 | 0.580 8 | 0.659 4 | 0.794 3 | [0.785 8, 0.802 7] |
| 1 | 0.880 4 | 0.911 8 | 0.895 8 | |||
| 2 | 0.731 6 | 0.825 9 | 0.775 9 | |||
| DeepSeek-V3.2-Exp | 0 | 0.668 8 | 0.819 9 | 0.736 6 | 0.802 5 | [0.794 2, 0.810 9] |
| 1 | 0.898 1 | 0.883 7 | 0.890 8 | |||
| 2 | 0.857 8 | 0.769 8 | 0.811 4 | |||
| Doubao-Seed-1.6 | 0 | 0.909 5 | 0.624 3 | 0.740 5 | 0.842 2 | [0.833 9, 0.850 5] |
| 1 | 0.852 8 | 0.902 9 | 0.876 9 | |||
| 2 | 0.806 9 | 0.907 1 | 0.854 0 |
表5 通用提示词生成式人工智能模型结果(SnowNLP示例)Table 5 Results of GAI models with general prompts (SnowNLP examples) |
| 模型 | 情感倾向 | 精确率 | 召回率 | F1值 | 整体准确率 | 整体准确率95% CI |
|---|---|---|---|---|---|---|
| Moonshot-Kimi-K2-Instruct | 0 | 0.749 6 | 0.577 9 | 0.652 3 | 0.783 9 | [0.776 6, 0.791 2] |
| 1 | 0.882 0 | 0.890 4 | 0.885 9 | |||
| 2 | 0.728 5 | 0.824 2 | 0.773 3 | |||
| DeepSeek-V3.2-Exp | 0 | 0.672 9 | 0.798 0 | 0.730 1 | 0.794 0 | [0.788 4, 0.799 5] |
| 1 | 0.897 7 | 0.870 8 | 0.884 0 | |||
| 2 | 0.857 5 | 0.696 2 | 0.768 2 | |||
| Doubao-Seed-1.6 | 0 | 0.905 9 | 0.635 0 | 0.746 5 | 0.836 3 | [0.830 7, 0.841 9] |
| 1 | 0.852 4 | 0.913 5 | 0.881 8 | |||
| 2 | 0.798 6 | 0.890 5 | 0.841 8 |
表6 企业声誉风险类别比例 (%)Table 6 Proportions of Corporate Reputation Risk Categories |
| 模型 | 生态伦理谴责 | 品牌信誉损害 | 责任追问诉求 | 消费抵制倾向 | 其他风险 |
|---|---|---|---|---|---|
| Doubao-Seed-1.6 | 14.33 | 43.84 | 12.84 | 23.46 | 5.53 |
| DeepSeek-V3.2-Exp | 14.65 | 44.27 | 11.89 | 23.25 | 5.94 |
| Moonshot-Kimi-K2-Instruct | 14.76 | 42.46 | 15.71 | 22.29 | 4.78 |
| 平均比例 | 14.58 | 43.52 | 13.48 | 23.00 | 5.42 |
表7 结构化提示词生成式人工智能模型结果(人工示例)Table 7 Results of Generative AI Models (Manual Examples) |
| 模型 | 情感倾向 | 精确率 | 召回率 | F1值 | 整体准确率 | 整体准确率95% CI |
|---|---|---|---|---|---|---|
| Moonshot-Kimi-K2-Instruct | 0 | 0.762 4 | 0.579 5 | 0.658 4 | 0.791 1 | [0.787 3, 0.794 8] |
| 1 | 0.882 3 | 0.909 3 | 0.895 6 | |||
| 2 | 0.730 8 | 0.821 9 | 0.773 6 | |||
| DeepSeek-V3.2-Exp | 0 | 0.669 4 | 0.814 7 | 0.734 9 | 0.801 2 | [0.790 0, 0.812 5] |
| 1 | 0.898 5 | 0.882 4 | 0.890 2 | |||
| 2 | 0.857 6 | 0.770 0 | 0.811 4 | |||
| Doubao-Seed-1.6 | 0 | 0.907 7 | 0.617 3 | 0.734 7 | 0.841 2 | [0.838 1, 0.844 3] |
| 1 | 0.857 3 | 0.886 9 | 0.871 7 | |||
| 2 | 0.806 5 | 0.900 4 | 0.850 9 |
| [1] |
中国互联网络信息中心 .第57 次中国互联网络发展状况统计报告[EB/OL].[2026-08-18].https://www.cnnic.net.cn/n4/2026/0304/c88-11549.html.
China Internet Network Information Center .The 57th statistical report on the development of China's internet[EB/OL].[2026-08-18].https://www.cnnic.net.cn/n4/2026/0304/c88-11549.html.
|
| [2] |
刘茂辉 .生成式人工智能时代虚假信息治理的困境与出路[J].现代商贸工业,2025(21):204-206.
|
| [3] |
江长斌,何珂,黄英辉,等 .基于生成式人工智能的高校网络舆情评论回复研究[J].情报科学,2026,44(4):42-50.
|
| [4] |
张艳丰,黄亚婷,易臣何 .生成式人工智能情境下智能生成网络舆情风险测度实证研究[J].情报资料工作,2026,47(2):19-29.
|
| [5] |
吴江,黄茜,贺超城,等 .基于引爆点理论的人工智能生成内容微博网络舆情传播与演化分析[J].现代情报,2023,43(7):145-161.
|
| [6] |
梁循,王涵玉,张森森,等 .人工智能生成内容对舆情生态的影响、挑战和治理[J].管理世界,2025,41(6):129-158.
|
| [7] |
|
| [8] |
伏虎 .多媒体网络突发事件舆情语义识别及危机响应研究[J].情报科学,2021,39(5):70-74,90.
|
| [9] |
邓君,孙绍丹,王阮,等 .基于Word2Vec和SVM的微博舆情情感演化分析[J].情报理论与实践,2020,43(8):112-119.
|
| [10] |
杨秀璋,郭明镇,候红涛,等 .融合情感词典的改进BiLSTM-CNN+Attention情感分类算法[J].科学技术与工程,2022,22(20):8761-8770.
|
| [11] |
曹宇,李天瑞,贾真,等 .BGRU:中文文本情感分析的新方法[J].计算机科学与探索,2019,13(6):973-981.
|
| [12] |
|
| [13] |
曾子明,孙守强,李青青 .基于融合策略的突发公共卫生事件网络舆情多模态负面情感识别[J].情报学报,2023,42(5):611-622.
|
| [14] |
张换香,李梦云,张景 .基于多模态信息融合的中文隐式情感分析[J].计算机工程与应用,2025,61(2):179-190.
|
| [15] |
|
| [16] |
|
| [17] |
陈郅睿,陆雪松.基于开源代码大语言模型提示的学生代码修复[J].华东师范大学学报(自然科学版),2024(5):93-103.
|
| [18] |
刘华玲,张子龙,彭宏帅 .面向闭源大语言模型的增强研究综述[J].计算机科学与探索,2025,19(5):1141-1156.
|
| [19] |
江长斌,陈子涵,黄英辉,等 .基于ChatGPT的高校突发事件网络舆情情感分析研究[J].数字图书馆论坛,2025,21(6):53-62.
|
| [20] |
殷芝玲,石晓雪,黄晓,等 .生成式人工智能技术的舆论动态与公众认知演化——基于ChatGPT的微博舆情研究[J].知识管理论坛,2025,10(4):335-347.
|
| [21] |
知微事见 .始祖鸟喜马拉雅烟花事件传播趋势[EB/OL].[2026-08-18].https://ef.zhiweidata.com/event/b95a4928a914870710092795/trend.(Zhiwei Shijian.Communication trend of the Arc'teryx Himalayan fireworks event[EB/OL].[2026-08-18].https://ef.zhiweidata.com/event/b95a4928a914870710092795/trend.)
|
/
| 〈 |
|
〉 |