| 梅澎,赖先进.面向公众情感分析的大型语言模型有效性与智慧城市治理策略[J].情报工程,2026,12(4):080-092 |
| 面向公众情感分析的大型语言模型有效性与智慧城市治理策略 |
| Effectiveness of Large Language Models for Public Sentiment Analysis and Smart City Governance Strategies |
| |
| DOI： |
| 中文关键词: 大型语言模型，智慧城市，技术治理 |
| 英文关键词: Large Language Model; Smart City; Technology Governance |
| 基金项目:中央党校（国家行政学院）2024 年度青年科研项目“基于数字治理平台的政府公共服务成效及满意度提升机制研究”（2024QN040）。 |
| 作者 | 单位 | | 梅澎 | 中共中央党校（国家行政学院）国家治理教研部（电子政务研究中心） 北京 100091 | | 赖先进 | 中共中央党校（国家行政学院）国家治理教研部（电子政务研究中心） 北京 100091 |
|
| 摘要点击次数: 8 |
| 全文下载次数: 7 |
| 中文摘要: |
| [目的/意义]进入数字化智能化时代，智慧城市建设可运用的数字技术更加多样。其中，大型语言模型不仅能够处理和生成自然语言，而且能够模拟和分析智慧城市中社会公众情感、提高城市公共服务质量和效率。[方法/过程]本研究以广泛使用的社交媒体平台新浪微博为数据源，采用数据挖掘技术来收集大量用户评论，并使用基于 Transformer 的双向编码器表征（Bidirectional Encoder Representations from Transformers，BERT）模型进行情感分类。通过将 BERT 模型与潜在狄利克雷分配（Latent Dirichlet Allocation，LDA）主题聚类模型和时间序列分析相结合，比较和验证了大型语言模型的情感分析功能。[结果/结论] 第一，BERT+LDA 模型取得了最佳性能，准确率和精密率分别为 90.49% 和 91.95%。该模型可以被选为评论文本情绪预测的主要工具。第二，在应用阶段，使用torch.load函数加载预训练模型并设置为评估模式，成功预测了微博评论的情绪，为智慧城市建设中政府监管和智能技术治理提供了新路径。第三，公众对大语言模型的隐私保护能力表示担忧，政府在大型语言模型和人工智能治理过程中应加强个人隐私和个人信息保护方面的监管工作。因此，基于大型语言模型，实现“用技术治理”与“对技术治理”相结合，是智慧城市建设和治理的有效策略。 |
| 英文摘要: |
| [Objective/Significance] In the era of digitalization and intelligence, there is a wider variety of digital technologies available for the construction of smart cities. Among them, large language models can not only process and generate natural language but also simulate and analyze the social public sentiment in smart cities, thereby improving the quality and efficiency of urban public services. [Methods/Process] This study uses Sina Weibo, a widely used social media platform, as the data source, employs data mining to collect a large number of user comments, and applies the Bidirectional Encoder Representations from Transformers (BERT) model for sentiment classification. By combining the BERT model with the Latent Dirichlet Allocation (LDA) topic clustering model and time-series analysis, the sentiment analysis capabilities of large language models are compared and verified. [Results/Conclusions] The results show three key findings. Firstly, the BERT+LDA model achieves optimal performance, with an accuracy of 90.49% and a precision of 91.95%, making it the preferred tool for sentiment prediction of comment texts. Secondly, during the application phase, the pre-trained model, loaded via the torch.load function and set to evaluation mode, successfully predicts the sentiment of Weibo comments, providing a new approach for government supervision and intelligent technology governance in smart city construction. Thirdly, the public expresses concerns about the privacy protections offered by large language models. Therefore, the government should strengthen its supervision over the protection of personal privacy and information in the governance of large language models and artificial intelligence. In conclusion, integrating governance through technology with governance over technology based on large language models constitutes an effective strategy for the construction and governance of smart cities. |
|
查看全文
查看/发表评论 下载PDF阅读器 |
| 关闭 |
|
|
|
