«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年05期

页码:

90-96

栏目:

人工智能

出版日期:

2025-05-10

文章信息/Info

Title:

Policy Text Topic Mining Based on Improved BERTopic Model

文章编号:

1673-629X(2025)05-0090-07

作者:

王雨琪1; 刘晨1; 刘建炜2*; 蔡宏民3

1. 大规模流数据集成与分析技术北京市重点实验室(北方工业大学),北京 100144;

2. 福建幼儿师范高等专科学校,福建 福州 350007;

3. 华南理工大学,广东 广州 510640

Author(s):

WANG Yu-qi1; LIU Chen1; LIU Jian-wei2*; CAI Hong-min3

1. Beijing Key Laboratory on Integration and Analysis of Large-scale Stream Data (North China University of Technology),Beijing 100144,China;

2. Fujian Preschool Education College,Fuzhou 350007,China;

3. South China University of Technology,Guangzhou 510640,China

关键词:

自然语言处理; 主题模型; 政策文本; BERTopic; 流行度偏差

Keywords:

natural language processing; topic model; policy texts; BERTopic; popularity bias

分类号:

TP393.1

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0410

摘要:

自然语言处理技术在文本分析中的应用,显著提高了从海量数据中提取关键信息的效率。 基于自然语言处理技术的主题分析方法也在文本分析领域中取得了一定成果。 然而,由于政策文本数据具有场景复杂、文本长和头部效应等挑战,现有的主题挖掘方法生成的主题结果仍有较大的进步空间。 针对政策文本主题建模的挑战,该文基于 BERTopic 方法,引入了动态文档嵌入优化器和流行度纠偏正则化项。 分别克服了 BERTopic 只能在固定维度挖掘主题导致的普适性不足以及受到词级别的流行度偏差影响导致的主题结果同质化问题,实现了对最佳主题聚类向量维度的自动选择和对热门词汇的有效纠偏。 通过对实验分析,改进后的 BERTopic 方法在主题一致性、主题多样性和综合质量指标上均显著优于原始 BERTopic 模型及先进的神经网络主题模型;在可视化结果上,生成的主题质量也显著优于原生模型。

Abstract:

The application of natural language processing technology in text analysis has significantly improved the efficiency of extracting key information from massive data. Topic analysis methods based on natural language processing technology have also gained some success in the field of text analysis. However,due to the challenges of policy text data such as complex scenes,long text,and popularity bias effects,present topic mining approaches have a lot of opportunity for improvement. To address the aforementioned issues of policy text topic modeling, we propose a dynamic document embedding optimizer and a popularity bias regularization term based on the BERTopic approach. It respectively overcomes the lack of universality caused by the BERTopic model’s ability to mine topics in fixed di-mensions and the high homogeneity of topic results caused by word-level popularity bias,and achieves automatic optimization of the optimal topic clustering vector dimensions and selection and effective correction of hot words. Through experimental analysis of policy texts,we found that the improved BERTopic is significantly better than the original BERTopic model and the state-of-the-art models in topic consistency,topic diversity,and comprehensive quality indicators. In the visualization results,the quality of the topics generated is also significantly better than that of the native model.

常用功能

更新日期/Last Update: 2025-05-10