基于PolyLoss函数的不平衡犯罪隐语文本增强方法

doi:10.11871/jfdc.issn.2096-742X.2026.01.013

数据与计算发展前沿 ›› 2026, Vol. 8 ›› Issue (1): 158-167.

CSTR: 32002.14.jfdc.CN10-1649/TP.2026.01.013

doi: 10.11871/jfdc.issn.2096-742X.2026.01.013

基于PolyLoss函数的不平衡犯罪隐语文本增强方法

李文邦(),颜靖华^*(),董泽

中国人民公安大学，信息网络安全学院，北京 100038

收稿日期:2025-04-01 出版日期:2026-02-20 发布日期:2026-02-02
通讯作者: 颜靖华
作者简介:李文邦，中国人民公安大学，硕士研究生，研究方向为警务数据分析。
本文主要工作为开展实验和论文撰写。
LI Wenbang, is a master's student at the People’s Public Security University of China. His research interest is Police data analysis.
In this paper, he is responsible for conducting experiments and writing the paper.
E-mail: 1053058921@qq.com|颜靖华，博士，副教授，研究方向为数据警务技术等。
本文主要工作为定制研究计划。
YAN Jinghua, Ph.D., is an associate professor. Her research interests include data policing technology.
In this paper, she is responsible for formulating the research plan.
E-mail: yanjing hua@ppsuc.edu.cn
基金资助:
高校基本科研业务费项目“基于仿真的指挥效能评估”(2022JKF02038)

An Implicit Transcript Enhancement Method for Unbalanced Crimes Based on PolyLoss Function

LI Wenbang(),YAN Jinghua^*(),DONG Ze

School of Information and Network Security, People’s Public Security University of China, Beijing 100038, China

Received:2025-04-01 Online:2026-02-20 Published:2026-02-02
Contact: YAN Jinghua

摘要/Abstract

摘要：

【目的】为解决犯罪隐语数据集存在的不平衡问题，进一步提升对犯罪隐语文本的分类效果。【方法】对于长文本，提出SimPoly方法；对于短文本，提出EDAPoly方法。【结果】实验结果表明，SimPoly和EDAPoly方法在解决犯罪隐语数据集不平衡问题上显著提升了分类模型的表现，相较于未进行文本增强的基线方法，在准确率、召回率和F1分数等指标上均有明显提升。【结论】不仅为罪隐语识别技术的实际应用提供了一种有效的解决方案，也为类似的不平衡文本分类任务提供了新的思路和支持。

关键词: 犯罪隐语, 文本增强, 损失函数, 不平衡文本处理

Abstract:

[Objective] This study aims to address the imbalance issue in the criminal cant dataset and further improve the classification performance of criminal cant texts. [Method] For long texts, we present SimPoly, while for short texts, EDAPoly is proposed. [Result] Experimental results demonstrate that SimPoly and EDAPoly significantly improve classification model performance on imbalanced criminal cant datasets, achieving notable gains in accuracy, recall, and F1-score compared to baseline methods without text augmentation. [Conclusion] The proposed method not only provides an effective solution for the practical application of criminal cant recognition technology, but also offers new ideas and support for similar imbalanced text classification tasks.

Key words: criminal cant, text augmentation, loss function, unbalanced text processing

李文邦,颜靖华,董泽. 基于PolyLoss函数的不平衡犯罪隐语文本增强方法[J]. 数据与计算发展前沿, 2026, 8(1): 158-167.

LI Wenbang,YAN Jinghua,DONG Ze. An Implicit Transcript Enhancement Method for Unbalanced Crimes Based on PolyLoss Function[J]. Frontiers of Data and Computing, 2026, 8(1): 158-167, https://cstr.cn/32002.14.jfdc.CN10-1649/TP.2026.01.013.

图/表 9

图1

图2

图3

图4

表1

表2

表3

表4

表5

参考文献 27

[1]	段晨杰. 公安情报分析工作面临的困境及应对措施[J]. 情报探索, 2018(12): 87-91.
[2]	黄攀. 犯罪隐语侦查情报应用的困境与对策[J]. 中国人民警察大学学报, 2023, 39(4): 5-11.
[3]	张宁. 网络犯罪隐语的演变趋势、侦查应用困境与纾解路径[J]. 江西警察学院学报, 2024(4): 57-63.
[4]	冯冉, 陈丹蕾, 化柏林. 文本数据的增强方法研究综述[J/OL]. 数据分析与知识发现, 1-17 [2024-12-02].http://kns.cnki.net/kcms/detail/10.1478.G2.20240911.1803.004.html.
[5]	林佳瑞, 程志刚, 韩宇, 等. 基于BERT预训练模型的灾害推文分类方法[J]. 图学学报, 2022, 43(3): 530-536.
[6]	姜钰棋, 侯智文, 王一帆, 等. 社交平台不平衡文本数据处理与应用研究[J]. 计算机科学与探索, 2024, 18(9): 2370-2383. doi: 10.3778/j.issn.1673-9418.2311063
[7]	VUJIČIĆ S S, MLADENOVIĆ M. An approach to automatic classification of hate speech in sports domain on social media[J]. Journal of Big Data, 2023, 10(1): 109. doi: 10.1186/s40537-023-00766-9
[8]	施国良, 陈宇奇. 文本增强与预训练语言模型在网络问政留言分类中的集成对比研究[J]. 图书情报工作, 2021, 65(13): 96-07. doi: 10.13266/j.issn.0252-3116.2021.13.010
[9]	许楠桸, 柯圆圆, 胡晓莉. 基于增强语言表示模型的网络新闻长文本分类的研究[J]. 江汉大学学报(自然科学版), 2024, 52(4): 37-44.
[10]	傅薛林, 金红, 郑玮浩, 等. 知识增强的BERT短文本分类算法[J]. 计算机工程与设计, 2024, 45(7):2027-2033.
[11]	FENG S Y, GANGAL V, KANG D, et al. GenAug: Data Augmentation for Finetuning Text Generators[C]// Proceedings of Deep Learning Inside Out (DeeLIO): The First Workshop on Knowledge Extraction and Integration for Deep Learning Architectures. 2020: 29-42.
[12]	KARIMI A, ROSSI L, PRATI A. AEDA: An Easier Data Augmentation Technique for Text Classification[C]// Findings of the Association for Computational Linguistics:EMNLP 2021 2021: 2748-2754.
[13]	YE J, XU N, WANG Y, et al. LLM-da: Data augmentation via large language models for few-shot named entity recognition[J]. arXiv preprint arXiv: 2402. 14568, 2024.
[14]	COULOMBE C. Text Data Augmentation Made Simple By Leveraging NLP Cloud APIs[J]. arXiv preprint arXiv: 2004. 14926, 2020.
[15]	淦亚婷, 安建业, 徐雪. 基于深度学习的短文本分类方法研究综述[J]. 计算机工程与应用, 2023, 59(4):43-53.
[16]	LENG Z, TAN M, LIU C, et al. PolyLoss: A Polynomial Expansion Perspective of Classification Loss Functions[C]// International Conference on Learning Representations.
[17]	MACDERMOTT Á, MOTYLINSKI M, IQBAL F, et al. Using deep learning to detect social media ‘trolls’[J]. Forensic Science International: Digital Investigation, 2022, 43: 301446. doi: 10.1016/j.fsidi.2022.301446
[18]	RESENDE DE M R, FELIX DE B D, DE FRANCO R F, et al. A framework for detecting intentions of criminal acts in social media: A case study on twitter[J]. Information, 2020, 11(3): 154. doi: 10.3390/info11030154
[19]	李贺, 杨心苗, 沈旺, 等. 启发式图结构增强的社交媒体短文本谣言检测研究[J]. 情报理论与实践, 2025, 48(3): 151-159.
[20]	施鑫, 蒋志伟. 网络时代构建犯罪隐语识别机制的困境和出路[J]. 长春市委党校学报, 2024(4): 36-40.
[21]	DEVLIN J, CHANG M W, LEE K, et al. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding[C]// Proceedings of NAACL-HLT. 2019.
[22]	LIU Y, OTT M, GOYAL N, ET AL. ROBERTA: A Robustly Optimized BERT Pretraining Approach[J]. ArXiv Preprint ArXiv:1907.11692, 2019.
[23]	ZHANG Z, HAN X, LIU Z, et al. ERNIE: Enhanced Language Representation with Informative Entities[C]// Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019: 1441-1451.
[24]	WEI J, ZOU K. EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks[C]// Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (EMNLP).
[25]	BAI J, XUE H, JIANG X, et al. Classification and recognition of milk somatic cell images based on PolyLoss and PCAM-Reset50[J]. Mathematical Biosciences and Engineering: MBE, 2023, 20(5): 9423-9442. doi: 10.3934/mbe.2023414
[26]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]// Proceedings of the IEEE International Conference on Computer Vision. 2017: 2980-2986.
[27]	PEKER O, UYSAL F, HARDALAç F. Boost loss functions for better change detection[C]// 2022 3rd International Informatics and Software Engineering Conference (IISEC). IEEE, 2022: 1-4.

参数名称	参数值
Learning rate	1e-5
Epoch	3
pad_size	128
batch_size	128
hidden_size	768

Experiment stages	Model	Precision	Accuracy	Recall	F1
第一阶段	BERT	78.88%	65.52%	86.76%	74.66%
	BERT_RCNN	79.10%	67.16%	81.67%	73.71%
	ROBERTA	76.71%	62.41%	86.75%	72.60%
	ERNIE	78.53%	64.43%	88.22%	74.47%
	ERNIE_RCNN	78.61%	65.55%	82.49%	73.05%
第二阶段	BERT	83.97%	78.82%	92.51%	85.12%
	BERT_RCNN	83.97%	79.51%	91.14%	84.93%
	ROBERTA	82.58%	77.52%	91.81%	84.07%
	ERNIE	83.44%	77.83%	93.38%	84.90%
	ERNIE_RCNN	83.53%	78.92%	91.37%	83.64%

Mode	γ	Precision	Accuracy	Recall	F1
BERT	0.8	84.00%	80.59%	90.19%	85.12%
	0.9	84.38%	80.47%	91.41%	85.59%
	1	84.44%	79.64%	93.15%	85.87%
	1.1	84.32%	80.07%	92.05%	85.64%
BERT_RCNN	0.8	84.11%	79.17%	93.27%	85.65%
	0.9	84.32%	79.19%	93.79%	85.87%
	1	84.06%	78.68%	94.08%	85.69%
	1.1	83.85%	80.95%	89.21%	84.88%
ROBERTA	0.8	81.41%	77.59%	89.52%	83.13%
	0.9	80.79%	76.63%	89.87%	82.72%
	1	81.79%	78.10%	89.52%	83.42%
	1.1	80.67%	75.59%	90.43%	82.35%
ERNIE	0.8	84.73%	79.66%	92.49%	85.60%
	0.9	84.59%	79.70%	92.01%	85.41%
	1	84.56%	79.75%	91.83%	85.36%
	1.1	84.20%	79.72%	90.93%	84.95%
ERNIE_RCNN	0.8	84.17%	79.49%	91.29%	84.98%
	0.9	84.03%	80.26%	89.42%	84.59%
	1	84.23%	79.39%	91.65%	85.08%
	1.1	84.44%	79.52%	91.95%	85.28%

Models	Precision	Accuracy	Recall	F1	γ	Epoch
BERT	84.44%	79.64%	93.15%	85.87%	1	3
BERT_RCNN	84.32%	79.19%	93.79%	85.87%	0.9	3
ROBERTA	81.79%	78.10%	89.52%	83.42%	1	3
ERNIE	84.73%	79.66%	92.49%	85.60%	0.8	3
ERNIE_RCNN	84.44%	79.52%	91.95%	85.28%	1.1	3

Experimentstages	Models	Precision	Accuracy	Recall	F1
第二阶段	BERT	5.09%	13.30%	5.75%	10.46%
	BERT_RCNN	4.87%	12.35%	9.47%	3.26%
	ROBERTA	5.87%	15.11%	4.43%	11.47%
	ERNIE	4.91%	13.40%	5.16%	10.30%
	ERNIE_RCNN	4.92%	13.37%	8.88%	10.59%
第三阶段	BERT	0.47%	0.82%	0.64%	0.75%
	BERT_RCNN	0.35%	-0.32%	2.65%	0.94%
	ROBERTA	-0.79%	0.58%	-2.29%	-0.65%
	ERNIE	1.29%	1.83%	-0.89%	0.70%
	ERNIE_RCNN	0.91%	-0.40%	0.58%	1.64%

基于PolyLoss函数的不平衡犯罪隐语文本增强方法

An Implicit Transcript Enhancement Method for Unbalanced Crimes Based on PolyLoss Function

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 9

参考文献 27

相关文章 1

编辑推荐

Metrics

本文评价