[论文解读] Comparative Performance of Machine Learning Algorithms in Cyberbullying Detection: Using Turkish Language Preprocessing Techniques

Emre Cihan Ateş, Erkan Bostancı+1 more|arXiv (Cornell University)|Jan 29, 2021

Hate Speech and Cyberbullying Detection参考文献 31被引用 15

一句话总结

本研究使用领域特定的自然语言处理技术，评估了19种机器学习算法在土耳其语社交媒体文本中的网络欺凌检测效果。轻量化梯度提升机（LGBM）以90.788%的准确率和90.949%的F1分数表现出最佳性能，证明其在土耳其语网络欺凌检测中优于其他模型。

ABSTRACT

With the increasing use of the internet and social media, it is obvious that cyberbullying has become a major problem. The most basic way for protection against the dangerous consequences of cyberbullying is to actively detect and control the contents containing cyberbullying. When we look at today's internet and social media statistics, it is impossible to detect cyberbullying contents only by human power. Effective cyberbullying detection methods are necessary in order to make social media a safe communication space. Current research efforts focus on using machine learning for detecting and eliminating cyberbullying. Although most of the studies have been conducted on English texts for the detection of cyberbullying, there are few studies in Turkish. Limited methods and algorithms were also used in studies conducted on the Turkish language. In addition, the scope and performance of the algorithms used to classify the texts containing cyberbullying is different, and this reveals the importance of using an appropriate algorithm. The aim of this study is to compare the performance of different machine learning algorithms in detecting Turkish messages containing cyberbullying. In this study, nineteen different classification algorithms were used to identify texts containing cyberbullying using Turkish natural language processing techniques. Precision, recall, accuracy and F1 score values were used to evaluate the performance of classifiers. It was determined that the Light Gradient Boosting Model (LGBM) algorithm showed the best performance with 90.788% accuracy and 90.949% F1 Score value.

研究动机与目标

• 为解决土耳其语社交媒体内容中网络欺凌检测缺乏综合研究的问题。
• 评估并比较多种机器学习算法在土耳其语文本数据集上的性能表现。
• 识别使用自然语言处理预处理技术对土耳其语网络欺凌信息分类中最有效的算法。
• 提升土耳其数字空间中自动化检测系统，以实现更安全的在线交流。

提出的方法

• 本研究应用针对土耳其语的自然语言处理技术对社交媒体文本进行预处理，包括分词、词形还原和停用词移除。
• 在整理后的土耳其语网络欺凌数据集上，训练并评估了19种多样化的机器学习分类器。
• 通过TF-IDF向量化方法进行特征提取，将文本转换为模型可输入的数值表示。
• 采用标准指标衡量性能：精确率、召回率、准确率和F1分数。
• 对所有分类器应用超参数调优以优化模型性能。
• 基于交叉验证评估结果，选定LGBM模型为表现最佳的算法。

实验结果

研究问题

• RQ1 哪种机器学习算法在检测土耳其语社交媒体文本中的网络欺凌行为时表现最佳？
• RQ2 不同自然语言处理预处理技术对土耳其语网络欺凌检测模型分类性能有何影响？
• RQ3 传统分类器与基于树的分类器在检测土耳其语网络欺凌内容方面的相对有效性如何？
• RQ4 在检测土耳其语网络欺凌时，不同模型的精确率、召回率和F1分数差异有多大？

主要发现

• 轻量化梯度提升机（LGBM）在检测土耳其语文本中的网络欺凌行为时达到最高准确率90.788%。
• LGBM还取得了90.949%的最佳F1分数，表明其在精确率与召回率之间具有良好的平衡。
• 其他表现优异的模型包括XGBoost和随机森林，其性能虽具竞争力但低于LGBM。
• 本研究证实，算法选择对土耳其语网络欺凌分类的检测性能具有显著影响。
• 针对土耳其语的自然语言处理预处理技术对提升模型泛化能力和性能至关重要。