- Detect All Abuse!Toward Universal Abusive Language Detection Models
- Classification Methods for Hate Speech Diffusion: Detecting the Spread of HateSpeech on Twitter
Detect All Abuse!Toward Universal Abusive Language Detection Models
介绍
ALD - abusive language detection
许多ALD的研究，数据有主观性，不能通用。
构建四种不同类型的辱骂性语言方面嵌入：directed target, generalised target, explicit content, and implicit content。使用异质图来分析每个作者的语言行为，并利用图卷积网络（GCNs）学习单词和文档嵌入
贡献：
- 发现大多数ALD算法都无法在不同的在线社区中接受不同类型的辱骂性语言方面
- 提出了一种新的ALD算法，该算法能够将多个方面的辱骂语言进行显式整合，并能检测出不同方面和不同领域的通用性辱骂语言行为
数据集
| Dataset | Source | Size | Composition |
|---|---|---|---|
| Waseem | 16.2k | Racism(11.97%), Sexism(19.43%), None(68.60%) | |
| HatEval | 13k | Hateful(42.08%), Non-hateful(57.92%) | |
| OffEval | 13.2k | Offensive(33.23%), Not-offensive(66.77%) | |
| Davids | 24.8k | Hate(5.77%), Offensive(77.43%), Neither(16.80%) | |
| Founta | 99k | Abusive(27.15%), Hateful(4.97%), Normal(53.85%), Spam(4.97%) | |
| FNUC | Fox News Discussion Threads | 1.5k | Hateful(28.50%), Non-hateful(71.50%) |
| StormW | Stormfront(forum) | 10.7k | Hate(10.93%), NoHate(89.07%) |
MACAS ALD模型
Multi-Aspect Cross Attention Super Joint 模型
ALD模型
TIS
TF-IDF + SVM,使用三个特征:
- TF-IDF的权重
- 代词和辱骂词汇的TF-IDF权重
- 与相邻帖子的相似性
RBF核来处理非线性超平面分离问题。min_df设置为2
OTH
One-Two Steps Hybrid CNN
使用Chars2vec作为character嵌入，Glove作为word嵌入
MFR
Multi-Features with RNN
TWL
Two-step Word-level LSTM
LTC
Latent Topic Clustering with Bi-GRU
CBT
Character-based Transformer
引用
Classification Methods for Hate Speech Diffusion: Detecting the Spread of HateSpeech on Twitter
基于网络图中的扩散来检测Twitter中hate speech的传播