• To address the lack of a standard harmful euphemism dataset, we selected Bilibili as the target platform because harmful euphemisms spread quickly and widely on the site.

• To construct our corpus, we developed an efficient web scraper using the Scrapy framework to collect video comments.

• To collect different harmful euphemisms on social media platforms, the dataset needed to cover a wide range of topics. We manually selected videos from various topics as seed videos and used the video IDs to collect corresponding comments through the web scraper. ​- To ensure the diversity and balance of the comments, the seed videos were selected from different channels on Bilibili, and the seed keywords were manually gathered based on Bilibili's search engine.

• The popularity of the videos was assessed using metrics such as views, likes, donations, comments, and shares.

• The initial seed keywords covered six topics: "Race"(“种族”), "Sexism"(“色情”), "General"(“日常”), "Gender"(“性别”), "LGBT"(“性少数群体”), and "Region"(“地区”).

• The field details of the TE-dataset:

• Some samples of the TE-dataset:

• The Summary of the TE-dataset：

• dataset download

Method 1：download files directly

• access the Link（https://github.com/yiyepianzhounc/TED-SCL）
• click "Raw" and download （The type of files is ".csv"，and the encoding type is "utf-8"）

Method 2：download by "Git" command

- git clone https://github.com/yiyepianzhounc/TED-SCL.git

    import pandas as pd
    df = pd.read_csv("path/to/dataset.csv")

• train.csv: for training
• eval.csv: for evaluating
• test.csv: for testing
• eu-pairs: Annotated dictionary of toxic euphemism terms
• Please read the train.csv, eval.csv, and test.csv firstly, then load the Annotated dictionary of toxic euphemism terms (eu-pairs.csv) and map toxic euphemism terms to their corresponding toxic meaning wordstarget.

article：A Toxic Euphemism Detection Framework for Online Social Network Based on Semantic Contrastive Learning and Dual Channel Knowledge Augmentation

• Corresponding author Professor Haizhou Wang： E-mail：[email protected]

• First Author Gang Zhou E-mail：[email protected]

To prevent privacy leakage, the following processing steps have been applied to this dataset before release:

• Special entity have been deleted or anonymized;
• Personally Identifiable Information (PII) that could be used to identify individuals (including IDs or nicknames on social media platforms) has been masked;
• The release of this dataset is for research purposes only and should not be used for any inappropriate applications;
• Please sign the "Authorization for Use of the Dataset."

• 为了解决缺乏标准有害委婉语数据集的问题，我们选择了哔哩哔哩作为目标平台，因为该网站上有害委婉语的传播迅速且广泛。​

• 为了构建我们的语料库，我们开发了一个高效的网页爬虫，使用Scrapy框架收集视频评论。​

• 为了研究社交媒体平台上不同的有害委婉语，数据集需要涵盖广泛的话题。我们手动选择了多种话题的视频作为种子视频，并通过每个视频的ID使用网络爬虫获取相应的评论。 ​- 为了确保评论的多样性和平衡性，种子视频从哔哩哔哩的不同频道中选择，基于哔哩哔哩搜索引擎的种子关键词。​

• 视频的受欢迎程度通过观看次数、点赞数、打赏数、评论数和分享数等指标进行评估。​

• 初始的种子关键词包括六个话题： “种族” (Race), “色情” (Sexism), “日常” (General), “性别” (Gender), “性少数群体” (LGBT), and “地区” (Region)。

• 数据集字段的基本情况说明：topic,content,toxic,length,euphemism

• 数据集的示例数据：

• 数据集的详细信息统计：

TE-Dataset中有害委婉语评论的平均长度和最小长度分别为33.7和3.33，​TE-Dataset中PTET的总数为424。

• 数据集下载

方法 1：直接下载单个文件

• 访问仓库主页（https://github.com/yiyepianzhounc/TED-SCL）
• 找到数据集文件点击下载 （数据集以.csv文件形式存在，编码方式为utf-8）
• 找点击文件名称进入文件详情页，点击右上角的 Download（或 Raw）按钮直接下载

方法 2：下载整个仓库（推荐）- 需要安装 Git。

• git clone https://github.com/yiyepianzhounc/TED-SCL.git
• 数据集导入 导入方式取决于数据格式和目标工具（如 Python、Excel 等）。以下是常见格式的示例：

    import pandas as pd
    df = pd.read_csv("path/to/dataset.csv")

• 数据集使用 train.csv用于模型训练，eval.csv用于模型评估，test.csv用于模型测试 注意：请读取完训练集、测试集或验证集后，再读取eu-pairs中有害委婉语与有害含义目标词的映射字典，然后拼接到原来读取到的dataframel，再进行进一步的数据操作。

article：A Toxic Euphemism Detection Framework for Online Social Network Based on Semantic Contrastive Learning and Dual Channel Knowledge Augmentation

• Corresponding author Professor Haizhou Wang： E-mail：[email protected]

• First Author Gang Zhou E-mail：[email protected]

• 对特殊词汇进行了删除或匿名化；
• 屏蔽了能够定位识别个人身份的PII信息（包括：社交网络平台中的ID或昵称信息）

• 签订《Authorization for Use of the Dataset》，严格确保不用于商业用途或其他风险。