美容仪器类快手直播带货违法监测预警数据

对采集的快手平台带货品类为美容仪类的达人直播视频内容进行转译分析，对达人口播语言内容进行处理、分析，根据达人在直播过程中对预先设置的违规敏感词（比如：全方位、超强、立刻见效、永久去除、院线级别、安全、无痛、不留疤痕、没有副作用等）违反的次数和频率，依据触发条件规则提出警告或处理。为服务辖区市场监督局管理区域内规范企业快手直播行为，提供数据支持。将采集完成的直播视频进行进行预处理，第一步：基于原始视频文件，以最大10分钟单位对原始视频进行切片。第二步：对于已完成的切片视频，进行视频内容转语音操作。第三步：对于已完成视频转语音操作的切片，进行语音转文本操作。第四步：使用OCR技术对原始视频中抓取的图片进行文字提取操作。第五步：将所得到的文字内容与违法预警关键词库进行匹配。最终运用多标准决策分析模型，对主播在直播过程中出现的违规语句进行分析计算，得出违法预警值和是否预警判断。违法预警值 ≤1 时，不触发预警提示，违法预警值 ＞1 时触发违法预警提示。

This dataset conducts transcription and analysis on livestream video content of Kuaishou platform influencers who promote beauty instrument products. We process and analyze the spoken language content of the influencers, and issue warnings or take targeted treatments in accordance with the triggering condition rules, based on the frequency and count of violations against pre-defined sensitive violation keywords (e.g., "comprehensive", "ultra-powerful", "immediate effect", "permanent removal", "professional salon-grade", "safe", "painless", "no scar left", "no side effects") during the livestream. This work aims to provide data support for market supervision bureaus in their respective jurisdictions to standardize the livestream behaviors of enterprises on Kuaishou.

The collected livestream videos undergo the following preprocessing steps:
1. Slice the original video files into segments with a maximum duration of 10 minutes per segment;
2. Extract audio from each sliced video and convert it into standalone audio files;
3. Perform speech-to-text transcription on the extracted audio files to generate text content;
4. Extract text from images captured from the original videos using Optical Character Recognition (OCR) technology;
5. Match the collected text content against the pre-built illegal warning keyword database.

Finally, a multi-criteria decision analysis (MCDA) model is applied to analyze and calculate the violation statements made by the hosts during the livestream, to derive the illegal warning score and make a warning judgment. Specifically, if the illegal warning score is ≤1, no warning alert will be triggered; if the illegal warning score > 1, an illegal warning alert will be triggered.