基于大模型语义校正的直播电商语音实时合规监测方法

A Real-Time Compliance Monitoring Method for Live E-Commerce Speech Based on Large Language Model Semantic Correction

摘要: 随着直播电商行业的爆发式发展，直播内容的合规性监管成为行业治理的重点与难点。由于直播场景下存在背景噪声大、主播语速快、口音多样及专业术语(SKU)繁杂等特点，传统的自动语音识别(ASR)技术在转写准确率上存在瓶颈，导致违规内容监测的漏报率与误报率居高不下。本文提出了一种基于大语言模型(LLM)校正的直播电商语音实时合规监测方法。该方法首先利用流式语音识别技术获取原始文本，随后引入经过电商领域基于提示工程的上下文学习的大模型对转写结果进行语义纠错与上下文重构，最后结合关键词匹配与语义分析实现对夸大宣传、广告法违禁词等风险内容的实时判别。实验结果表明，该方法在电商直播数据集上的字错率(WER)相比传统方法降低了53.3%，违规内容检出率(Recall)提升了26.1%。该研究有效地解决了复杂声学环境下的内容风控难题，为电商平台的自动化治理提供了新的技术路径。

Abstract: With the explosive growth of the live e-commerce industry, the compliance supervision of live streaming content has become a key focus and challenge. Due to high background noise, fast speech rates, diverse accents, and complex professional terminology (SKU) in live streaming scenarios, traditional Automatic Speech Recognition (ASR) technologies face bottlenecks in transcription accuracy, leading to high false-negative and false-positive rates in violation monitoring. This paper proposes a real-time compliance monitoring method for live e-commerce voice based on Large Language Model (LLM) correction. This method first uses streaming speech recognition technology to obtain the original text, then introduces an LLM fine-tuned with e-commerce domain knowledge to perform semantic error correction and context reconstruction on the transcription results. Finally, it combines keyword matching and semantic analysis to achieve real-time identification of risk content such as exaggerated publicity and prohibited words under the Advertising Law. Experimental results show that the Word Error Rate (WER) of this method on e-commerce live streaming datasets is reduced by 53.3% compared to traditional methods, and the detection rate (Recall) of compliant content is improved by 26.1%. This research effectively solves the problem of content risk control in complex acoustic environments and provides a new technical path for the automated governance of e-commerce platforms.