Abstract The European Food Safety Authority (EFSA) develops methodologies and tools for the detection of emerging risks in food and feed. This includes the identification of drivers of emerging risks, such as food frauds, which requires innovative approaches. In this study, an unsupervised machine learning technique called the Latent Dirichlet Allocation (LDA) topic model, was applied on a media corpus in the view of detecting rapidly specific food fraud incidents in the media, i.e. on the Europe Media Monitor Medical Information System (EMM/MEDISYS). LDA topic model can explore large collection of documents discovering the themes associated with the corpus and organize and summarize text documents identifying topics comprised in them, where a topic is defined as a pattern of words with their probability to belong to it. As a specific food fraud incident, beeswax adulteration was taken as an example. Beeswax can be adulterated for financial gain, and, although it is a product from apiculture, it might enter the food chain when it is introduced as honeycomb in honey pots. With the beeswax example, a total of 2276 news articles were retrieved on EMM/MEDISYS and classified into 10 topics showing different levels of relatedness to beeswax adulteration. A manual screening of all articles allowed to validate the classification made by the topic model. The topics that were found the most relevant contained indeed articles on beeswax adulteration incidents reported from official sources. In addition, those topics contained signals of potential emerging risks in the cosmetic and food wrapping sectors. The remaining topics highlighted the emergence of new beeswax market opportunities which supported the identified signals. It is concluded that the LDA topic model can be used to process rapidly information in the media, support the definition of more specific food fraud filters on EMM/MEDISYS and be of direct use for all stakeholders involved in the monitoring, assessment and management of food frauds.

中文翻译：

---

一种识别和跟踪媒体中蜂蜡掺假新风险的主题模型方法

摘要 欧洲食品安全局 (EFSA) 开发了用于检测食品和饲料中新出现的风险的方法和工具。这包括识别食品欺诈等新兴风险的驱动因素，这需要创新的方法。在这项研究中，一种称为潜在狄利克雷分配（LDA）主题模型的无监督机器学习技术被应用于媒体语料库，以快速检测媒体中的特定食品欺诈事件，即欧洲媒体监测医疗信息系统（ EMM/MEDISYS）。LDA 主题模型可以探索大量文档，发现与语料库相关的主题，并组织和总结识别其中包含的主题的文本文档，其中一个主题被定义为一个单词的模式，它们的概率属于它。作为具体的食品造假事件，以蜂蜡掺假为例。蜂蜡可以为了经济利益而掺假，虽然它是养蜂业的产品，但当它作为蜜罐中的蜂窝引入时，它可能会进入食物链。以蜂蜡为例，在 EMM/MEDISYS 上共检索到 2276 篇新闻文章，并将其分为 10 个主题，显示出与蜂蜡掺假的不同程度的相关性。手动筛选所有文章以验证主题模型所做的分类。发现最相关的主题确实包含有关官方来源报告的蜂蜡掺假事件的文章。此外，这些主题包含化妆品和食品包装行业潜在新风险的信号。其余主题强调了新蜂蜡市场机会的出现，这些机会支持已确定的信号。得出的结论是，LDA 主题模型可用于快速处理媒体中的信息，支持在 EMM/MEDISYS 上定义更具体的食品欺诈过滤器，并可直接供参与食品监测、评估和管理的所有利益相关者使用。欺诈。