Using Open-Weight BART-MNLI for Zero Shot Classification of Food Recall Documents 

Status: Completed |  Funding: Howard University 

Food safety is a critical global concern, with foodborne illnesses impacting millions annually and posing significant health and economic burdens. Rapid and accurate detection of food hazards is essential, yet the vast volume of food-incident reports makes manual monitoring impractical. Advancements in AI and Natural Language Processing (NLP) offer promising solutions, enabling automated classification of food hazards from recall documents, social media reports, and official notices. However, existing models often struggle with imbalanced data and lack explainability, which is essential for regulatory compliance and trust.This study investigates the effectiveness of zero-shot classification for food hazard detection using the BART-large-MNLI model. Specifically, we explore whether this approach can categorize hazards and products without task-specific fine-tuning and how well it performs on unbalanced textual data. Our method predicts hazard and product categories, along with their exact vectors, from short food recall texts. The key contribution of this work is evaluating the effectiveness of zero-shot classification using a well-studied pre-trained open-weight model, providing a scalable and explainable approach to automated food hazard detection. 

Contributors

Progress Timeline

• 2024