The Food Hazard Detection Challenge

A collaborative project for SemEval 2025 Task 9: The Food Hazard Detection Challenge, leveraging advanced natural language processing (NLP) models. This repository contains implementations, experiments, and reports focused on detecting food hazards based on textual data.

Team Members

• Matheus Campos
• Daniel Menezes
• Matheus Laureano

Objective

This project aims to develop an efficient solution for detecting food hazards from text, using Transformers and Natural Language Processing (NLP) techniques.
The core approach includes:

• Data analysis and preprocessing: Cleaning raw text and engineering additional features.
• Model training and evaluation: Developing a DistilBERT-based model with integrated temporal features.

---

Repository Structure

├── data/                 # Raw data, preprocessed datasets, and preparation scripts
├── notebooks/            # Jupyter notebooks for experimentation and analysis
├── report/               # Collaborative LaTeX report (Overleaf)
└── README.md             # Project documentation (this file)

---

Implemented Architecture

The implemented architecture is based on DistilBERT, enhanced with extensions that integrate additional (temporal) features into the classification head. Key components include:

1. Base Model: DistilBERT

• The DistilBERT pre-trained model generates robust embeddings from descriptive text.

2. Custom Classification Head

The classification head is modified to integrate textual and additional features:

• Feature Engineering: Temporal features (e.g., incident date) are normalized and incorporated.
• Representation Fusion:
  • Text embeddings are concatenated with the additional features.
  • A custom cross-attention mechanism allows the model to dynamically adjust weights between textual and temporal signals.
  • The combined representation is passed through a linear layer to generate logits.
• Loss Function: Uses CrossEntropyLoss for multi-class classification.

3. Optimization and Regularization

• Optimizer: AdamW for improved weight regularization.
• Scheduler: Linear decay with warm-up to dynamically adjust the learning rate.
• Dropout: Applied after feature fusion to reduce overfitting.

Model Data Flow

1. Input:
   • Descriptive text
   • Additional features (normalized and, if applicable, cyclically encoded).
2. Processing:
   • Text is transformed into embeddings using DistilBERT.
   • The embeddings are concatenated with the additional features.
3. Output:
   • Classification logits for each food hazard category.

---

Results

Key Findings:

• Accuracy: 0.77
• Temporal feature integration: Improved model accuracy by [X]% over the baseline.
• Visualizations: See the results/ folder for detailed charts and metrics.

Prediction Examples:

• Input: "Recall Notification: FSIS-033-94 - Sausage contaminated"
  • Output: Risk Category: Biological
• Input: "Plastic fragments in chicken breast - High risk"
  • Output: Risk Category: Foreign Body

---

Contributions

Each team member contributed to different aspects of the project:

• Matheus Campos: Data engineering, model architecture development.
• Daniel Menezes: Validation module, result visualization, and report structuring.
• Matheus Laureano: Data preprocessing, notebook experimentation, and training pipeline.

Detailed contributions can be found in the report.