Start with an interpretable baseline
Established a TF-IDF and logistic-regression baseline before comparing stronger statistical and neural approaches.
Machine learning / NLPPublic source & report
Food Hazard Detection
Finding an effective text-classification pipeline for imbalanced food-recall data.
The context
SemEval-2025 Task 9 requires predicting a hazard category and a product category from recall reports. The repository reports a final Subtask 1 score of 0.78003.
View repository ↗Machine learning / NLP↗
Reported final Subtask 1 score0.78003
10Hazard categories
22Product categories
SVCFinal classifier
How it connects
The workflow.
- 01Recall text
- 02Word / char TF-IDF
- 03LinearSVC
- 04Hazard + product labels
The final pipeline combines text features, title boosting and country information for classification.
Engineering details
Inside the work.
Features suited to the domain
Combined word and character n-grams with title boosting. Country one-hot features were added to the product-category pipeline.
Compare approaches empirically
The tested LinearSVC pipelines outperformed the tested DistilBERT approach in this project. The reported scores progressed from 0.68695 for the baseline to 0.78003 for the final submission.
Document the evaluation setting
The final approach uses transductive TF-IDF: its vocabulary is fitted using training, validation and unlabeled test text. The reported competition score is specific to that evaluation setup.
What I took from it.
- Model complexity does not automatically translate into better results.
- Character features can capture useful variation in domain-specific vocabulary.
- Feature engineering and an explicit evaluation setup are central to interpreting a result.
See for yourself
Source & documentation.
Next projectGreek++ Compiler↗