A Fine-Tuned Bidirectional Encoder Representations From Transformers Model for Food Named-Entity Recognition:...

Area of Science:

• Food science
• Biomedical informatics
• Natural language processing

Background:

• Food science research faces challenges in understanding food interactions with health entities.
• Limited resources exist for machine learning in food science, hindering information extraction.
• An annotated corpus for food entities was published in 2019, aiding research.

Purpose of the Study:

• To investigate the fine-tuning of Bidirectional Encoder Representations from Transformers (BERT) for food information extraction.
• To develop and evaluate corpus-based food named-entity recognition (NER) methods.

Main Methods:

• Introduced FoodNER, a collection of 15 models fine-tuned from 3 pretrained BERT models.
• Utilized 5 groups of semantic resources: food vs. non-food, Hansard, FoodOn, and SNOMED CT tags.
• Evaluated models on tasks including food vs. non-food entity recognition and food group classification.

Main Results:

• Achieved state-of-the-art performance in distinguishing food vs. non-food entities, with macro F1 scores from 93.30% to 94.31%.
• Demonstrated strong results in predicting semantic tags, with macro F1 scores ranging from 73.39% to 78.96%.

Conclusions:

• FoodNER models offer robust capabilities for extracting and annotating food entities.
• The system supports 5 distinct annotation tasks, including food vs. non-food and hierarchical food group classification.
• This work advances the field of food information extraction through advanced NLP techniques.