sgarbi/

bert-fda-nutrition-ner

Training Data Description

Data Source

• Labeling Source: U.S. Food and Drug Administration (FDA), FoodData Central. FDA FoodData Central. The dataset includes detailed nutritional data, such as ingredient lists, nutritional values, serving sizes, and other essential label information.
• Yelp Restaurant Reviews: Utilized the Yelp Review Full dataset from Hugging Face, augmented with Mistral 7B for general tagging, to enrich the model's understanding of restaurant-related nutritional mentions.
• Amazon Food Reviews: Similar to the Yelp dataset, this model also incorporates the Amazon Food Reviews dataset from Hugging Face, augmented with Mistral 7B, enhancing its capability to recognize and classify a wide range of nutritional information from diverse food product reviews correlated with FDA data.

Preprocessing and Augmentation Steps

• Extraction: Key textual data, encompassing nutritional facts and ingredient lists, were extracted from the FDA dataset.
• Normalization: All text underwent normalization for consistency, including converting to lowercase and removing redundant formatting.
• Entity Tagging: Significant nutritional elements were manually tagged, creating a labeled dataset for training. This includes macronutrients, vitamins, minerals, and various specific dietary components.
• Tokenization and Formatting: The data was tokenized and formatted to meet the BERT model's input requirements.
• Introducing Noise: To enhance the model's ability to handle real-world, imperfect data, deliberate noise was introduced into the training set. This included:
  • Sentence Swaps: Random swapping of sentences within the text to promote the model's understanding of varied sentence structures.
  • Introducing Misspellings: Deliberately inserting common spelling errors to train the model to recognize and correctly process misspelled words frequently encountered in real-world scenarios such as inaccurate document scans.

Considerations

• The model was trained only on publicly available data from food product labels. No private or sensitive data was used.
• Labeling tasks were performed by Mistral 7B-Instruct served by mistral.ai (https://docs.mistral.ai/). It is probable that models experienced hallucinations during labeling data, which could result in an imprecise taxonomy classification.
• The tool only extracts nutritional entities from text; it should not be used for nutrition or health recommendations. Qualified experts should provide any nutrition advice.
• The language and phrasing on certain types of food product labels may introduce biases to the model.
• This model was created for exploring the BERT architecture and NER tasks.

Label Map

Github