FoodMenu-NER-pt-br

FoodMenu-NER-pt-br is a Named Entity Recognition (NER) model specialized for Brazilian Portuguese (pt-br) texts, designed to extract ingredients and their quantities from food-related content such as menus and recipes.

Technical Specifications

Model Architecture

• Base Model: pt_core_news_lg (spaCy)
• Pipeline: Preserves all components except "ner", "trf_wordpiecer" and "trf_tok2vec"
• Fine-tuning: Transfer Learning focused on NER layer

Hyperparameters

• Number of Iterations: 60
• Batch Size: Dynamic (4 to 16, using compounding)
• Dropout Rate: 0.3

Entity Classes

• QUANTIDADE: Identification of measurements and quantities
• INGREDIENTE: Recognition of food ingredients

Project Structure

FoodMenu-NER-pt-br/
├── data/
│   └── dataset.json         
├── model/
│   └── food_item_ner_pt/   
├── src/
│   ├── utils.py             
│   ├── evaluate_model.py    
│   ├── inference.py         
│   └── train.py             
├── .gitignore
├── requirements.txt
└── README.md

Dataset

Statistics

• Total Examples: 243
• Data Split: 80% training, 20% validation
• Entity Distribution:
  • QUANTIDADE: 301 instances
  • INGREDIENTE: 1130 instances

Data Format

{
  "text": "500g de macarrão com molho",
  "entities": [
    {
      "start": 0,
      "end": 4,
      "label": "QUANTIDADE",
      "text": "500g"
    },
    {
      "start": 8,
      "end": 16,
      "label": "INGREDIENTE",
      "text": "macarrão"
    }
  ]
}

Training and Evaluation

To train the model, run:

python src/train.py

To evaluate the trained model:

python src/evaluate_model.py

Evaluation Metrics

Inference

import spacy

nlp = spacy.load("model/food_item_ner_pt")

text = "250g de filé de frango grelhado, temperado com limão, azeite e ervas finas. Servido com arroz basmati e legumes assados como abóbora e cenoura"
doc = nlp(text)

for ent in doc.ents:
    print(f"{ent.label_}: {ent.text}")

Command Line Usage

You can also run predictions using the command line:

python src/inference.py --text "250g de filé de frango grelhado, temperado com limão, azeite e ervas finas. Servido com arroz basmati e legumes assados como abóbora e cenoura"

Available Arguments:

Argument	Type	Required	Default	Description
--text	string	Yes	-	Input text for entity recognition
--model	string	No	model/food_item_ner_pt	Path to the trained model directory
--display	string	No	cli	Output format: cli for terminal or html for browser

Examples:

Basic usage with CLI output:

python src/inference.py --text "500g de açúcar e 2 ovos"

Using a custom model path:

python src/inference.py --text "300ml de leite" --model "path/to/custom/model"

HTML visualization in browser:

python src/inference.py --text "1kg de farinha de trigo" --display html

Complex recipe analysis:

python src/inference.py --text "Para o bolo: 3 ovos, 200g de açúcar, 250ml de leite e 300g de farinha. Para a cobertura: 100g de chocolate e 50ml de creme de leite."

Inference Output:

Known Limitations

• Model optimized only for culinary texts
• May struggle with regional culinary slang
• Performance not tested on very long texts

License

This project is released under the MIT License - refer to the LICENSE file for details.