🍲 FastText Word Embeddings for Indian Food Recipes

This project demonstrates the use of FastText for creating and exploring word embeddings, with a special focus on Indian food recipes. We'll use pre-trained models for English and Hindi, and then train a custom model on Indian food recipe data.

📋 Table of Contents

1. Installation
2. Downloading Pre-trained Models
3. Using Pre-trained FastText Models
   • English Model
   • Hindi Model
4. Custom Training on Indian Food Recipes
   • Data Preprocessing
   • Training the Model
   • Exploring the Custom Model
5. Conclusion
6. Further Resources

🛠 Installation

First, install the required libraries:

pip install fasttext pandas

📥 Downloading Pre-trained Models

FastText provides pre-trained word vectors for 157 languages. To download the models used in this project:

1. Visit the FastText website
2. Scroll down to the "Pre-trained word vectors" section
3. Download the following files:
   • For English: cc.en.300.bin.gz
   • For Hindi: cc.hi.300.bin.gz

After downloading, extract the .bin files from the .gz archives.

gunzip cc.en.300.bin.gz
gunzip cc.hi.300.bin.gz

Move the extracted .bin files to your project directory or a designated models folder.

🚀 Using Pre-trained FastText Models

English Model

Load the pre-trained English model and explore word embeddings:

import fasttext

# Load the pre-trained English model
model_en = fasttext.load_model('path/to/cc.en.300.bin')

# Get nearest neighbors for 'good'
print(model_en.get_nearest_neighbors('good'))

# Check the shape of the word vector
print(model_en.get_word_vector("good").shape)

# Get analogies
print(model_en.get_analogies("berlin", "germany", "france"))

Output:

[(0.7517593502998352, 'bad'),
 (0.7426098585128784, 'great'),
 (0.7299689054489136, 'decent'),
 ...]

(300,)

[(0.7303731441497803, 'paris'),
 (0.6408537030220032, 'france.'),
 (0.6393311023712158, 'avignon'),
 ...]

Hindi Model

Load the pre-trained Hindi model and explore word embeddings:

# Load the pre-trained Hindi model
model_hi = fasttext.load_model('path/to/cc.hi.300.bin')

# Get nearest neighbors for "अच्छा" (good)
print(model_hi.get_nearest_neighbors("अच्छा"))

Output:

[(0.6697985529899597, 'बुरा'),
 (0.6132625341415405, 'अच्छे'),
 (0.608695387840271, 'अच्चा'),
 ...]

🥘 Custom Training on Indian Food Recipes

Data Preprocessing

1. Load the dataset:

import pandas as pd
import re

df = pd.read_csv("Cleaned_Indian_Food_Dataset.csv")
print(df.shape)
print(df.head(3))

2. Define a preprocessing function:

def preprocess(text):
    text = re.sub(r'[^\w\s\']',' ', text)
    text = re.sub(r'[ \n]+', ' ', text)
    return text.strip().lower()

3. Export preprocessed data to a text file:

df.to_csv("food_recipes.txt", columns=["TranslatedInstructions"], header=None, index=False)

Training the Model

Train a custom FastText model on the preprocessed data:

import fasttext

model = fasttext.train_unsupervised("food_recipes.txt")

Exploring the Custom Model

Explore the custom-trained model:

# Get nearest neighbors for 'paneer'
print(model.get_nearest_neighbors("paneer"))

# Get nearest neighbors for 'halwa'
print(model.get_nearest_neighbors("halwa"))

Output:

[(0.6676578521728516, 'tikka'),
 (0.6331593990325928, 'bhurji'),
 (0.6316412687301636, 'tikkas'),
 ...]

[(0.7327786087989807, 'khoya'),
 (0.7155830264091492, 'sheera'),
 (0.6999987363815308, 'rabri'),
 ...]

🎉 Conclusion

This project demonstrates the versatility of FastText for working with word embeddings. We've shown how to use pre-trained models for English and Hindi, as well as how to train a custom model on domain-specific data. These techniques can be applied to various natural language processing tasks, particularly those involving specialized vocabularies or multilingual contexts.

The custom-trained model now provides word embeddings specifically tailored to Indian food recipes, allowing for more accurate and relevant word associations within this domain. This can be particularly useful for applications such as recipe recommendation systems, ingredient substitution suggestions, or culinary trend analysis.

📚 Further Resources

• FastText Official Website
• FastText GitHub Repository
• Tutorial on Word Embeddings
• Research Paper: Enriching Word Vectors with Subword Information

Happy embedding! 🚀🍽️