Quantum-Enhanced Bigram Language Identifier

A Hybrid Quantum-Classical Approach for Multilingual Text Classification in the AQC Hackathon

Project Overview

This repository presents a sophisticated hybrid quantum-classical framework for language identification, specifically tailored for English (eng), French (fra), and Twi (twi). The system leverages classical bigram-based feature extraction with quantum-optimized weighting to achieve high accuracy and efficiency.

The core innovation lies in using a variational quantum circuit to learn optimal feature weights, which are then applied in a classical inference pipeline. This design addresses the hackathon challenge by employing quantum computing to enhance accuracy through non-linear feature combinations, while ensuring scalability for production use.

Key contributions:

• Feature Engineering: Combined word and character bigram scoring with additive smoothing for robust handling of unseen patterns.
• Quantum Optimization: A parameterized quantum circuit optimizes feature weights via cross-entropy minimization.
• Hybrid Inference: Quantum-trained weights enable fast, classical prediction on large corpora.
• Evaluation Methodology: Dual testing on held-out dataset and 100,000 synthetically generated sentences for realistic performance assessment.

This work demonstrates the potential of quantum computing in natural language processing tasks, particularly for low-resource languages like Twi.

Features

• Multi-Level Bigram Modeling: Word and character-level features for comprehensive linguistic analysis.
• Quantum Weight Learning: 4-qubit variational circuit with rotational and entangling gates to derive optimal feature weights.
• Efficient Prediction: Batch processing supports 100k+ sentences with low latency.
• Robust Smoothing: Additive smoothing (α=0.1) to handle sparse or unseen bigrams.
• Realistic Evaluation: Synthetic sentence generation using Faker for English/French and custom word bank for Twi.
• Modular Design: Clean classes for feature extraction, quantum optimization, and hybrid prediction.

Installation

Prerequisites

• Python 3.8+
• Google Colab (recommended for quantum simulation)

Steps

1. Clone the repository:

   git clone https://github.com/freddyfavour/quantum-language-identifier.git
   cd quantum-language-identifier
   

2. Install dependencies:

   pip install qiskit qiskit-aer scipy pandas numpy faker
   

   Or create a requirements.txt file

   qiskit==1.0.2
   qiskit-aer==0.13.2
   scipy==1.11.4
   pandas==2.1.4
   numpy==1.26.2
   faker==22.0.0
   scikit-learn==1.3.2
   

3. Place bigrams.csv in your working directory or Google Drive.

Usage

In Google Colab

1. Mount Drive:

   from google.colab import drive
   drive.mount('/content/drive')

2. Load and run the notebook cells in sequence.

Prediction Example

extractor = FeatureExtractor(alpha=0.1)
extractor.build_models(df)  # df is the loaded bigrams.csv

# Train quantum optimizer (on your training data)
quantum = QuantumWeightOptimizer(num_qubits=4, layers=4)
quantum.train(X_train, y_train)  # X_train, y_train from prepared data

model = HybridLanguageIdentifier(extractor, quantum.weights)

# Single prediction
print(model.predict("This is an English sentence."))

# Batch prediction
batch = ["Bonjour le monde", "Akwaaba to the world", "Hello world"] * 1000
preds = model.predict_batch(batch)

Training

The system trains the quantum circuit on a sample of the dataset (or generated sentences) to learn feature weights.

Methodology

1. Feature Extraction

• Word bigrams capture sequential patterns.
• Character bigrams handle orthographic nuances (e.g., Twi accents).
• Scores computed as averaged log-probabilities with smoothing.

2. Quantum Weight Optimization

• Variational circuit: RX, RY, RZ rotations + CX entanglements.
• Measurement probabilities mapped to 6 feature weights.
• Optimized using COBYLA on cross-entropy loss.

3. Hybrid Prediction

• Weighted sum of features using quantum-derived weights.
• Aggregation: Word + character scores per language → argmax.

4. Evaluation

• Dataset Test: Held-out bigrams from CSV.
• Generated Test: 100k synthetic full sentences using Faker (English/French) and Twi word bank.
• Metrics: Accuracy, classification report, confusion matrix.

Results

• CSV Test Accuracy: ~0.95–0.99 (reference metric).
• Generated Sentences Accuracy: ~0.85–0.92 (realistic generalization).
• Training Time: ~3 minutes for 1M samples.
• Inference Speed: Thousands of sentences per second.

Detailed reports and confusion matrices are printed during execution.

Limitations and Future Work

• Twi sentence generation relies on limited word bank — could integrate more diverse corpora.
• Quantum simulation on classical hardware; potential for real QPU deployment.
• Explore deeper circuits or alternative encodings for further accuracy gains.

Team

• Salam Musa
• Favour Alfred
• Kashi

License

MIT License. See LICENSE for details.

---

Submitted to the AQC Hackathon
Date: 5th January, 2025