translators

README: English to Croatian and Croatian to English Translator

Overview

This project involves building a translator capable of converting text:

1. From English to Croatian
2. From Croatian to English

Two models are implemented for this task:

• LSTM-Based Translator
• Seq-to-Seq Translator (without attention)

The project includes dataset preparation, model training, evaluation using BLEU and CHRF scores, and a comparison of the models' performance.

---

How the Translator Works

1. Translation Models

LSTM-Based Translator

• The LSTM (Long Short-Term Memory) model uses recurrent layers to process sequences of words.
• The input text is tokenized and converted into numerical representations (word embeddings).
• The LSTM processes the sequence word by word, generating an output sequence in the target language.
• The final output sequence is reconstructed into text.

Seq-to-Seq Translator

• The Seq-to-Seq (Sequence-to-Sequence) model uses two main components:
  • Encoder: Encodes the input sequence into a fixed-length vector representation.
  • Decoder: Decodes the vector representation into the target language sequence.
• Unlike the LSTM-based translator, the Seq-to-Seq model does not use attention mechanisms.
• This simpler structure may lead to lower accuracy on long sentences but works effectively for shorter sequences.

2. Dataset Preparation

• The dataset is sourced from the HuggingFace dataset repository.
• The data is prepared in a parallel format to align English and Croatian sentences:

  <s>
  <en> This is a sample sentence in English. </en>
  <cr> Ovo je primjer rečenice na hrvatskom jeziku. </cr>
  </s>
  

• Both English-to-Croatian and Croatian-to-English pairs are created for training.

3. Training Process

• Input Data: Sentences are tokenized into word or subword tokens.
• Model Training:
  • For each pair of input and target sequences, the model learns to predict the target tokens from the input tokens.
  • Loss is calculated based on the difference between the predicted sequence and the target sequence.
• Validation: The model's performance is evaluated on a separate validation dataset to monitor overfitting or underfitting.

---

Evaluation Metrics

1. BLEU Scores:
   • Measure n-gram overlap between the predicted and reference translations.
   • Higher scores indicate better translation accuracy.
2. CHRF Scores:
   • Evaluate translation quality at the character level, useful for morphologically rich languages like Croatian.
3. Training Curves:
   • Training and validation loss are plotted to assess the model's learning progress.

---

File Structure

English-Croatian-Translator/
├── data/
│   ├── Cleaned_Croatian.json   #Dataset in parallel format
├── notebooks/
│   ├── lstm_translator.ipynb # LSTM-based translator implementation
│   ├── seq2seq_translator.ipynb # Seq-to-Seq translator implementation
├── results/
│   ├── bleu_scores.csv       # BLEU scores for both models
│   ├── chrf_scores.csv       # CHRF scores for both models
│   ├── training_curves/
│   │   ├── lstm_training_curve.png # LSTM training/validation loss curve
│   │   ├── seq2seq_training_curve.png # Seq-to-Seq training/validation loss curve
│   ├── bleu_comparison_plot.png  # BLEU score comparison plot
│   ├── chrf_comparison_plot.png  # CHRF score comparison plot
├── README.md                 # Project README file

---

How to Run

1. Dataset Preparation:
   • Ensure the dataset is formatted in the parallel structure mentioned above.
   • Upload the dataset files (`cleaned_croatian.json') to the project folder or Colab.
2. Training:
   • Open the provided Colab notebooks:
     • lstm_translator.ipynb
     • seq2seq_translator.ipynb
   • Run all cells to train the models.
3. Evaluation:
   • Calculate BLEU and CHRF scores using the test dataset.
   • Plots for training curves, BLEU, and CHRF comparisons will be automatically generated and saved.
4. Results:
   • Review the generated score files (bleu_scores.csv and chrf_scores.csv) and plots in the results/ folder.

---

References

• HuggingFace Dataset Repository
• BLEU Score Metrics
• CHRF Scores Guide

This project enables translation between English and Croatian using two types of models, showcasing the strengths and weaknesses of LSTM-based and Seq-to-Seq architectures.