Skip to main content

Back-Translation Augmentation

The idea: Generate synthetic parallel data by translating existing target-language text back into the source language, then use these synthetic pairs to train or prompt a forward model. This expands your parallel corpus cheaply — but with caveats about quality.

:::info This is a cookbook, not a finished implementation This guide sketches the strategy and its critical pitfalls. Back-translation is powerful but can amplify errors if not done carefully. :::

When to Use This​

  • You have monolingual target-language text but limited parallel data
  • You want to expand a training corpus for fine-tuning without manual translation
  • You need more few-shot examples but can't get human translations fast enough
  • You're willing to quality-filter the synthetic data aggressively

How It Works​

[Target-language text]          "awâsisak mêtawêwak"
        │
        ▼
[Back-translate to source]      "The children are playing"  (via LLM or MT API)
        │
        ▼
[Create synthetic pair]         ("The children are playing", "awâsisak mêtawêwak")
        │
        ▼
[Quality filter]                Keep only high-confidence pairs
        │
        ▼
[Use for training/prompting]    Expand your parallel corpus
  1. Collect monolingual text — target-language books, articles, transcripts, social media
  2. Back-translate — use an LLM or MT API to translate each sentence to the source language
  3. Quality filter — round-trip (translate back again) and compare; keep pairs where round-trip ≈ original
  4. Use the synthetic corpus — for fine-tuning, few-shot examples, or coaching data

Quality Filtering: The Round-Trip Test​

# Pseudo-code for round-trip quality filtering
for target_text in monolingual_corpus:
    # Back-translate: target → source
    synthetic_source = translate(target_text, "crk", "en")
    
    # Forward-translate: source → target
    round_trip = translate(synthetic_source, "en", "crk")
    
    # Compare round-trip to original
    chrf_score = compute_chrf(target_text, round_trip)
    
    if chrf_score > 0.70:  # High similarity = high-quality pair
        parallel_corpus.append((synthetic_source, target_text))

Critical Pitfall: Error Amplification​

:::warning Back-translation amplifies existing model biases If your back-translation model consistently makes the same errors, your synthetic corpus will encode those errors as "correct." This creates a feedback loop: train on bad data → produce worse translations → generate worse synthetic data. Always quality-filter aggressively and mix synthetic data with verified human translations. :::

Where to Find Monolingual Text​

  • Community newsletters, newspapers, and publications
  • Government documents in the target language (e.g., Nunavut Hansard for Inuktitut)
  • Educational materials and textbooks
  • Religious texts (widely available for many languages)
  • Social media (with appropriate permissions and quality filtering)
  • Transcribed audio/video from language programs

Pros and Cons​

✅ Expands training data cheaply❌ Amplifies model errors if not filtered
✅ Uses abundant monolingual text❌ Quality ceiling limited by back-translation model
✅ Easy to generate at scale❌ Round-trip filtering is compute-intensive
✅ Complements other approaches❌ Synthetic data is never as good as human translation

Combines Well With​

See Also​