Skip to main content
The BLEU (bilingual evaluation understudy) metric is one of the Deterministic Metric Galtea exposes to objectively compare a model’s output against ground‑truth references. It is most appropriate when you expect the model to reproduce wording very close to the reference (e.g., machine translation or tightly constrained generation).

​
Evaluation Parameters

To compute the bleu metric, the following parameters are required:
  • actual_output: The model’s generated text.
  • expected_output: The reference (or gold) text to compare against.

​
How Is It Calculated?

Conceptually, BLEU is computed as:
  1. N‑gram Precision
    Compute modified precision for n‑grams of sizes 1, 2, 3, and 4: how many of the candidate’s n‑grams also appear in the reference(s), clipped by the maximum count in the references.
  2. Geometric Mean of Precisions
    Combine the n‑gram precisions with a geometric mean to balance contributions from different n‑gram sizes.
  3. Brevity Penalty (BP)
    Penalize overly short candidates:     BP={1if c>re(1−rc)if c≤rBP = \begin{cases} 1 & \text{if } c > r \\ e^{(1 - \frac{r}{c})} & \text{if } c \le r \end{cases} where cc is candidate length and rr is reference length.
  4. Final Score BLEU=BP⋅exp⁡(∑n=14wnlog⁡pn)BLEU = BP \cdot \exp\left(\sum_{n=1}^{4} w_n \log p_n \right) where pnp_n are the modified n‑gram precisions and wnw_n are their weights (uniform, wn=14w_n = \frac{1}{4} ).
BLEU returns a value between 0 and 1 in this implementation.
  • ≥ 0.6 – Very strong overlap / near-reference quality.
  • 0.3 – 0.6 – Moderate overlap; acceptable for some machine translation tasks.
  • < 0.3 – Low overlap; likely poor or very different phrasing.

​
Suggested Test Case Types

Use BLEU when you have deterministic, reference-style outputs, such as:
  • Machine Translation (primary use case).
  • Data-to-Text Generation where the wording is tightly constrained.
  • Template-like Summaries or Headline Generation where near-exact phrasing matters.
  • Paraphrase Detection (strict) when you want to reward lexical overlap (with caution).