The METEOR (Metric for Evaluation of Translation with Explicit ORdering) metric is one of the Classic Metric Type Galtea exposes for evaluating machine translation, summarization, and paraphrasing tasks. It aims to correlate more closely with human judgments compared to BLEU.

​
Evaluation Parameters

To compute the meteor 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?

METEOR improves upon BLEU by considering semantic and morphological matches:
  1. Alignment
    Tokens are matched between candidate and reference using:
    • Exact matches
    • Stems (e.g., “run” vs. “running”)
    • Synonyms (e.g., “big” vs. “large”)
  2. Precision & Recall
    Both are calculated from the aligned tokens.
  3. Fragmentation Penalty
    A penalty is applied if matches are scattered (fragmented alignment).
  4. Final Score
    The score is computed as:     METEOR=Fmean×(1−Penalty)METEOR = F_{mean} \times (1 - Penalty) where F_mean is the harmonic mean of precision and recall.
METEOR returns a score between 0 and 1:
  • ≥ 0.6 – High-quality translation/summary with semantic fidelity.
  • 0.3 – 0.6 – Moderate quality; some paraphrasing or structural divergence.
  • < 0.3 – Low-quality or semantically incorrect output.

​
Suggested Test Case Types

Use METEOR when evaluating:
  • Machine Translation with varied phrasings.
  • Abstractive Summarization where synonyms are common.
  • Paraphrase Detection with semantic variation.