BLEURT is a regression model built on top of BERT and trained in two stages. (1) Pre-training on millions of synthetic sentence pairs produced by randomly perturbing Wikipedia sentences (e.g. token masking, permutations, backtranslation); these pairs are not human-rated — labels are generated by a collection of existing signals and models from the literature (e.g. BLEU, ROUGE, BERTScore, backtranslation likelihood, and textual entailment signals). This teaches the model general similarity assessment and makes it robust to distribution shift. (2) Fine-tuning on a small set of human ratings (WMT Metrics Shared Task, WebNLG). At inference the model takes a (candidate, reference) pair, encodes it with the BERT/RemBERT encoder, and a regression head over the sentence representation returns a single number — the predicted quality score (in practice roughly 0 for a random output and 1 for a perfect one).
Traditional metrics such as BLEU and ROUGE measure surface-level n-gram overlap between a model output and a reference, so they correlate poorly with human judgment and fail to capture paraphrases or semantic equivalence. BLEURT addresses this by learning to predict human ratings and to capture meaning-level similarity rather than literal word overlap, while remaining robust to out-of-distribution data.
A pre-trained Transformer encoder that encodes the (candidate, reference) pair into contextual representations. Monolingual checkpoints use BERT, while the multilingual BLEURT-20 uses RemBERT.
Official
A linear layer on top of the sentence representation (the [CLS] token) that maps the representation to a single number — the predicted quality score.
The pre-training stage on millions of synthetic sentence pairs (Wikipedia perturbations) automatically labelled by existing metrics and models (BLEU, ROUGE, BERTScore, backtranslation, textual entailment). Provides robustness and generalization with few human ratings.
Official
Fine-tuning on a small set of real human ratings (WMT Metrics Shared Task, WebNLG), calibrating predictions to actual quality judgments.
Official
Raw BLEURT values are regression predictions and can fall outside the 0–1 range; ~0 and ~1 are only rough reference points (random vs perfect output).
Different checkpoints (BERT-base, BERT-large, BLEURT-20, BLEURT-20-D12) produce values on different scales; comparing scores across checkpoints is invalid.
BLEURT requires a reference and is fine-tuned on human ratings (mostly WMT), so it can carry domain/language biases and perform worse out of distribution.
Sellam, Das and Parikh introduce a learned BERT-based metric, pre-trained on synthetic sentence pairs and fine-tuned on human ratings, with higher correlation to human judgment than BLEU.
The paper 'Learning Compact Metrics for MT' (Pu et al., EMNLP 2021) advances RemBERT-based multilingual metrics and distillation; it underpins the recommended BLEURT-20 checkpoint and its lighter variants (e.g. BLEURT-20-D12).
Time complexity: O(n² · d). Space complexity: O(n² + n · d).
Choice of the pre-trained encoder: BERT (monolingual) or RemBERT (multilingual, BLEURT-20). Determines language coverage and quality.
Number of parameters / layers. Larger checkpoints yield better correlation at the cost of speed; distilled variants (e.g. BLEURT-20-D12) lower cost.
Token limit for the concatenated candidate + reference pair; longer inputs are truncated, which can affect scoring of very long texts.
BLEURT uses a dense Transformer encoder (BERT/RemBERT) — all parameters are active for every input; there is no conditional routing.
The encoder processes all tokens of the pair in parallel in a single pass; scoring many pairs batches easily.
Scoring is a Transformer encoder forward pass; GPUs with tensor cores greatly accelerate matrix multiplications and batching of pairs.
The reference implementation uses TensorFlow and maps well onto TPUs for training and large-scale scoring.
BLEURT runs on CPU but slower; useful for small numbers of pairs or when no accelerator is available.
