Struct Bleu 

pub struct Bleu { /* private fields */ }

BLEU evaluator (BLEU-4 by default).

Implementations

impl Bleu

pub fn new() -> Self

Creates the default BLEU-4 evaluator.

pub fn with_max_n(self, n: usize) -> Self

Uses BLEU-n (default 4)

pub fn with_char_level(self, v: bool) -> Self

Character-level tokenization: whitespace-less languages such as Chinese split per char (otherwise the whole sentence becomes one token and BLEU breaks)

pub fn with_smoothing(self, v: bool) -> Self

Enables smoothing: an order with no n-gram match gets a small value instead of a whole-zero, so short sentences are not cut off wholesale

pub fn corpus_bleu( &self, predictions: &[&str], references: &[&str], ) -> Result<f64, EvalError>

Corpus-level BLEU: aggregates n-gram match counts across examples, then computes a single geometric mean + corpus-level brevity penalty.

P2-1: a sentence-level brevity penalty cuts short sentences off wholesale (e.g. “the cat” scores a hard zero under BLEU-4); corpus aggregation computes the penalty from total lengths and merges per-order n-gram counts, so short sentences still contribute low-order precision. With with_smoothing, an order with no match gets a small value instead of a whole zero.

predictions and references must have the same length (one-to-one), otherwise EvalError::LengthMismatch is returned.

Trait Implementations

impl Default for Bleu

fn default() -> Self

Returns the “default value” for a type.

impl Evaluator for Bleu

fn eval<'life0, 'life1, 'life2, 'life3, 'async_trait>( &'life0 self, _input: &'life1 str, prediction: &'life2 str, reference: &'life3 str, ) -> Pin<Box<dyn Future<Output = Result<Score, EvalError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait, 'life3: 'async_trait,

Scores a single prediction

fn name(&self) -> &str

Evaluator name (used in report summaries)