sentenceGLEU static method

double sentenceGLEU(List<List<String>> references, List<String> hypothesis, { int minN = 1, int maxN = 4, })

Computes the sentence-level GLEU score.

references are one or more reference token lists. hypothesis is the generated token list. minN and maxN define the n-gram size range (default 1–4).

Implementation

static double sentenceGLEU(
  List<List<String>> references,
  List<String> hypothesis, {
  int minN = 1,
  int maxN = 4,
}) {
  if (references.isEmpty) throw ArgumentError('references cannot be empty.');
  if (hypothesis.isEmpty) throw ArgumentError('hypothesis cannot be empty.');

  final hypCounts = MatchCounter<NGram<String>>(
      hypothesis.createAllNGrams(minN: minN, maxN: maxN));
  final truePosFalsePos = hypCounts.sum();

  var corpusNMatch = 0;
  var corpusNAll = 0;

  for (final reference in references) {
    final refCounts = MatchCounter<NGram<String>>(
        reference.createAllNGrams(minN: minN, maxN: maxN));
    final truePosFalseNeg = refCounts.sum();
    final overlapCounts = hypCounts.intersect(refCounts);
    final truePos = overlapCounts.sum();
    final nAll = math.max(truePosFalsePos, truePosFalseNeg);
    if (nAll > 0) {
      corpusNMatch += truePos;
      corpusNAll += nAll;
    }
  }

  if (corpusNAll == 0) return 0.0;
  return corpusNMatch / corpusNAll;
}