【python】测试中文文本的BLEU

解决问题：测试中文文本的BLEU指标，可用于评估大模型性能。

# -*- coding: utf-8 -*-
import nltk
from nltk.translate.bleu_score import SmoothingFunction
import jieba

# 将句子分词并转换为n-gram格式
def sentence_to_ngrams(sentence, n):
    words = jieba.lcut(sentence)
    return set(nltk.ngrams(words, n))

# 计算BLEU指标
def calculate_bleu(reference, candidate, max_n=4):
    smooth = SmoothingFunction().method4
    scores = []
    for n in range(1, max_n + 1):
        reference_ngrams = sentence_to_ngrams(reference, n)
        print("ref-------------------")
        print(reference_ngrams)
        print("-------------------")
        candidate_ngrams = sentence_to_ngrams(candidate, n)
        print("can-------------------")
        print(candidate_ngrams)
        print("-------------------")
        scores.append(nltk.translate.bleu_score.sentence_bleu([reference_ngrams], candidate_ngrams, smoothing_function=smooth))
        

    return scores

# 测试数据
reference = "桌子上有只猫"
candidate = "这个桌子上有个狸花猫"
# 计算BLEU指标
bleu_scores = calculate_bleu(reference, candidate)

# 输出结果
for n, score in enumerate(bleu_scores, start=1):
    print(f'BLEU-{n} Score:', score)

运行结果
BLEU-1

BLEU-2

BLEU-3

BLEU-4

得分