BLEU的php实现

参考:http://www.nltk.org/_modules/nltk/align/bleu_score.htmlhttps://blog.csdn.net/Allocator/article/details/79657792https://blog.csdn.net/guolindonggld/article/details/56966200

参考NLTK和BLEU的公式实现的php版本BLEU计算,目前能够计算一个candidate对应一个reference的BLEU。(虽然通常会有多个reference)

贴下代码

class BLEU
{
    public function bleu_score($candidate, $reference)
    {
        // 计算BP惩罚因子
        $bp = $this->brevity_penalty($candidate, $reference);
        // 计算修正的n元精确率(Modified n-gram Precision)
        $mp = $this->modified_precision($candidate, $reference, 3);
        return $bp*exp($mp);
    }

    /**
     * 计算BP
     * @param $candidate
     * @param $reference
     * @return float
     */
    private function brevity_penalty($candidate, $reference): float
    {
        $candidate_ = explode(" ", $candidate);
        $reference_ = explode(" ", $reference);

        $can_length = sizeof($candidate_);
        $ref_length = sizeof($reference_);

        if ($can_length >= $ref_length) {
            return 1.0;
        } else {
            // BP = e^(1-r/c)
            return exp((1 - $ref_length / $can_length));
        }
    }

    /**
     * 计算MP
     * @param $candidate
     * @param $reference
     * @param $n
     * @return float|int
     */
    private function modified_precision($candidate, $reference, $n = 3)
    {
        $precisions = [];
        for ($i=1; $i<=$n; $i++) {
            // 分别计算每个ngram的precision
            $candidate_grams = $this->ngram($candidate, $i);
            $reference_grams = $this->ngram($reference, $i);
            $can_gram_count = 0;
            foreach ($candidate_grams as $gram) {
                $can_gram_count += substr_count($candidate, $gram);
            }
            $ref_gram_count = 0;
            foreach ($reference_grams as $gram) {
                $ref_gram_count += substr_count($reference, $gram);
            }
            $p = $ref_gram_count / $can_gram_count;
            array_push($precisions, $p);
        }

        return $this->log_sum($precisions) / sizeof($precisions);
    }

    /**
     * 计算数组中每个数log后的加和
     * @param array $precisions
     * @return float|int
     */
    private function log_sum(array $precisions)
    {
        $sum = 0;
        foreach ($precisions as $precision) {
            $sum += log($precision);
        }
        return $sum;
    }

    /**
     * 将一个句子分解成ngram
     * @param $sentence
     * @param $n
     * @return array
     */
    private function ngram($sentence, $n)
    {
        $sentence_ = explode(" ", trim($sentence));
        $ngrams = [];
        foreach ($sentence_ as $idx => $gram) {
            if ($idx <= sizeof($sentence_) - $n) {
                $ngram = "";
                for ($i = $idx; $i<$idx+$n; $i++) {
                    $ngram .= " " .$sentence_[$i];
                }
                $ngram = ltrim($ngram);
                array_push($ngrams, $ngram);
            }
        }
        return $ngrams;
    }
}

食用方法

(new BLEU())→bleu_score($candidate, $reference)

  • 1
    点赞
  • 0
    收藏
    觉得还不错? 一键收藏
  • 1
    评论
评论 1
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值