基于bertService的二次精排

一、bertService安装

可以自行百度，网络安装方案很多

二、bertService启动

# -*- coding: utf-8 -*-
from bert_serving.server import BertServer
from bert_serving.server.helper import get_args_parser
def main():
    args = get_args_parser().parse_args(['-model_dir', '/Users/alibaba/Downloads/fintune_model',
                                         '-port', '86500',
                                         '-port_out', '86501',
                                         '-max_seq_len', '512',
                                         '-num_work', '2',
                                         '-cpu'])

    bs = BertServer(args)
    bs.start()
if __name__ == "__main__":
    main()

注意事项 -model_dir：需要填写你本地model的地址，显示如下表示服务启动成功

 三、bertService二次排序

# 导入bert客户端
from bert_serving.client import BertClient
import numpy as np
import pandas as pd

class SimilarModel:
    def __init__(self):
        # ip默认为本地模式，如果bert服务部署在其他服务器上，修改为对应ip
        self.bert_client = BertClient(port=86500, port_out=86501, show_server_config=True, timeout=1000000)

    def close_bert(self):
        self.bert_client.close()

    def get_sentence_vec(self,sentence):
        '''
        根据bert获取句子向量
        :param sentence:
        :return:
        '''
        return self.bert_client.encode([sentence])[0]

    def cos_similar(self,sen_a_vec, sen_b_vec):
        '''
        计算两个句子的余弦相似度
        :param sen_a_vec:
        :param sen_b_vec:
        :return:
        '''
        vector_a = np.mat(sen_a_vec)
        vector_b = np.mat(sen_b_vec)
        num = float(vector_a * vector_b.T)
        denom = np.linalg.norm(vector_a) * np.linalg.norm(vector_b)
        cos = num / denom
        return cos

if __name__=='__main__':
    # 从候选集condinates 中选出与sentence_a 最相近的句子
    query = open("/Users/alibaba/Desktop/query.txt",'r').readlines()
    condinates = open("/Users/alibaba/Desktop/recall.txt",'r').readlines()
    sentence_a = query[0]
    bert_result = []
    bm25_score = []
    fir_recall_title_list = []
    query_list = []
    bert_client = SimilarModel()
    i = 0
    for sentence_b in condinates:
        i = i+1
        print("计算第  " + str(i)   + "  个文本相似度")
        query_list.append( sentence_a )
        sentence_a_vec = bert_client.get_sentence_vec(sentence_a)
        fir_recall_title = sentence_b.split("    ")[1]
        fir_recall_title_list.append(fir_recall_title)
        bm25_score.append( sentence_b.split("    ")[2].replace("\n", "") )
        fir_recall_title_vec = bert_client.get_sentence_vec(fir_recall_title)
        cos_sim = bert_client.cos_similar(sentence_a_vec,fir_recall_title_vec)
        bert_result.append(round(cos_sim, 3))
    bert_client.close_bert()
    second_sort_res = pd.DataFrame(
                {
                    "query": query_list,
                    "reacll": fir_recall_title_list,
                    "bm25_score": bm25_score,
                    "bert_result": bert_result
                }
            )
    second_sort_res['bm25_score'] = second_sort_res['bm25_score'].astype(float)
    second_sort_res['bert_result'] = second_sort_res['bert_result'].astype(float)
    if  '停车' in second_sort_res['query'][0]:
        second_sort_res = second_sort_res.sort_values(by="bert_result" , ascending=False)
    else:
        second_sort_res = second_sort_res.sort_values(by="bm25_score" , ascending=False)
    second_sort_res = second_sort_res.drop_duplicates(inplace=False)
    second_sort_res.iloc[0:50,:].to_csv("/Users/alibaba/Desktop/second_sort_resut.csv",index=False)
    print("二次排序完成, 召回排名前 " + str(len(condinates)) + " 工单")

四、粗排参考博文

Lucene bm25 结合 jieba中文分词搜索_深挖技术点滴-CSDN博客_jieba lucene

五、思考（bert是否适合做相似度计算）

1. bert模型不适合做搜索召回，因为它的机理实现的是基于类别相似最小、不相似最大，通过此类方法进行fintune后，结果的间隙不明显，无法区分
2. 网上资料可以用bert各个pool层输出作为相似度计算，bert-as-service采用的是这种
3. 可以考虑使用双塔模型做搜索召回