基于ES实现电商语义搜索

整体效果

背景描述

在电商领域，如果仅仅依赖传统词匹配方式进行商品检索，会导致一些商品召回错误影响用户体验；如搜索【蛋糕奶油】结果查询出“奶油蛋糕“相关商品，搜索【车手】相关的宣传海报，结果找到了“汽车手机版”宣传海报

架构设计

特征提取部分采用传统的DSSM模型进行相似度训练，正样本数据集来自七日内用户搜索词和点击商品的标题及标签，负样本来自正样本随机组合生成的数据

提取后的特征采用ES的dense_vector来存储，目前最新版本支持knn检索，较传统script查询性能提升不少；索引结构如下：


PUT my-index
{
  "mappings": {
    "properties": {
      "doc_vector": {
        "type": "dense_vector",
        "dims": 128,
        "index": true,
        "similarity": "cosine"
      },
      "title": {
        "type": "text",
        "analyzer": "ik_max_word"
      },
      "big_tags": {
        "type": "text",
        "analyzer": "ik_max_word"
      },
      "id": {
        "type": "keyword" 
      }
    }
  }
}

工程实现

准备数据

从日志里获取搜索词和商品id

import requests

url = "http://****:8123/?user=&password=&database=app"

rep = requests.post(url, data="SELECT  decodeURLComponent(extractURLParameter(ref, 'keywords')) kw, splitByChar('/', path(url))[-1] id from app.scene_tracker where e_t='page_view' and p_l='PC' and ref like '%keywords%' and (url like '%detail%' or url like '%create%' ) and length(id)>4  ").text

kw_array = []
id_array = []
for line in rep.split('\n'):
    array = line.split('\t')
    try:
        id_array.append(array[1])
        kw_array.append(array[0])
    except:
        print(line)

补全标题、标签等信息

import json
headers={
    "Content-Type": "application/json"
}

temp='''
{
    "_source": ["big_tags", "id", "title"],
    "size":10000,
  "query": {
    "ids": {
     "values": {}
    }
  }
}'''
id_set = set(id_array)
id_map_tit = {}
for i in range(0, len(id_set), 10000):
    query = temp.replace('{}', str(list(id_set)[i:i+10000]).replace("'","\""))
    response = requests.post("http://****:9200/index/_search", data=query.encode(), auth=('guoyanchao',''), headers=headers)
    json_data = json.loads(response.text)
#     print(response.text)
    for item in json_data.get('hits').get('hits'):
        doc = item.get('_source')
        id_map_tit[item.get('_id')]=doc.get('title')+' '+doc.get('big_tags')

基于正样本构建负样本

import random
import numpy as np
from collections import Counter

query_total = []
id_total = []

for query, id in zip(kw_array, id_array):
    if id not in id_map_tit or query=='':
        continue
    id_total.append(id)
    query_total.append(query)
    
lab_total = np.ones(len(query_total), dtype=int)   
cn = len(lab_total) 
ids = list(id_map_tit.keys())
for i in range(cn):
    k = query_total[i] 
    query_total.append(k)
    id_total.append(ids[random.randint(0,len(id_map_tit)-1)])

lab_total=np.append(lab_total, np.zeros(cn, dtype=int))

特征转换

中文分词，这里采用的是百度的lac，大家也可以使用其它分词器，或者单字分词（考虑到中文词组合无穷尽，大家线上可采用单字分词约1.5w维，个人也推荐使用这种方式虽然词义表达上有所降低，但更便于线上维护）

from tensorflow.keras.preprocessing.text import Tokenizer
from LAC import LAC
lac = LAC(mode='seg')

query_terms = []
title_terms = []
for query, id in zip(query_total, id_total):
    words = []
    title = id_map_tit[id]
    query_terms.append(lac.run(query))
    title_terms.append(lac.run(title))

# MAX_WORDS=15
tokenizer = Tokenizer()
tokenizer.fit_on_texts(query_terms+title_terms)
vocab = tokenizer.word_index
vocab

生成词序列：通过词袋将中文映射为数字，并拓展为固定序列，便于模型输入

from tensorflow.keras.preprocessing.sequence import pad_sequences

query_max_len = 16
doc_max_len = 128
# query_vec = pad_sequences(query_terms, dtype='object', maxlen=query_max_len).tolist()
# doc_vec = pad_sequences(title_terms, dtype='object', maxlen=doc_max_len).tolist()
query_seq = tokenizer.texts_to_sequences(query_terms)
query_vec = pad_sequences(query_seq, maxlen=query_max_len)
doc_seq = tokenizer.texts_to_sequences(title_terms)
doc_vec = pad_sequences(doc_seq, maxlen=doc_max_len)

构建双塔模型并训练
考虑到输入都是短文本，故表示层才用了经典的长短记忆模型LSTM来构建双塔用于特征提取，输出层采用的是余弦相似，大家也可以求解向量点积，模型每日更新

import tensorflow as tf
from sklearn.model_selection import train_test_split 
embed_dim = 64
NEG, batch_size = 20, 128
 
class CosineLayer():
    """ 自定义batch内负采样并做cosine相似度的层 """
    def __call__(self, inputs):
        def _cosine(x):
            query_encoder, doc_encoder = x
            doc_encoder_fd = doc_encoder
            for i in range(NEG):
                ss = tf.gather(doc_encoder, tf.random.shuffle(tf.range(tf.shape(doc_encoder)[0])))
                doc_encoder_fd = tf.concat([doc_encoder_fd, ss], axis=0)
            query_norm = tf.tile(tf.sqrt(tf.reduce_sum(tf.square(query_encoder), axis=1, keepdims=True)),[NEG + 1, 1])
            doc_norm = tf.sqrt(tf.reduce_sum(tf.square(doc_encoder_fd), axis=1, keepdims=True))
            query_encoder_fd = tf.tile(query_encoder, [NEG + 1, 1])
            prod = tf.reduce_sum(tf.multiply(query_encoder_fd, doc_encoder_fd, name="sim-multiply"), axis=1, keepdims=True)
            norm_prod = tf.multiply(query_norm, doc_norm)
            cos_sim_raw = tf.truediv(prod, norm_prod)
            cos_sim = tf.transpose(tf.reshape(tf.transpose(cos_sim_raw), [NEG + 1, -1])) * 20
            
            prob = tf.nn.softmax(cos_sim, name="sim-softmax")
            hit_prob = tf.slice(prob, [0, 0], [-1, 1], name="sim-slice")
            loss = -tf.reduce_mean(tf.math.log(hit_prob), name="sim-mean")
            return loss
        output_shape = (1,)
        value = tf.keras.layers.Lambda(_cosine, output_shape=output_shape)([inputs[0], inputs[1]])
        return value

    
query_input = tf.keras.layers.Input(shape=(query_max_len, ), name="query_input")
doc_input = tf.keras.layers.Input(shape=(doc_max_len, ), name="doc_input")
 
embedding = tf.keras.layers.Embedding(len(vocab)+1, embed_dim)
query_embed = embedding(query_input)
doc_embed = embedding(doc_input)

query_encoder = tf.keras.layers.LSTM(128, name="query_tower")(query_embed)
doc_encoder = tf.keras.layers.LSTM(128, name="doc_tower")(doc_embed)
 
cos_sim = CosineLayer()([query_encoder, doc_encoder])
 
# output = tf.keras.layers.Dense(1, activation='sigmoid')(output)
model = tf.keras.models.Model(inputs=[query_input, doc_input], outputs=cos_sim)
# model.compile(loss="binary_crossentropy", optimizer='adam')
model.compile(optimizer="adam", loss=lambda y_true, y_pred: y_pred)

# query tower
query_model = tf.keras.models.Model(inputs=query_input, outputs=query_encoder)
# doc tower
doc_model = tf.keras.models.Model(inputs=doc_input, outputs=doc_encoder)


train_X, test_X, train_y, test_y = train_test_split(list(zip(query_vec, doc_vec)), lab_total, test_size=0.3)
# 将每一列转化为nparray进行输入
train_q, train_p = zip(*train_X)
train_inputs = [ np.array(train_q), np.array(train_p) ]
test_q, test_p = zip(*test_X)
test_inputs = [ np.array(test_q), np.array(test_p) ]
print(train_q[:1], train_p[:1])
# train the model
print("[INFO] training model...")
model.fit(
    train_inputs, train_y,
    validation_data=(test_inputs, test_y),
    epochs=40, batch_size=32, verbose=1)

from tensorflow.keras.utils import plot_model
plot_model(model, to_file='model.png', show_shapes=True)

导出特征

通过获取训练好的doc_tower将商品特征导入到ES，另外也可以保存 doc_model 用于实时索引商品特征

import requests
 
for id, vec in dict(zip(id_total,doc_vec)).items():
    doc_embedding = doc_model(np.reshape(vec, [1, doc_max_len]) )
    data='''{{
      "id":"{}",
      "title":"{}",
      "doc_vector":{}
    }}'''.format(id, id_map_tit[id], list(doc_embedding.numpy().flatten()))

    headers={
        "Content-Type": "application/json"
    }
    response = requests.post("http://****:9200/my-index/_doc/"+str(id), data=data.encode(), auth=('guoyanchao',''), headers=headers)

线上服务

获取query特征

import requests
from tensorflow.keras.preprocessing.sequence import pad_sequences
from LAC import LAC
lac = LAC(mode='seg')

query_layer_model = tf.keras.models.Model(
    inputs=[query_input],
    outputs=model.get_layer(name='query_tower').output
)
query = "婚礼邀请函"
query_seq = tokenizer.texts_to_sequences([lac.run(query)])
qvec = pad_sequences(query_seq, maxlen=query_max_len).tolist() 
 
query_embedding = query_layer_model(np.reshape(qvec, [1,query_max_len]))
query_embedding.numpy().flatten()

ES查询语句，这里使用knn检索其中filter部分可搭配其它DSL语句做商品的近一步筛选以提升匹配精度

GET my-index/_knn_search
{
  "knn": {
    "field": "doc_vector",
    "query_vector":[],
    "k": 20,
    "num_candidates": 1000
  } ,
  "filter": {
    "match": {
      "title": ""
    }
  },
  "_source": ["title" ]
}

传统的script检索

GET my-index/_search
{
  "_source": ["title"], 
  "query": {
    "script_score": {
      "query" : {
         "match_all": {}
      },
      "script": {
        "source": """
          double value = dotProduct(params.query_vector, 'doc_vector');
          return sigmoid(1, Math.E, -value); 
        """,
        "params": {
          "query_vector": []
        }
      }
    }
  }
}

后期优化

目前模型数据只是考虑了query和商品文本描述，后期会增加更多维度的训练特征，如用户短期兴趣偏好特征，商品属性、价格以及封面图特征等
匹配层采用欧式距离要优于点积和余弦

# -- 欧式距离
output = tf.sqrt(tf.reduce_sum(tf.square(query_encoder - doc_encoder), axis = 1))
output = tf.expand_dims(output, 1)
output = tf.keras.layers.Dense(1, activation='sigmoid')(output)

表示层优化模型结构，将LSTM换成双向结构

forward_layer = tf.keras.layers.LSTM(128, return_sequences=True)
backward_layer = tf.keras.layers.LSTM(128, activation='tanh', return_sequences=True, go_backwards=True)

query_encoder = tf.keras.layers.Bidirectional(forward_layer, backward_layer=backward_layer)(query_embed)
query_encoder = tf.keras.layers.Flatten()(query_encoder)
merged = tf.keras.layers.Dropout(0.1)(query_encoder)
merged = tf.keras.layers.BatchNormalization()(merged)
merged = tf.keras.layers.Dense(128, activation='tanh')(merged)
query_encoder = tf.keras.layers.Dense(128, activation='tanh', name="query_tower")(merged)
  
doc_encoder = tf.keras.layers.Bidirectional(forward_layer, backward_layer=backward_layer)(doc_embed)
doc_encoder = tf.keras.layers.Flatten()(doc_encoder)
merged = tf.keras.layers.Dropout(0.1)(doc_encoder)
merged = tf.keras.layers.BatchNormalization()(merged)
merged = tf.keras.layers.Dense(128, activation='tanh')(merged)
doc_encoder = tf.keras.layers.Dense(128, activation='tanh', name="doc_tower")(merged)
 
output = tf.reduce_sum(query_encoder*doc_encoder, axis = 1)
output = tf.expand_dims(output, 1)
  
output = tf.keras.layers.Dense(1, activation='sigmoid', use_bias=False)(output)

模型结构