短文本转向量的一种实现方式

文章目录

    • 前言
    • 实现思路
    • word2vec相关配置
    • 代码
    • 补充资料
    • 完整代码

前言

下文实现仅仅是比较粗糙的一种方式,可以改进的点还有很多,是真的很多!重点是,不讲解原理,就是这么没道理…

实现思路

  1. 分词。分词还是jieba好。word2vec模型训练选取gensim。
  2. 使用大语料进行基础词典word2vec模型的训练。
  3. 使用特定领域(针对业务)语料对word2vec模型进行增量训练。
  4. 文本分词后使用AVG-W2V方式获取短文本向量,维度取决于word2vec维度大小,即所有词向量求平均。

word2vec相关配置

  • w2v.properties
#一些经验
#架构(sg):skip-gram(慢、对罕见字有利)vs CBOW(快)
#训练算法(hs):分层softmax(对罕见字有利)vs 负采样(对常见词和低纬向量有利)
#欠采样频繁词(sample):可以提高结果的准确性和速度(适用范围1e-3到1e-5)
#文本大小(window):skip-gram通常在10附近,CBOW通常在5附近
#大语料下,建议提高min_count,减少iter
#内存占用大约公式:词汇数*8*size/1000/1000/1000(GB)
#硬盘占用大约公式:词汇数*8/1000/1000/1000(GB)(实际上考虑到其模型的其他文件,最好再*10的大小)

# 训练算法,0为CBOW算法,1为skip-gram算法,默认为0
sg=1
# 特征向量的维度
size=300
# 词窗大小
window=5
# 最小词频
min_count=5
# 初始学习速率
alpha=0.025
# 0为负采样,1为softmax,默认为0
hs=1
#迭代次数
iter=10

代码

  • 大语料基础训练相关代码
# -*- coding:utf-8 -*-
"""
Description: 基于百度百科大语料的word2vec模型

@author: WangLeAi
@date: 2018/9/18
"""
import os
from util.DBUtil import DbPoolUtil
from util.JiebaUtil import jieba_util
from util.PropertiesUtil import prop
from gensim.models import word2vec


class OriginModel(object):
    def __init__(self):
        self.params = prop.get_config_dict("config/w2v.properties")
        self.db_pool_util = DbPoolUtil(db_type="mysql")
        self.train_data_path = "gen/ori_train_data.txt"
        self.model_path = "model/oriw2v.model"

    @staticmethod
    def text_process(sentence):
        """
        文本预处理
        :param sentence:
        :return:
        """
        # 过滤任意非中文、非英文、非数字
        # regex = re.compile(u'[^\u4e00-\u9fa50-9a-zA-Z\-·]+')
        # sentence = regex.sub('', sentence)
        words = jieba_util.jieba_cut(sentence)
        return words

    def get_train_data(self):
        """
        获取训练数据,此处需要自行修改,最好写入文件而不是直接取到内存中!!!!!
        :return:
        """
        print("创建初始语料训练数据")
        sql = """ """
        sentences = self.db_pool_util.loop_row(origin_model, "text_process", sql)
        with open(self.train_data_path, "w", encoding="utf-8") as f:
            for sentence in sentences:
                f.write(" ".join(sentence) + "\n")

    def train_model(self):
        """
        训练模型
        :return:
        """
        if not os.path.exists(self.train_data_path):
            self.get_train_data()
        print("训练初始模型")
        sentences = word2vec.LineSentence(self.train_data_path)
        model = word2vec.Word2Vec(sentences=sentences, sg=int(self.params["sg"]), size=int(self.params["size"]),
                                  window=int(self.params["window"]), min_count=int(self.params["min_count"]),
                                  alpha=float(self.params["alpha"]), hs=int(self.params["hs"]), workers=6,
                                  iter=int(self.params["iter"]))
        model.save(self.model_path)
        print("训练初始模型完毕,保存模型")


origin_model = OriginModel()

  • 额外语料进行训练
# -*- coding:utf-8 -*-
"""
Description:word2vec fine tuning
基于对应类型的额外语料进行微调

@author: WangLeAi
@date: 2018/9/11
"""
import os
from util.DBUtil import DbPoolUtil
from util.JiebaUtil import jieba_util
from util.PropertiesUtil import prop
from gensim.models import word2vec
from algorithms.OriginModel import origin_model


class Word2VecModel(object):
    def __init__(self):
        self.db_pool_util = DbPoolUtil(db_type="mysql")
        self.train_data_path = "gen/train_data.txt"
        self.origin_model_path = "model/oriw2v.model"
        self.model_path = "model/w2v.model"
        self.model = None
        # 未登录词进入需考虑最小词频
        self.min_count = int(prop.get_config_value("config/w2v.properties", "min_count"))

    @staticmethod
    def text_process(sentence):
        """
        文本预处理
        :param sentence:
        :return:
        """
        # 过滤任意非中文、非英文、非数字等
        # regex = re.compile(u'[^\u4e00-\u9fa50-9a-zA-Z\-·]+')
        # sentence = regex.sub('', sentence)
        words = jieba_util.jieba_cut(sentence)
        return words

    def get_train_data(self):
        """
        获取训练数据,此处需要自行修改,最好写入文件而不是直接取到内存中!!!!!
        :return:
        """
        print("创建额外语料训练数据")
        sql = """ """
        sentences = self.db_pool_util.loop_row(w2v_model, "text_process", sql)
        with open(self.train_data_path, "a", encoding="utf-8") as f:
            for sentence in sentences:
                f.write(" ".join(sentence) + "\n")

    def train_model(self):
        """
        训练模型
        :return:
        """
        if not os.path.exists(self.origin_model_path):
            print("无初始模型,进行初始模型训练")
            origin_model.train_model()
        model = word2vec.Word2Vec.load(self.origin_model_path)
        print("初始模型加载完毕")
        if not os.path.exists(self.train_data_path):
            self.get_train_data()
        print("额外语料训练")
        extra_sentences = word2vec.LineSentence(self.train_data_path)
        model.build_vocab(extra_sentences, update=True)
        model.train(extra_sentences, total_examples=model.corpus_count, epochs=model.iter)
        model.save(self.model_path)
        print("额外语料训练完毕")

    def load_model(self):
        """
        载入模型
        :return:
        """
        print("载入词嵌入模型")
        if not os.path.exists(self.model_path):
            print("无词嵌入模型,进行训练")
            self.train_model()
        self.model = word2vec.Word2Vec.load(self.model_path)
        print("词嵌入模型加载完毕")

    def get_word_vector(self, words, extra=0):
        """
        获取词语向量,需要先载入模型
        :param words:
        :param extra:是否考虑未登录词,0不考虑,1考虑
        :return:
        """
        if extra:
            if words not in self.model:
                more_sentences = [[words, ] for i in range(self.min_count)]
                self.model.build_vocab(more_sentences, update=True)
                self.model.train(more_sentences, total_examples=self.model.corpus_count, epochs=self.model.iter)
                self.model.save(self.model_path)
        rst = None
        if words in self.model:
            rst = self.model[words]
        return rst

    def get_sentence_vector(self, sentence, extra=0):
        """
        获取文本向量,需要先载入模型
        :param sentence:
        :param extra: 是否考虑未登录词,0不考虑,1考虑
        :return:
        """
        words = jieba_util.jieba_cut_flag(sentence)
        if not words:
            words = jieba_util.jieba_cut(sentence)
        if not words:
            print("存在无法切出有效词的句子:" + sentence)
            # raise Exception("存在无法切出有效词的句子:" + sentence)
        if extra:
            for item in words:
                if item not in self.model:
                    more_sentences = [words for i in range(self.min_count)]
                    self.model.build_vocab(more_sentences, update=True)
                    self.model.train(more_sentences, total_examples=self.model.corpus_count, epochs=self.model.iter)
                    self.model.save(self.model_path)
                    break
        return self.get_sentence_embedding(words)

    def get_sentence_embedding(self, words):
        """
        获取短文本向量,仅推荐短文本使用
        句中所有词权重总和求平均获取文本向量,不适用于长文本的原因在于受频繁词影响较大
        长文本推荐使用gensim的doc2vec
        :param words:
        :return:
        """
        count = 0
        vector = None
        for item in words:
            if item in self.model:
                count += 1
                if vector is not None:
                    vector = vector + self.model[item]
                else:
                    vector = self.model[item]
        if vector is not None:
            vector = vector / count
        return vector


w2v_model = Word2VecModel()

  • 测试方式
# -*- coding:utf-8 -*-
"""
Description:

@author: WangLeAi
@date: 2018/9/18
"""
import os
from algorithms.Word2VecModel import w2v_model


def main():
    root_path = os.path.split(os.path.realpath(__file__))[0]
    if not os.path.exists(root_path + "/model"):
        os.mkdir(root_path + "/model")
    w2v_model.load_model()
    print(w2v_model.get_sentence_vector("不知不觉间我已经忘记了爱"))


if __name__ == "__main__":
    main()

补充资料

  1. 文本相似度算法相关资料(力推!):戳我
  2. DBUtils相关内容可以看我之前的博文,有一点小改动:戳我

完整代码

下载地址

你可能感兴趣的:(人工智能)