NLP trains a model (LSI, LDA, TFIDF) that can find the most similar sentence

A full set of codes, not much explanation, plug and play~

English sentence preprocessing module

# 英文句子处理模块
from nltk.corpus import stopwords as pw
import sys 
import re
cacheStopWords=pw.words("english")

def English_processing(sentence):
    if sentence:
        sentence = sentence.lower()  # 大写转小写
           
        for ch in "“”!?.\;'',()<>\{}/-1234567890$&#%~":
            sentence = sentence.lower().replace(ch," ")  # 去除符号
            
        sentence=''.join([word+" " for word in sentence.split() if word not in cacheStopWords]) # 去除停用词
        
        sentence=''.join([word+" " for word in sentence.split() if word not in ['br','w','b','bc']]) # 去除指定特殊词
        
        return sentence

Similarity model training-evaluation

import gc
import tqdm
import numpy as np
from gensim import corpora, models, similarities
from collections import defaultdict
import time


class SentenceSimilarity():

    def __init__(self, sentences,min_frequency= 1):
        self.sentences = []
        for i in range(0, len(sentences)):
            self.sentences.append(English_processing(sentences[i]))
        self.sentences_num = len(self.sentences)
        
        self.min_frequency = min_frequency

    # 获取切过词的句子
    def get_cuted_sentences(self):
        cuted_sentences = []

        for sentence in self.sentences:
            cuted_sentences.append(sentence.strip().split())

        return cuted_sentences

    # 构建其他复杂模型前需要的简单模型
    def simple_model(self):
        self.texts = self.get_cuted_sentences()
        
        # 删除低频词
        frequency = defaultdict(int)
        for text in self.texts:
            for token in text:
                frequency[token] += 1
        self.texts = [[token for token in text if frequency[token] > self.min_frequency] for text in self.texts]
        self.dictionary = corpora.Dictionary(self.texts)
        
        self.corpus_simple = [self.dictionary.doc2bow(text) for text in self.texts]

    # tfidf模型
    def TfidfModel(self):
        self.simple_model()

        # 转换模型
        self.model = models.TfidfModel(self.corpus_simple)
        self.corpus = self.model[self.corpus_simple]

        # 创建相似度矩阵
        self.index = similarities.MatrixSimilarity(self.corpus)

    # lsi模型
    def LsiModel(self):
        self.simple_model()

        # 转换模型
        self.model = models.LsiModel(self.corpus_simple)
        self.corpus = self.model[self.corpus_simple]

        # 创建相似度矩阵
        self.index = similarities.MatrixSimilarity(self.corpus)

    # lda模型
    def LdaModel(self):
        self.simple_model()

        # 转换模型
        self.model = models.LdaModel(self.corpus_simple)
        self.corpus = self.model[self.corpus_simple]

        # 创建相似度矩阵
        self.index = similarities.MatrixSimilarity(self.corpus)

    # 对新输入的句子（比较的句子）进行预处理
    def sentence2vec(self, sentence):
        sentence = English_processing(sentence)
        vec_bow = self.dictionary.doc2bow(sentence.strip().split())
        return self.model[vec_bow]

    def bow2vec(self):
        vec = []
        length = max(self.dictionary) + 1
        for content in self.corpus:
            sentence_vectors = np.zeros(length)
            for co in content:
                sentence_vectors[co[0]] = co[1]  # 将句子出现的单词的tf-idf表示放入矩阵中
            vec.append(sentence_vectors)
        return vec

    # 求最相似的句子
    # input: test sentence
    def similarity(self, sentence):
        sentence_vec = self.sentence2vec(sentence)
        sims = self.index[sentence_vec]
        sim = max(enumerate(sims), key=lambda item: item[1])

        index = sim[0]
        score = sim[1]
        sentence = self.sentences[index]

        return index,score  # 返回最相似的句子的下标和相似度得分

        # 求最相似前k个句子
    def similarity_k(self, sentence, k):
        sentence_vec = self.sentence2vec(sentence)
        t1 = time.time()
        sims = self.index[sentence_vec]
        t2 = time.time()
        print('特征检索耗时：{:.4f}ms, 检索样本总数：{}'.format(t2-t1, self.sentences_num))
        sim_k = sorted(enumerate(sims), key=lambda item: item[1], reverse=True)[:k]

        indexs = [i[0] for i in sim_k]
        scores = [i[1] for i in sim_k]
        return indexs, scores

Expand use

Training model, input train_data, the training model can be replaced

Similar_model = SentenceSimilarity(train_data,min_frequency = 1)
Similar_model.simple_model()
Similar_model.LsiModel()
#Similar_model.LdaModel()。
#Similar_model.TfidfModel()

Predict the sentence, input sentence, return the subscript index of the most similar sentence in train_data, and the similarity score score

index,score = Similar_model.similarity(sentence)