千言数据集：文本相似度——提取TFIDF以及统计特征，训练和预测

以下学习笔记来源于 Coggle 30 Days of ML（22年1&2月）
链接：https://coggle.club/blog/30days-of-ml-202201

比赛链接：https://aistudio.baidu.com/aistudio/competition/detail/45/0/task-definition

提取TFIDF以及统计特征，训练和预测

导入所需库

import numpy as np
import pandas as pd
import jieba
import Levenshtein #计算编辑距离
from tqdm import tqdm
import warnings
import lightgbm as lgb
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import roc_auc_score, f1_score, precision_score, recall_score,accuracy_score
import glob

读取文件并做相应处理

#读取tsv文件的方法
def read_tsv(input_file,columns):
    with open(input_file,"r",encoding="utf-8") as file:
        lines = []
        count = 1
        for line in file:
            if len(line.strip().split("\t")) != 1:
                lines.append([count]+line.strip().split("\t"))
                count += 1
        df = pd.DataFrame(lines)
        df.columns = columns
    return df

• 读取训练数据

bq_train=read_tsv('./bq_corpus/train.tsv',['index','text1','text2','label'])
lcqmc_train=read_tsv('./lcqmc/train.tsv',['index','text1','text2','label'])
pawsx_train=read_tsv('./paws-x-zh/train.tsv',['index','text1','text2','label'])

• 读取预测数据

bq_test=read_tsv('./bq_corpus/test.tsv',['index','text1','text2'])
lcqmc_test=read_tsv('./lcqmc/test.tsv',['index','text1','text2'])
pawsx_test=read_tsv('./paws-x-zh/test.tsv',['index','text1','text2'])

• 一些处理文本的函数

#文本处理，有些可能用不到
import re
import string
import jieba
with open("dict/stop_words.utf8",encoding="utf-8") as f:
    stopword_list=f.readlines()

def tokenize_text(text):
    tokens=jieba.cut(text)
    tokens=[token.strip() for token in tokens]
    return tokens

def remove_special_characters(text):
    tokens=tokenize_text(text)
    pattern=re.compile('[{}]'.format(re.escape(string.punctuation)))
    filtered_tokens=filter(None,[pattern.sub('',token) for token in tokens])
    filtered_text=''.join(filtered_tokens)
    return filtered_text

#去除停用词
def remove_stopwords(text):
    tokens=tokenize_text(text)
    filtered_tokens=[token for token in tokens if token not in stopword_list]
    filtered_text=''.join(filtered_tokens)
    return filtered_text

def normalize_corpus(corpus,tokenize=False):
    normalize_corpus=[]
    for text in corpus:
        text=remove_special_characters(text)
        text=remove_stopwords(text)
        if tokenize:
            normalize_corpus.append(tokenize_text(text))
        else:
            normalize_corpus.append(text)
    return normalize_corpus

• 将训练集文本作为语料库

df_train=pd.concat([bq_train,lcqmc_train,pawsx_train])
df_test=pd.concat([bq_test,lcqmc_test,pawsx_test])
corpus=df_train['text1'].values.tolist()+df_train['text2'].values.tolist()
tokenized_corpus=[' '.join(jieba.lcut(text)) for text in corpus]

统计特征

构建统计特征，包括：
句子A包含的字符个数、句子B包含的字符个数
句子A与句子B的编辑距离
句子A与句子B共有单词的个数
句子A与句子B共有字符的个数
句子A与句子B共有单词的个数 / 句子A字符个数
句子A与句子B共有单词的个数 / 句子B字符个数

#计算编辑距离，可以直接使用Levenshtein的distance计算
def edit_distance(text1,text2):
    n,m=len(text1),len(text2)
    import numpy as np
    dp=np.zeros((n+1,m+1))
    for i in range(1,n+1):
        dp[i][0]=i
    for j in range(1,m+1):
        dp[0][j]=j  
    for i in range(1,n+1):
        for j in range(1,m+1):
            tmp=int(text1[i-1]!=text2[j-1])
            dp[i][j]=min(dp[i-1][j]+1,dp[i][j-1]+1,dp[i-1][j-1]+tmp)
    return dp[n][m]

# 两个列表共有变量的个数
def both_num(list1,list2):
    dict1,dict2,ans={},{},0
    for i in list1:
        dict1[i]=list1.count(i)
    for i in list2:
        dict2[i]=list2.count(i)
    for k,v in dict1.items():
        tmp=0 if k not in list2 else dict2[k]
        ans+=min(v,tmp)
    return ans

#text1和text2长度差
def len_diff(text1,text2):
    return abs(len(text1)-len(text2))

# text1和text2共有单词的个数
def both_words_num(text1,text2):
    a,b=jieba.lcut(text1),jieba.lcut(text2)
    return both_num(a,b)

# text1和text2共有字符的个数
def both_chars_num(text1,text2):
    a,b=[i for i in text1],[i for i in text2]
    return both_num(a,b)

#text1与text2共有单词的个数 / text1字符个数
def both_words_divideby_char1(text1,text2):
    return both_words_num(text1,text2)/len(text1)

#text1与text2共有单词的个数 / text2字符个数
def both_words_divideby_char2(text1,text2):
    return both_words_num(text1,text2)/len(text2)

提取TFIDF特征

计算TFIDF，并对句子A和句子B进行特征转换，计算句子A与句子B的TFIDF向量的内积距离

from sklearn.feature_extraction.text import TfidfVectorizer

def tfidf_extractor(corpus,ngram_range=(1,1)):
    vectorizer=TfidfVectorizer(min_df=1,
                              norm='l2',
                              smooth_idf=True,
                              use_idf=True,
                              ngram_range=ngram_range)
    features=vectorizer.fit_transform(corpus)
    return vectorizer,features


def participle_text(text):
    words_list=jieba.lcut(text)
    return ' '.join(words_list)


#得到tfidf特征向量的内积距离
def get_tfidfvec_dis(tfidf_vectorizer,text1,text2):
    fit_text1=tfidf_vectorizer.transform([participle_text(text1)])
    fit_text2=tfidf_vectorizer.transform([participle_text(text2)])
    vec1=fit_text1.toarray()[0]
    vec2=fit_text2.toarray()[0]
    return np.dot(vec1,vec2)

def tfidfvec_dis_list(tfidf_vectorizer,df):
    dis_list=[]
    for text1,text2 in zip(df['text1'],df['text2']):
        dis_list.append(get_tfidfvec_dis(tfidf_vectorizer,text1,text2))
    return dis_list

tfidf_vectorizer,tfidf_train_features=tfidf_extractor(tokenized_corpus)

示例：

所有特征

def feature_eng(data):
    """
    统计特征
    """
    data['len_diff']=[len_diff(t1,t2) for t1,t2 in zip(data['text1'],data['text2'])]
    data['both_words']=[both_words_num(t1,t2) for t1,t2 in zip(data['text1'],data['text2'])]
    data['both_chars']=[both_chars_num(t1,t2) for t1,t2 in zip(data['text1'],data['text2'])]
    data['both_words_div1']=[both_words_divideby_char1(t1,t2) for t1,t2 in zip(data['text1'],data['text2'])]
    data['both_words_div2']=[both_words_divideby_char2(t1,t2) for t1,t2 in zip(data['text1'],data['text2'])]
    data['tf_idf_dis']=tfidfvec_dis_list(tfidf_vectorizer,data)
    data['edit_dis']=[Levenshtein.distance(t1,t2) for t1,t2 in zip(data['text1'],data['text2'])]
    
    return data

df_train=feature_eng(df_train)
df_test=feature_eng(df_test)

处理完之后的样子：

训练并预测

eval_fun = accuracy_score
def run_oof(clf, X_train, y_train, X_test, kf):
    print(clf)
    preds_train = np.zeros((len(X_train), 2), dtype = np.float64)
    preds_test = np.zeros((len(X_test), 2), dtype = np.float64)
    train_loss = []; test_loss = []

    i = 1
    for train_index, test_index in kf.split(X_train, y_train):
        x_tr = X_train[train_index]; x_te = X_train[test_index]
        y_tr = y_train[train_index]; y_te = y_train[test_index]
        clf.fit(x_tr, y_tr, eval_set = [(x_te, y_te)], verbose = False)
        train_loss.append(eval_fun(y_tr, np.argmax(clf.predict_proba(x_tr)[:], 1)))
        test_loss.append(eval_fun(y_te, np.argmax(clf.predict_proba(x_te)[:], 1)))

        preds_train[test_index] = clf.predict_proba(x_te)[:]
        preds_test += clf.predict_proba(X_test)[:]

        print('{0}: Train {1:0.7f} Val {2:0.7f}/{3:0.7f}'.format(i, train_loss[-1], test_loss[-1], np.mean(test_loss)))
        print('-' * 50)
        i += 1
    print('Train: ', train_loss)
    print('Val: ', test_loss)
    print('-' * 50)
    print('Train{0:0.5f}_Test{1:0.5f}\n\n'.format(np.mean(train_loss), np.mean(test_loss)))
    preds_test /= n_fold
    return preds_train, preds_test

params = {
    'objective':'binary',
    'boosting_type':'gbdt',
    'metric':'auc',
    'n_jobs':-1,
    'learning_rate':0.05,
    'num_leaves': 2**6,
    'max_depth':8,
    'tree_learner':'serial',
    'colsample_bytree': 0.8,
    'subsample_freq':1,
    'subsample':0.8,
    'num_boost_round':5000,
    'max_bin':255,
    'verbose':-1,
    'seed': 2021,
    'bagging_seed': 2021,
    'feature_fraction_seed': 2021,
    'early_stopping_rounds':100,
}
n_fold=10
skf = StratifiedKFold(n_splits = n_fold, shuffle = True)
train_pred, test_pred = run_oof(lgb.LGBMClassifier(**params),
                                df_train.iloc[:,1:].values,
                                df_train.iloc[:,0].values,
                                df_test.iloc[:,:].values,
                                skf)

训练过程：

将结果保存到文件，并提交

df_test['prediction']=np.argmax(test_pred, 1)
bq_test['prediction']=df_test['prediction'][:len(bq_test)].tolist()
lcqmc_test['prediction']=df_test['prediction'][len(bq_test):len(bq_test)+len(lcqmc_test)].tolist()
pawsx_test['prediction']=df_test['prediction'][len(bq_test)+len(lcqmc_test):].tolist()

bq_test['index']=[i for i in range(len(bq_test))]
lcqmc_test['index']=[i for i in range(len(lcqmc_test))]
pawsx_test['index']=[i for i in range(len(pawsx_test))]
bq_test.to_csv('submit/bq_corpus.tsv', sep='\t',
                  index=False, columns=["index","prediction"], mode="w")
lcqmc_test.to_csv('submit/lcqmc.tsv', sep='\t', 
                  index=False, columns=["index","prediction"], mode="w")
pawsx_test.to_csv('submit/paws-x.tsv', sep='\t', 
                  index=False, columns=["index","prediction"], mode="w")

0.6495的分数，再接再厉。