前言

追求最前沿的技术是每个NLPer的基本操守！

一、相关知识介绍

Jigsaw Rate Severity of Toxic Comments原文.

1-1、比赛描述

# require: 测试集是一些评论，比赛要求按照评论毒性的严重程度对评论进行排名，毒性强的评论应该获得比较高的数值
# train: 训练数据的话，以单模一为例：我们使用之前的jigsaw-toxic-comment-classification-challenge的训练集，
# train: 将训练集处理为有毒评论和无毒评论，将评论处理之后，使用分类器进行有标签的数据训练
# validation：对验证集进行向量化处理、文本清洗，再使用模型来进行概率预测。

二、模型介绍

TF-idf+贝叶斯.

2-1、单模1——（TF-idf+贝叶斯）

import pandas as pd
from sklearn.naive_bayes import MultinomialNB
from sklearn.feature_extraction.text import TfidfVectorizer# 读取挑战赛的训练集
df = pd.read_csv("../input/jigsaw-toxic-comment-classification-challenge/train.csv")
# 新增一列，为数据各列相加的和。
df['y'] = (df[['toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']].sum(axis=1) > 0 ).astype(int)
# 重命名列名
df = df[['comment_text', 'y']].rename(columns={'comment_text': 'text'})# 数据不均衡，进行下采样
min_len = (df['y'] == 1).sum()
df_y0_undersample = df[df['y'] == 0].sample(n=min_len, random_state=201)
df = pd.concat([df[df['y'] == 1], df_y0_undersample])from nltk.stem import SnowballStemmer, WordNetLemmatizer
import re
from nltk.corpus import stopwords# SnowballStemmer：词干提取
stemmer = SnowballStemmer('english')
lemmatizer = WordNetLemmatizer()
# 调取停用词库
all_stopwords = stopwords.words('english')def clean(comment):"""对数据进行处理1、把所有除了英文字母之外的字符都转换为空格2、所有字母都转化为小写3、分割开句子，方便处理每个单词4、遍历句子的每一个单词，去除掉停用词，不是停用词的提取词干5、将处理后的每个单词连在一起。""" comment = re.sub('[^a-zA-Z]', ' ', comment)comment = comment.lower()comment = comment.split()comment = [stemmer.stem(word) for word in comment if not word in set(all_stopwords)]comment = [lemmatizer.lemmatize(word) for word in comment]comment = ' '.join(comment)return comment
# 对每条评论都进行数据清洗
df['text'] = df['text'].apply(clean)# 使用tf-idf来对处理后的文本进行处理
vec = TfidfVectorizer()
X = vec.fit_transform(df['text'])# 使用多项式贝叶斯分类器来训练数据
model = MultinomialNB()
model.fit(X, df['y'])# 读取比赛的验证集
df_val = pd.read_csv("../input/jigsaw-toxic-severity-rating/validation_data.csv")
# 使用训练好的模型来分别训练验证集上的弱毒评论和强度评论
X_less_toxic = vec.transform(df_val['less_toxic'].apply(clean))
X_more_toxic = vec.transform(df_val['more_toxic'].apply(clean))# 使用验证集来评判模型的好坏
# predict_proba返回的是一个n行k列的数组，第i行第j列上的数值是模型预测第i个预测样本的标签为j的概率, 如果是二分类，就是两列，第一个概率值是预测为0的概率，第二个概率值是预测为1的概率。
p1 = model.predict_proba(X_less_toxic)
p2 = model.predict_proba(X_more_toxic)(p1[:, 1] < p2[:, 1]).mean()
# 0.6675634382888269
# 即模型的准确率为0.67# 读取要提交的评论,训练评论。
df_sub = pd.read_csv("../input/jigsaw-toxic-severity-rating/comments_to_score.csv")
X_test = vec.transform(df_sub['text'])
p3 = model.predict_proba(X_test)# 读取评论为毒性的概率
df_sub['score'] = p3[:, 1]# 评论唯一的个数是多少。
# df_sub['score'].nunique()# 提交
df_sub[['comment_id', 'score']].to_csv("submission.csv", index=False)

TF-idf+岭回归.

2-2、单模2——（TF-idf+岭回归）

%%time
import pandas as pd
import numpy as npfrom sklearn.linear_model import Ridge
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import mean_squared_error
from scipy.stats import rankdatadef ridge_cv (vec, X, y, X_test, folds, stratified ):kf = StratifiedKFold(n_splits=FOLDS,shuffle=True,random_state=12)val_scores = []rmse_scores = []X_less_toxics = []X_more_toxics = []preds = []for fold, (train_index,val_index) in enumerate(kf.split(X,stratified)):X_train, y_train = X[train_index], y[train_index]X_val, y_val = X[val_index], y[val_index]model = Ridge()model.fit(X_train, y_train)rmse_score = mean_squared_error ( model.predict (X_val), y_val, squared = False) rmse_scores.append (rmse_score)X_less_toxic = vec.transform(df_val['less_toxic'])X_more_toxic = vec.transform(df_val['more_toxic'])p1 = model.predict(X_less_toxic)p2 = model.predict(X_more_toxic)X_less_toxics.append ( p1 )X_more_toxics.append ( p2 )# Validation Accuracyval_acc = (p1< p2).mean()val_scores.append(val_acc)pred = model.predict (X_test)preds.append (pred)print(f"FOLD:{fold}, rmse_fold:{rmse_score:.5f}, val_acc:{val_acc:.5f}")mean_val_acc = np.mean (val_scores)mean_rmse_score = np.mean (rmse_scores)p1 = np.mean ( np.vstack(X_less_toxics), axis=0 )p2 = np.mean ( np.vstack(X_more_toxics), axis=0 )val_acc = (p1< p2).mean()print(f"OOF: val_acc:{val_acc:.5f}, mean val_acc:{mean_val_acc:.5f}, mean rmse_score:{mean_rmse_score:.5f}")preds = np.mean ( np.vstack(preds), axis=0 )return p1, p2, predstoxic = 1.0
severe_toxic = 2.0
obscene = 1.0
threat = 1.0
insult = 1.0
identity_hate = 2.0def create_train (df):df['y'] = df[["toxic","severe_toxic","obscene","threat","insult","identity_hate"]].max(axis=1)df['y'] = df["y"]+df['severe_toxic']*severe_toxicdf['y'] = df["y"]+df['obscene']*obscenedf['y'] = df["y"]+df['threat']*threatdf['y'] = df["y"]+df['insult']*insultdf['y'] = df["y"]+df['identity_hate']*identity_hatedf = df[['comment_text', 'y', 'toxic', 'severe_toxic', 'obscene', 'threat', 'insult', 'identity_hate']].rename(columns={'comment_text': 'text'})#undersample non toxic comments  on Toxic Comment Classification Challengemin_len = (df['y'] >= 1).sum()df_y0_undersample = df[df['y'] == 0].sample(n=int(min_len*1.5),random_state=201)df = pd.concat([df[df['y'] >= 1], df_y0_undersample])return dfdf_val = pd.read_csv("../input/jigsaw-toxic-severity-rating/validation_data.csv")
df_test = pd.read_csv("../input/jigsaw-toxic-severity-rating/comments_to_score.csv")jc_train_df = pd.read_csv("../input/jigsaw-toxic-comment-classification-challenge/train.csv")
print(f"jc_train_df:{jc_train_df.shape}")jc_train_df = create_train(jc_train_df)df = jc_train_df
print(df['y'].value_counts())FOLDS = 7vec = TfidfVectorizer(analyzer='char_wb', max_df=0.5, min_df=3, ngram_range=(4, 6) )
X = vec.fit_transform(df['text'])
y = df["y"].values
X_test = vec.transform(df_test['text'])stratified = np.around ( y )
jc_p1, jc_p2, jc_preds =  ridge_cv (vec, X, y, X_test, FOLDS, stratified )juc_train_df = pd.read_csv("../input/jigsaw-unintended-bias-in-toxicity-classification/train.csv")
print(f"juc_train_df:{juc_train_df.shape}")
juc_train_df = juc_train_df.query ("toxicity_annotator_count > 5")
print(f"juc_train_df:{juc_train_df.shape}")juc_train_df['y'] = juc_train_df[[ 'severe_toxicity', 'obscene', 'sexual_explicit','identity_attack', 'insult', 'threat']].sum(axis=1)juc_train_df['y'] = juc_train_df.apply(lambda row: row["target"] if row["target"] <= 0.5 else row["y"] , axis=1)
juc_train_df = juc_train_df[['comment_text', 'y']].rename(columns={'comment_text': 'text'})
min_len = (juc_train_df['y'] > 0.5).sum()
df_y0_undersample = juc_train_df[juc_train_df['y'] <= 0.5].sample(n=int(min_len*1.5),random_state=201)
juc_train_df = pd.concat([juc_train_df[juc_train_df['y'] > 0.5], df_y0_undersample])df = juc_train_df
print(df['y'].value_counts())FOLDS = 7vec = TfidfVectorizer(analyzer='char_wb', max_df=0.5, min_df=3, ngram_range=(4, 6) )
X = vec.fit_transform(df['text'])
y = df["y"].values
X_test = vec.transform(df_test['text'])stratified = (np.around ( y, decimals = 1  )*10).astype(int)
juc_p1, juc_p2, juc_preds =  ridge_cv (vec, X, y, X_test, FOLDS, stratified )rud_df = pd.read_csv("../input/ruddit-jigsaw-dataset/Dataset/ruddit_with_text.csv")
print(f"rud_df:{rud_df.shape}")
rud_df['y'] = rud_df['offensiveness_score'].map(lambda x: 0.0 if x <=0 else x)
rud_df = rud_df[['txt', 'y']].rename(columns={'txt': 'text'})
min_len = (rud_df['y'] < 0.5).sum()
print(rud_df['y'].value_counts())FOLDS = 7
df = rud_df
vec = TfidfVectorizer(analyzer='char_wb', max_df=0.5, min_df=3, ngram_range=(4, 6) )
X = vec.fit_transform(df['text'])
y = df["y"].values
X_test = vec.transform(df_test['text'])stratified = (np.around ( y, decimals = 1  )*10).astype(int)
rud_p1, rud_p2, rud_preds =  ridge_cv (vec, X, y, X_test, FOLDS, stratified )jc_max = max(jc_p1.max() , jc_p2.max())
juc_max = max(juc_p1.max() , juc_p2.max())
rud_max = max(rud_p1.max() , rud_p2.max())p1 = jc_p1/jc_max + juc_p1/juc_max + rud_p1/rud_max
p2 = jc_p2/jc_max + juc_p2/juc_max + rud_p2/rud_maxval_acc = (p1< p2).mean()
print(f"Ensemble: val_acc:{val_acc:.5f}")preds2 = jc_preds/jc_max + juc_preds/juc_max + rud_preds/rud_max

参考文章：
一文了解倒排表.

总结

kaggle比赛——Jigsaw Rate Severity of Toxic Comments（NLP类型）——分析获奖模型笔记相关推荐

在参加了39场Kaggle比赛之后，有人总结了一份图像分割炼丹的「奇技淫巧」
点击上方"视学算法",选择加"星标"或"置顶" 重磅干货,第一时间送达本文转载自:机器之心一个经历了 39 场 Kaggle 比赛的团队 ...
Kaggle比赛分类与winner资料汇总（更新中）
Kaggle比赛分类与winner资料汇总(更新中) 1.介绍把比赛分为四类,Data Mining.Images.NLP.Speech Recognition,举几个例子: Data Mining ...
kaggle 比赛分类_黑色素瘤分类在kaggle比赛中获得奖牌
kaggle 比赛分类 Using deep learning to identify melanomas from skin images and patient meta-data 使用深度学习从 ...
竞赛老陪跑怎么办？来自一位Kaggle比赛失败者的含泪总结
大数据文摘出品来源:medium 编译:zeroInfinity.笪洁琼 Kaggle比赛应该是数据竞赛中公认含金量最高的那个.每场比赛,参加的队伍至少上千人,也并非每次都次都能脱引而出,一不小心就 ...
通俗理解kaggle比赛大杀器xgboost + XGBOOST手算内容转
通俗理解kaggle比赛大杀器xgboost 转 https://blog.csdn.net/v_JULY_v/article/details/81410574 XGBOOST有手算内容 htt ...
Kaggle比赛冠军经验分享：如何用 RNN 预测维基百科网络流量
Kaggle比赛冠军经验分享:如何用 RNN 预测维基百科网络流量 from:https://www.leiphone.com/news/201712/zbX22Ye5wD6CiwCJ.html 导语 ...
Kaggle常用函数总结原创 2017年07月03日 21:47:34 标签： kaggle 493 kaggle比赛也参加了好几次，在这里就把自己在做比赛中用到的函数汇总到这，方便自己以后查阅
Kaggle常用函数总结原创 2017年07月03日 21:47:34 标签: kaggle / 493 编辑删除 kaggle比赛也参加了好几次,在这里就把自己在做比赛中用到的函数汇总到这,方便 ...
kaggle比赛模型融合指南
kaggle比赛模型融合指南转载 2017年10月13日 16:29:32
EL之Bagging：kaggle比赛之利用泰坦尼克号数据集建立Bagging模型对每个人进行获救是否预测
EL之Bagging:kaggle比赛之利用泰坦尼克号数据集建立Bagging模型对每个人进行获救是否预测目录输出结果设计思路核心代码输出结果设计思路核心代码 bagging_clf = ...

kaggle比赛——Jigsaw Rate Severity of Toxic Comments（NLP类型）——分析获奖模型笔记

目录

前言

一、相关知识介绍

1-1、比赛描述

二、模型介绍

2-1、单模1——（TF-idf+贝叶斯）

2-2、单模2——（TF-idf+岭回归）

总结

kaggle比赛——Jigsaw Rate Severity of Toxic Comments（NLP类型）——分析获奖模型笔记相关推荐

最新文章

热门文章