载入数据集

首先我们载入kaggle的数据集
PhraseId为一个id，SentenceId为句子ID ，这里数据集，将一个句子分解开了，第一个为句子，后面的均是他的分解。

# 解压数据集
import zipfile
def extract_file(path):
    with zipfile.ZipFile(path,'r') as zipref:
        zipref.extractall('./data')

paths = [
    '/kaggle/input/sentiment-analysis-on-movie-reviews/train.tsv.zip',
    '/kaggle/input/sentiment-analysis-on-movie-reviews/test.tsv.zip', 
]
_ = [extract_file(path) for path in paths]

然后我们只需要读入句子和情感标签即可

train_tsv =pd.read_csv('./data/train.tsv', sep='\t', header=0)
# train_tsv = pd.read_csv("/kaggle/working/data/train.tsv")
phrase = np.array(train_tsv["Phrase"])
sentiment =  np.array(train_tsv["Sentiment"])

BOW特征工程

这里使用sklearn的CountVectorizer模型，将模型转换成一个np的矩阵。

from sklearn.feature_extraction.text import CountVectorizer

vectorizer  = CountVectorizer(stop_words='english', max_features=1000)
docs = phrase
# bag是一个稀疏的矩阵。因为词袋模型就是一种稀疏的表示。
bag = vectorizer .fit_transform(docs)
vocabulary = vectorizer.get_feature_names_out()
word_counts = bag.toarray()

划分数据集

这里将数据划分80%的训练集，20%测试集

from sklearn.model_selection import train_test_split
x_train, x_test, y_train, y_test = train_test_split(word_counts,sentiment, random_state=22,test_size=0.2)

训练

导入sklarn的LogisticRegression模型

# 损失函数： liblinear  lbfgs  newton-cg  sag
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score
classifier = LogisticRegression(solver="lbfgs",max_iter=500 )
classifier.fit(x_train, y_train)

y_pred = classifier.predict(x_test)
print(accuracy_score(y_pred,y_test))

结果精度只有0.58

提交

def submission(model,vectorizer):
    test_csv =pd.read_csv('./data/test.tsv', sep='\t', header=0)
    # phrase = np.array(test_csv["Phrase"])
    
    X = vectorizer.fit_transform(test_csv["Phrase"].values.astype('U'))

    y_pred = model.predict(X)
    
    test_csv['Sentiment'] = pd.Series(y_pred)
    submission = pd.concat([test_csv["Phrase"], test_csv['Sentiment']], axis=1)
    submission.to_csv("submission.csv", index=False)
    print("Done!!!!!!!!!!!!!!!!!!!!!!!!!!!")
    
    
submission(classifier,vectorizer)

只有0.488,太低了，接下来尝试深度学习词嵌入的方法。