
import pandas as pd


review = pd.read_excel('data/data_all.xlsx')


review.head()


with open('review.txt', 'w', encoding='utf-8') as f:
    f.write('\n'.join(review['review_spell_check']))


from sentencepiece import SentencePieceTrainer
sp = SentencePieceTrainer.Train('--input=review.txt --model_prefix=review --vocab_size=3000')


from sentencepiece import SentencePieceProcessor
sp = SentencePieceProcessor()
sp.Load("review.model")

True


sp.encode_as_pieces(review.loc[123, 'review'])

['▁여러',
 '지',
 '점',
 '▁가',
 '봤',
 '는데',
 '▁느낌',
 '은',
 '▁다',
 '▁비슷',
 '하고',
 '▁광화문점',
 '이',
 '▁서울',
 '여행',
 '하기',
 '에',
 '▁위치',
 '는',
 '▁좋은',
 '것',
 '같',
 '아요',
 '▁대신',
 '▁다',
 '른',
 '지',
 '점',
 '들',
 '보다',
 '▁조금',
 '은',
 '▁비싼',
 '▁',
 '감이',
 '▁있',
 '네',
 '요']


from sklearn.feature_extraction.text import CountVectorizer


cv = CountVectorizer(lowercase=False, tokenizer=sp.encode_as_pieces) #tokenizer를 안해주면, 띄어쓰기 단위로 잘라줌


tdm = cv.fit_transform(review['review_spell_check'])


from sklearn.model_selection import train_test_split


x = tdm
y = review['helpful']


x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=42)


import joblib


cv.tokenizer = None # joblib로 저장하기 위해 필요


joblib.dump((cv, x_train, x_test, y_train, y_test), 'hotel_subword.pkl')

['hotel_subword.pkl']


cv, x_train, x_test, y_train, y_test = joblib.load('hotel_subword.pkl')


x_train.shape

(1974, 2996)


import tensorflow as tf


model = tf.keras.models.Sequential()


model.add(tf.keras.layers.Dense(
    1,  #1,0 중 하나가 나온다! y자체는 열이 하나 ->Dense(1, )
    input_shape = (2996, ), #input data는 3018개 : 튶플로 넣어주기
    activation = 'sigmoid',
    kernel_regularizer = tf.keras.regularizers.l1_l2(0, 0.001) #ㅣ1은 0만, l2는 0.001만
))


model.summary()

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense (Dense)                (None, 1)                 2997      
=================================================================
Total params: 2,997
Trainable params: 2,997
Non-trainable params: 0
_________________________________________________________________


model.compile(optimizer = 'adam', loss='binary_crossentropy', metrics = ['accuracy'])


model.fit(x_train.toarray(), y_train.values, epochs=100, validation_split=0.1,
          callbacks=[tf.keras.callbacks.EarlyStopping()])

Epoch 1/100
56/56 [==============================] - 2s 13ms/step - loss: 0.6719 - accuracy: 0.5976 - val_loss: 0.6259 - val_accuracy: 0.6919
Epoch 2/100
56/56 [==============================] - 0s 7ms/step - loss: 0.6040 - accuracy: 0.7105 - val_loss: 0.6012 - val_accuracy: 0.7273
Epoch 3/100
56/56 [==============================] - 0s 4ms/step - loss: 0.5805 - accuracy: 0.8178 - val_loss: 0.5837 - val_accuracy: 0.7879
Epoch 4/100
56/56 [==============================] - 0s 3ms/step - loss: 0.5654 - accuracy: 0.8499 - val_loss: 0.5709 - val_accuracy: 0.8131
Epoch 5/100
56/56 [==============================] - 0s 3ms/step - loss: 0.5350 - accuracy: 0.8580 - val_loss: 0.5592 - val_accuracy: 0.8232
Epoch 6/100
56/56 [==============================] - 0s 4ms/step - loss: 0.5144 - accuracy: 0.8745 - val_loss: 0.5496 - val_accuracy: 0.8232
Epoch 7/100
56/56 [==============================] - 0s 4ms/step - loss: 0.4930 - accuracy: 0.8873 - val_loss: 0.5415 - val_accuracy: 0.8283
Epoch 8/100
56/56 [==============================] - 0s 4ms/step - loss: 0.4804 - accuracy: 0.8946 - val_loss: 0.5345 - val_accuracy: 0.8131
Epoch 9/100
56/56 [==============================] - 0s 4ms/step - loss: 0.4673 - accuracy: 0.8821 - val_loss: 0.5283 - val_accuracy: 0.8283
Epoch 10/100
56/56 [==============================] - 0s 3ms/step - loss: 0.4711 - accuracy: 0.8850 - val_loss: 0.5226 - val_accuracy: 0.8232
Epoch 11/100
56/56 [==============================] - 0s 4ms/step - loss: 0.4448 - accuracy: 0.8943 - val_loss: 0.5181 - val_accuracy: 0.8182
Epoch 12/100
56/56 [==============================] - 0s 4ms/step - loss: 0.4425 - accuracy: 0.8806 - val_loss: 0.5141 - val_accuracy: 0.8283
Epoch 13/100
56/56 [==============================] - 0s 4ms/step - loss: 0.4353 - accuracy: 0.8842 - val_loss: 0.5096 - val_accuracy: 0.8283
Epoch 14/100
56/56 [==============================] - 0s 4ms/step - loss: 0.4235 - accuracy: 0.8859 - val_loss: 0.5073 - val_accuracy: 0.8333
Epoch 15/100
56/56 [==============================] - 0s 4ms/step - loss: 0.4072 - accuracy: 0.8934 - val_loss: 0.5039 - val_accuracy: 0.8333
Epoch 16/100
56/56 [==============================] - 0s 4ms/step - loss: 0.4106 - accuracy: 0.8962 - val_loss: 0.5014 - val_accuracy: 0.8283
Epoch 17/100
56/56 [==============================] - 0s 3ms/step - loss: 0.3998 - accuracy: 0.9025 - val_loss: 0.4988 - val_accuracy: 0.8283
Epoch 18/100
56/56 [==============================] - 0s 4ms/step - loss: 0.4008 - accuracy: 0.8995 - val_loss: 0.4971 - val_accuracy: 0.8283
Epoch 19/100
56/56 [==============================] - 0s 4ms/step - loss: 0.3836 - accuracy: 0.8987 - val_loss: 0.4957 - val_accuracy: 0.8232
Epoch 20/100
56/56 [==============================] - 0s 3ms/step - loss: 0.3869 - accuracy: 0.9041 - val_loss: 0.4943 - val_accuracy: 0.8182
Epoch 21/100
56/56 [==============================] - 0s 7ms/step - loss: 0.3801 - accuracy: 0.9082 - val_loss: 0.4929 - val_accuracy: 0.8232
Epoch 22/100
56/56 [==============================] - 0s 5ms/step - loss: 0.3793 - accuracy: 0.9113 - val_loss: 0.4916 - val_accuracy: 0.8182
Epoch 23/100
56/56 [==============================] - 0s 3ms/step - loss: 0.3823 - accuracy: 0.8915 - val_loss: 0.4908 - val_accuracy: 0.8182
Epoch 24/100
56/56 [==============================] - 0s 4ms/step - loss: 0.3744 - accuracy: 0.9134 - val_loss: 0.4895 - val_accuracy: 0.8131
Epoch 25/100
56/56 [==============================] - 0s 3ms/step - loss: 0.3625 - accuracy: 0.9146 - val_loss: 0.4889 - val_accuracy: 0.8131
Epoch 26/100
56/56 [==============================] - 0s 3ms/step - loss: 0.3693 - accuracy: 0.8981 - val_loss: 0.4874 - val_accuracy: 0.8182
Epoch 27/100
56/56 [==============================] - 0s 3ms/step - loss: 0.3575 - accuracy: 0.9222 - val_loss: 0.4868 - val_accuracy: 0.8182
Epoch 28/100
56/56 [==============================] - 0s 7ms/step - loss: 0.3583 - accuracy: 0.9063 - val_loss: 0.4863 - val_accuracy: 0.8232
Epoch 29/100
56/56 [==============================] - 0s 6ms/step - loss: 0.3696 - accuracy: 0.9027 - val_loss: 0.4853 - val_accuracy: 0.8232
Epoch 30/100
56/56 [==============================] - 0s 5ms/step - loss: 0.3601 - accuracy: 0.9044 - val_loss: 0.4852 - val_accuracy: 0.8232
Epoch 31/100
56/56 [==============================] - 0s 5ms/step - loss: 0.3545 - accuracy: 0.9117 - val_loss: 0.4852 - val_accuracy: 0.8182
Epoch 32/100
56/56 [==============================] - 0s 4ms/step - loss: 0.3641 - accuracy: 0.8983 - val_loss: 0.4848 - val_accuracy: 0.8131
Epoch 33/100
56/56 [==============================] - 0s 4ms/step - loss: 0.3624 - accuracy: 0.9114 - val_loss: 0.4841 - val_accuracy: 0.8131
Epoch 34/100
56/56 [==============================] - 0s 4ms/step - loss: 0.3561 - accuracy: 0.9129 - val_loss: 0.4842 - val_accuracy: 0.8131

<tensorflow.python.keras.callbacks.History at 0x244feef0860>


model.evaluate(x_test.toarray(), y_test.values, verbose=0) #손실, accuracy

[0.5522161722183228, 0.8036437034606934]


weights, _ = model.trainable_weights


token_weight = pd.DataFrame({'토큰': cv.get_feature_names(), '가중치': weights.numpy().flat})


token_weight.sort_values('가중치').head()


token_weight.sort_values('가중치').tail()

	site	hotel	score	review	date	star	length	review_spell_check	helpful	attitude
0	야놀자	신라스테이 광화문	10.0	깔끔합니다	2019. 03. 30	5.0	5	깔끔합니다	0	2
1	야놀자	신라스테이 광화문	6.0	다니기는 참 좋은데 청소를 아예 안 하시네요	2020. 09. 13	3.0	24	다니기는 참 좋은데 청소를 아예 안 하시네요	1	0
2	야놀자	나인트리 프리미어 명동2	10.0	깔끔한 내부가 좋았습니다	2019. 11. 25	5.0	13	깔끔한 내부가 좋았습니다	0	2
3	아고다	나인트리 프리미어 호텔 명동 2	10.0	가격이 저렴해서 어딘가 아쉬운점은 한두개씩 있겠거니 하고 큰 기대없이 갔는데 생각외...	NaN	NaN	285	가격이 저렴해서 어딘가 아쉬운 점은 한두 개씩 있겠거니 하고 큰 기대 없이 갔는데 ...	1	2
4	아고다	나인트리 프리미어 호텔 명동 2	7.6	우선 뷰가 공사현장 보이고 주유소 보여서 그랬구요 새벽 6시쯤 넘어서는 공사장 소...	NaN	NaN	93	우선 뷰가 공사현장 보이고 주유소 보여서 그랬고요 새벽 6시쯤 넘어서는 공사장 소...	1	0

	토큰	가중치
1434	▁좋아요	-0.741736
337	▁깔끔	-0.447911
295	▁굿	-0.410221
1605	▁친절합니다	-0.371433
274	▁괜찮아	-0.368910

	토큰	가중치
1689	▁편의점	0.457330
738	▁방음	0.504998
193	▁객실	0.540762
398	▁넓	0.561374
851	▁빼고	0.578300

Sentencepiece 적용¶

훈련, 테스트 데이터 나누기¶

모형¶

계수 확인¶