word2vec+LSTM 情感识别，帮忙看看哪出了问题

[KarlRixon]

数据来源是京东新款手机的评论和打分，目前收集到 2500 条数据，但打分小于 5 分的只有不到 40 条

训练模型层次是：嵌入层-》 LSTM-》 Dense-》 Dense-》输出层 嵌入层的初始数据为 word2vec 训练的词向量 输入的训练数据为词索引，标记为打分

部分代码如下：

def main():
    x_train = pad_seq()
    y_train = star()
    x_train, y_train, x_test, y_test = set_data(x_train, y_train)
    model = Sequential()
    model.add(Embedding(input_dim=input_dim+1, output_dim=output_dim, input_length=k, embeddings_initializer=my_init))
    model.add(LSTM(128, dropout=0.2, recurrent_dropout=0.2, activation='sigmoid'))
    model.add(Dense(256, activation='relu'))
    model.add(Dense(128, activation='relu'))
    model.add(Dense(1,activation='sigmoid'))
    model.compile(loss="binary_crossentropy", optimizer="adam", metrics=["accuracy"])
    model.fit(x_train, y_train, batch_size=batch_size, epochs=15)
    score, acc = model.evaluate(x_test, y_test, batch_size=batch_size)
    print('Test score:', score)
    print('Test accuracy:', acc)

然而训练结果是这样的。。。

Epoch 15/15

  32/2015 [..............................] - ETA: 0s - loss: -63.2713 - acc: 0.0000e+00
 160/2015 [=>............................] - ETA: 0s - loss: -63.4706 - acc: 0.0000e+00
 288/2015 [===>..........................] - ETA: 0s - loss: -63.5481 - acc: 0.0000e+00
 384/2015 [====>.........................] - ETA: 0s - loss: -63.2713 - acc: 0.0026    
 480/2015 [======>.......................] - ETA: 0s - loss: -63.3046 - acc: 0.0021
 608/2015 [========>.....................] - ETA: 0s - loss: -63.4024 - acc: 0.0016
 736/2015 [=========>....................] - ETA: 0s - loss: -63.3796 - acc: 0.0027
 864/2015 [===========>..................] - ETA: 0s - loss: -63.3821 - acc: 0.0023
 992/2015 [=============>................] - ETA: 0s - loss: -63.3838 - acc: 0.0020
1120/2015 [===============>..............] - ETA: 0s - loss: -63.3852 - acc: 0.0018
1248/2015 [=================>............] - ETA: 0s - loss: -63.3991 - acc: 0.0016
1376/2015 [===================>..........] - ETA: 0s - loss: -63.4104 - acc: 0.0015
1504/2015 [=====================>........] - ETA: 0s - loss: -63.3879 - acc: 0.0020
1632/2015 [=======================>......] - ETA: 0s - loss: -63.4081 - acc: 0.0018
1760/2015 [=========================>....] - ETA: 0s - loss: -63.4344 - acc: 0.0017
1888/2015 [===========================>..] - ETA: 0s - loss: -63.4233 - acc: 0.0021
2015/2015 [==============================] - 1s 470us/step - loss: -63.4214 - acc: 0.0020

 32/504 [>.............................] - ETA: 2s
504/504 [==============================] - 0s 412us/step
Test score: -63.769539061046785
Test accuracy: 0.0

Process finished with exit code 0

model 如下：

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
embedding_1 (Embedding)      (None, 20, 50)            49350     
_________________________________________________________________
lstm_1 (LSTM)                (None, 128)               91648     
_________________________________________________________________
dense_1 (Dense)              (None, 256)               33024     
_________________________________________________________________
dense_2 (Dense)              (None, 128)               32896     
_________________________________________________________________
dense_3 (Dense)              (None, 1)                 129       
=================================================================
Total params: 207,047
Trainable params: 207,047
Non-trainable params: 0
_________________________________________________________________

[wz74666291]

你试着把 fit 的 step 减小一些试试，比如 10e-5

[hanbing135]

完全看不懂

[aREMbosAl]

label 是打分？但是用的 loss 是二分类的

[ayase252]

loss 为负？这违反代价函数定义了吧。
仔细考虑一下输出是什么

[douglas1997]

我就想知道 Epoch 15 的时候后面的 Acc 指标为什么一直是 0.001~0.002.... 压根就没有优化好。

[zzj0311]

loss 是负的是什么鬼。。这拿啥写的，sklearn ？

[zzj0311]

所以你这个输入是评分输出是一个二分类？那不可能对的嘛

[KarlRixon]

@zzj0311 用的 Keras

[KarlRixon]

@zzj0311 我想起了，1-3 分是负面评价，4-5 分是正面评价

[KarlRixon]

@ayase252 1-3 分是负面评价，4-5 分是正面评价，可能是我少了这步处理，直接把打分喂入了