具有 python 和 keras 的 2 个以上类的文本分类

Question

I'm trying categorization with a neuronal network.

my data looks like this:

Sentence	category
sentence 1	0
sentence 2	1
sentence 3	2
sentence 4	3

therefore I have 4 different categories. I seperate the sentences and the labels in 2 different lists and put the first list (with the sencenes) in a tokenizer.

Edit: Because someone asked about data samples. Here is a link to the Dropbox (data_samples.txt ) and some more explanations. 0 = not needed, 1 = title, 2= name from author, 3 = pages. The sentence is with | from the label separated. There are lot of special character (like -- >,") because of the lemmatization but the tokenizer skips these characters. The content is from different websites and I only use it in the context of study (no for money or in bad intentions)

y_train = array(labels) # it has to be numpy array otherwise it will cause errors, labels is a list of integers
x_train = tokenizer.texts_to_sequences(newLines) # newLines is a list with all sentences
x_train = pad_sequences(x_train, maxlen=sequenceLength)
vocabSize = len(tokenizer.word_index) + 1

An now i will train my model

mymodel = Sequential()
mymodel.add(Embedding(input_dim=vocabSize, output_dim=100, input_length=sequenceLength))
mymodel.add(Conv1D(32, 3, padding='same', activation='relu'))
mymodel.add(MaxPooling1D())
mymodel.add(Flatten())
mymodel.add(Dense(250, activation='relu'))
mymodel.add(Dense(4, activation='softmax'))
mymodel.compile(optimizer="adam", loss="categorical_crossentropy", metrics=["accuracy"])
history = mymodel.fit(x_train, y_train, epochs=30, batch_size=8)

But if i start the programm i get the following error:

• ValueError: Shapes (None, 1) and (None, 4) are incompatible

I know it is because of the last Dense Layer. If I write 1 instead of 4, I don't get an error. But I thought the last dense layer need a 4 as a parameter, because I have 4 categories. I think the last dense layer with a 1 instead of a 4 is not correct, because loss is 0.0000e+00. And that look not correct xD

What did I do wrong?

Answer 1

replace loss="categorical_crossentropy" with "sparse_ categorical_crossentropy".

The model produces a value per class; the sparse_ loss knows how to compare a single value in range [0-3] to these 4 score values.