如何在 keras 中将预测序列转换回文本

Question

我有一个序列到序列学习模型，它工作正常，能够预测一些输出。问题是我不知道如何将输出转换回文本序列。

这是我的代码。

from keras.preprocessing.text import Tokenizer,base_filter
from keras.preprocessing.sequence import pad_sequences
from keras.models import Sequential
from keras.layers import Dense
txt1="""What makes this problem difficult is that the sequences can vary in length,
be comprised of a very large vocabulary of input symbols and may require the model 
to learn the long term context or dependencies between symbols in the input sequence."""
#txt1 is used for fitting 
tk = Tokenizer(nb_words=2000, filters=base_filter(), lower=True, split=" ")
tk.fit_on_texts(txt1)
#convert text to sequence
t= tk.texts_to_sequences(txt1)
#padding to feed the sequence to keras model
t=pad_sequences(t, maxlen=10)
model = Sequential()
model.add(Dense(10,input_dim=10))
model.add(Dense(10,activation='softmax'))
model.compile(loss='categorical_crossentropy', optimizer='adam',metrics=['accuracy'])
#predicting new sequcenc
pred=model.predict(t)
#Convert predicted sequence to text
pred=??

Answer 1

您可以直接使用逆tokenizer.sequences_to_texts函数。

    text = tokenizer.sequences_to_texts(<list_of_integer_equivalent_encodings>)

我已经测试了上述内容，它按预期工作。

PS.：要格外小心，使参数成为整数编码的列表，而不是One Hot的列表。

Answer 2

这是我找到的解决方案：

reverse_word_map = dict(map(reversed, tokenizer.word_index.items()))

Answer 3

我必须解决同样的问题，所以这就是我最终的做法（灵感来自@Ben Usemans 反向字典）。

# Importing library
from keras.preprocessing.text import Tokenizer
# My texts
texts = ['These are two crazy sentences', 'that I want to convert back and forth']
# Creating a tokenizer
tokenizer = Tokenizer(lower=True)
# Building word indices
tokenizer.fit_on_texts(texts)
# Tokenizing sentences
sentences = tokenizer.texts_to_sequences(texts)
>sentences
>[[1, 2, 3, 4, 5], [6, 7, 8, 9, 10, 11, 12, 13]]
# Creating a reverse dictionary
reverse_word_map = dict(map(reversed, tokenizer.word_index.items()))
# Function takes a tokenized sentence and returns the words
def sequence_to_text(list_of_indices):
    # Looking up words in dictionary
    words = [reverse_word_map.get(letter) for letter in list_of_indices]
    return(words)
# Creating texts 
my_texts = list(map(sequence_to_text, sentences))
>my_texts
>[['these', 'are', 'two', 'crazy', 'sentences'], ['that', 'i', 'want', 'to', 'convert', 'back', 'and', 'forth']]

Answer 4

您可以将索引映射回字符的字典。

index_word = {v: k for k, v in tk.word_index.items()} # map back
seqs = tk.texts_to_sequences(txt1)
words = []
for seq in seqs:
    if len(seq):
        words.append(index_word.get(seq[0]))
    else:
        words.append(' ')
print(''.join(words)) # output
>>> 'what makes this problem difficult is that the sequences can vary in length  
>>> be comprised of a very large vocabulary of input symbols and may require the model  
>>> to learn the long term context or dependencies between symbols in the input sequence '

但是，在问题中，您尝试使用字符序列来预测 10 个类的输出，这不是序列到序列模型。在这种情况下，您不能只是将预测（或pred.argmax(axis=1)）转回字符序列。

Answer 5

    p_test = model.predict(data_test).argmax(axis =1)
#Show some misclassified examples
misclassified_idx = np.where(p_test != Ytest)[0]
len(misclassified_idx) 
i= np.random.choice(misclassified_idx)
print((i))
print((df_test[i]))
print('True label %s Predicted label %s' , (Ytest[i], p_test[i]))
df_test is the original text
data_test is sequence of integer