如何在 Tensorflow-keras 中对 nlp 使用预测？

Question

I have bit of problem when predicting named entity recognition set. After i trained and tested all went good. Now i want to test on raw data like strings .

I tried to use

model.predict(['Elon musk is good guy , he owns spacex, tesla.'])

but it throws erorr,

UnimplementedError:  Cast string to float is not supported
     [[node functional_29/Cast (defined at <ipython-input-210-e13dae4a124d>:1) ]] [Op:__inference_predict_function_223088]

Function call stack:
predict_function

I have token2index and

tag2index , dictionaries built from trained set. I tried to convert it and use these but predicted shows 0 on all ,

word = ['Elon musk is good guy , he owns spacex, tesla.']
word_index = [[token2idx[word] for word in word]]
X = pad_sequences(sequences=word_index, maxlen=7, padding='post')
predicted = np.argmax(model.predict(X), axis=-1) 
print(predicted)

gives array([[0, 0, 0, 0, 0, 0, 0]]) which is not true. Even tried a snippet sentence of x_train[0] but it throws like this. Thanks you for helping.

Answer 1

I guess you want to predict words, right?

Then you should split your words:

sentence = 'Elon musk is good guy , he owns spacex, tesla.'
word_index = [[token2idx[word] for word in sentence.split(' ')]]
X = pad_sequences(sequences=word_index, maxlen=7, padding='post')
predicted = np.argmax(model.predict(X), axis=-1) 
print(predicted)

Update

As the discussion showed, the issue was in the model having a high accuracy during learning, but the output was always zero.

Since your y-class sizes are not even distributed, the models learns, that improving the prediction for one class will improve the accuracy very fast very quickly. So your y data is something like this: [0,0,0,0,0,0,0,1,0,0,0,0,0,3,0] . With three classes : 0,1,3 the model learns quickly to predict zeros well, since that increases the accuracy the most. But

the model does only learn to predict 0, which already gives it a high accuracy. EG when one sequence contains of 20 words, so 20 y values and 19 are 0, the model will reach an accuracy of 95% by just predicting 0 all the time. So a high accuracy is in that case no measure for the quality of the model, since for increasing the performance of the model for all classes, a jump from 95% to 98% does improve the model way more than the jump from 50% to 95%.