Spoken language understanding

1. A method of spoken language understanding, the method comprising:

receiving audio data for a spoken language expression;

encoding the audio data using a basic encoder of a multi-stage encoder to obtain character features, wherein the basic encoder is trained to generate the character features during a first training phase by appending a softmax layer to the basic encoder and comparing an output of the softmax layer to ground-truth character training data, and wherein the softmax layer is removed prior to encoding the audio data;

encoding the character features using a sequential encoder of the multi-stage encoder to obtain token features, wherein the sequential encoder is trained to generate the token features during a second training phase based on ground-truth token training data; and

decoding the token features to generate semantic information representing the spoken language expression.

2. The method of claim 1 , further comprising:

generating a spectrogram based on the audio data; and

dividing the spectrogram into a plurality of frames, wherein the multi-stage encoder takes the frames as input.

3. The method of claim 2 , further comprising:

generating a sequence of character feature vectors using the basic encoder, wherein each of the sequence of character feature vectors corresponds to one of the frames; and

generating a sequence of token feature vectors based on the sequence of character feature vectors using the sequential encoder.

4. The method of claim 3 , wherein:

the character feature vectors are generated independently of each other.

5. The method of claim 3 , further comprising:

generating a first token feature vector; and

generating a second token feature vector based at least in part on the first token feature vector.

6. The method of claim 1 , further comprising:

identifying a decoding position;

computing a ratio between an output sequence of the multi-stage encoder and a subsequent input sequence; and

computing a sum of encoder states based on the decoding position and the ratio, wherein the decoding is based on the sum of encoder states.

7. The method of claim 1 , further comprising:

generating a response to the spoken language expression based on the semantic information.

8. The method of claim 1 , wherein:

the semantic information comprises contextual information.

9. The method of claim 1 , wherein:

the semantic information includes attribute names, attribute values, or both.

10. A method of training a neural network for spoken language understanding, the method comprising:

training a basic encoder to generate character features based on a spoken language expression in a first phase by appending a softmax layer to the basic encoder and comparing an output of the softmax layer to ground-truth character training data;

removing the softmax layer from the basic encoder;

training a sequential encoder to generate token features based on the spoken language expression in a second phase based on ground-truth token training data;

combining the basic encoder, the sequential encoder and a decoder in sequence to produce an end-to-end neural network for spoken language understanding; and

training the end-to-end network to generate semantic information for the spoken language expression.

11. The method of claim 10 , further comprising:

appending one or more linear layers with a log-softmax output function to the basic encoder;

predicting the character features for the spoken language expression using the one or more linear layers;

comparing the predicted character features to ground-truth character features; and

adjusting parameters of the basic encoder based on the comparison.

12. The method of claim 11 , further comprising:

removing the log-softmax output function prior to combining the basic encoder, the sequential encoder, and the decoder.

13. The method of claim 10 , further comprising:

appending a linear layer and a sequential decoder to layers of a basic model;

predicting token features for the spoken language expression using the linear layer and the sequential decoder;

comparing the predicted token features to ground-truth token features; and

adjusting parameters of the layers of the basic model, the linear layer, and the sequential decoder based on the comparison.

14. The method of claim 10 , further comprising:

predicting semantic information for the spoken language expression using the end-to-end network;

comparing the predicted semantic information to ground-truth semantic information; and

updating parameters of the basic encoder, the sequential encoder and the decoder based on the comparison.

15. The method of claim 10 , further comprising:

training the end-to-end network is based on a connectionist temporal classification (CTC) loss.

16. An apparatus for spoken language understanding, comprising:

a basic encoder configured to generate character features based on audio data for a spoken language expression, wherein the basic encoder is trained in a first training phase by appending a softmax layer to the basic encoder and comparing an output of the softmax layer to ground-truth character training data, and wherein the softmax layer is removed prior to generating character features;

a sequential encoder configured to generate token features based on an output of the basic encoder, wherein the sequential encoder is trained during a second training phase based on ground-truth token training data; and

a decoder configured to generate semantic information for the spoken language expression based on an output of the sequential encoder, wherein the decoder is trained together with the basic encoder and the sequential encoder during a third training phase.

17. The apparatus of claim 16 , wherein:

the basic encoder comprises one or more convolutional neural network (CNN) layers and one or more recurrent neural network (RNN) layers.

18. The apparatus of claim 16 , wherein:

the basic encoder is trained using a linear layer with a log-softmax output function.

19. The apparatus of claim 16 , wherein:

the sequential encoder comprises one or more CNN layers, one or more RNN layers, a linear layer, and a sequential decoder.

20. The apparatus of claim 16 , wherein:

the apparatus is trained using an incremental training process including the first training phase for training the basic encoder, the second training phase for training the sequential encoder, and the third training phase for training the neural network as a whole.