Summary generation method, summary generation model training method, and computer device

1. An abstract description generating method, the method being implemented by a computer device and comprising:

obtaining an input data item;

inputting the input data item into a pre-trained abstract description generation model;

obtaining an eigenvector of the input data item through an encoding network of the pre-trained abstract description generation model;

decoding the eigenvector through a decoding network of the pre-trained abstract description generation model to generate a time sequence of hidden states; and

generating an abstract description of the input data item from the time sequence of hidden states,

wherein:

the encoding network and the decoding network are preliminarily-trained during a first training stage based on supervision of a first loss function representing an abstract description labeling loss to obtain preliminary model parameters for the encoding network and the decoding network; and

the preliminarily-trained encoding network and the decoding network are further trained during a second training stage by

obtaining a backward-derived hidden state of a previous time moment through backward derivation according to a hidden state of a current moment in a training time sequence of hidden state outputted by the encoding network; and

further training the abstract description generation model including the preliminarily-trained encoding network and the decoding network based on supervision of a second loss function representing a difference between the backward-derived hidden state of the previous time moment and actual hidden state of the previous time moment outputted by the decoding network.

2. The method according to claim 1 , further comprising:

inputting a labeled training sample into the pre-trained abstract description generation model prior to it being trained;

preliminarily training the encoding network and the decoding network of the abstract description generation model based on a supervision of the first loss function, wherein the encoding network generates an eigenvector of the training sample, and the decoding network decodes the eigenvector using a recurrent neural network (RNN) to obtain a time sequence of hidden states for the training sample;

obtaining backward-derived hidden states of a previous time moment each through backward derivation from each of the hidden state in the time sequence of hidden states for the training sample outputted by the decoding network during the preliminary training;

obtaining a value of the second loss function according to difference between the backward-derived hidden states and actual hidden states outputted by the decoding network at same time moments; and

obtaining final model parameters of the abstract description generation model including the encoding network and the decoding network determined based on a supervision of the second loss function to reach a preset threshold value.

3. The method according to claim 2 , wherein obtaining backward-derived hidden states through backward derivation comprises: back deriving each of the backward-derived hidden states using a reconfiguration network by processing the hidden states at each time moment outputted by the decoding network as input to the reconfiguration network at a corresponding time moment.

4. The method according to claim 3 , wherein obtaining one of the backward-derived hidden states comprises:

inputting a hidden state at current time moment outputted by the decoding network and a hidden state of a previous time moment of the current time moment in the reconfiguration network into a long short-term memory (LSTM) hidden layer at the current time moment in the reconfiguration network, to obtain a hidden state at the current time moment in the reconfiguration network; and

inputting the hidden state of the current time moment in the reconfiguration network into a fully connected layer, to obtain the backward-derived hidden state of the previous time moment.

5. The method according to claim 2 , wherein obtaining backward-derived hidden states through backward derivation comprises: obtaining, by using a back propagation algorithm, each of the backward-derived hidden states of the previous time moment through backward derivation from each of the hidden state in the time sequence of the hidden states outputted by the preliminarily-trained decoding network.

6. The method according to claim 2 , wherein preliminarily training the encoding network and the decoding network of the pre-trained abstract description generation model based on the supervision of the first loss function comprises:

inputting the labeled training sample into the encoding network, and extracting a feature of the training sample, to obtain an eigenvector of the training sample;

inputting the eigenvector into the decoding network to obtain the time sequence of hidden states for the training sample;

for each time moment in the time sequence of hidden states, generating a word according to corresponding hidden state in the sequence of hidden states;

obtaining a value of the first loss function according to the generated word and an actual label word corresponding to each time moment; and

obtaining preliminary model parameters of the abstract description generation model by minimizing the first loss function to below a predetermined threshold value.

7. An abstract description generation model training method, the method being implemented by a computer device and comprising:

inputting a labeled training sample into the abstract description generation model;

preliminarily training an encoding network and a decoding network of the abstract description generation model based on a supervision of a first loss function, wherein the encoding network generates an eigenvector of the training sample, and the decoding network decodes the eigenvector using a recurrent neural network (RNN) to obtain a time sequence of hidden states for the training sample;

obtaining backward-derived hidden states of a previous time moment each through backward derivation from each of the hidden state in the time sequence of hidden states for the training sample outputted by the decoding network during the preliminary training;

obtaining a value of a second loss function according to difference between the backward-derived hidden states and actual hidden states outputted by the decoding network at same time moments; and

obtaining final model parameters of the abstract description generation model including the encoding network and the decoding network determined based on a supervision of the second loss function to reach a preset threshold value.

8. The method according to claim 7 , wherein obtaining backward-derived hidden states through backward derivation comprises: back deriving each of the backward-derived hidden states using a reconfiguration network by processing the hidden states at each time moment outputted by the decoding network as input to the reconfiguration network at a corresponding time moment.

9. The method according to claim 8 , wherein obtaining one of the backward-derived hidden states comprises:

inputting a hidden state at current time moment outputted by the decoding network and a hidden state of a previous time moment of the current time moment in the reconfiguration network into a long short-term memory (LSTM) hidden layer at the current time moment in the reconfiguration network, to obtain a hidden state at the current time moment in the reconfiguration network; and

inputting the hidden state of the current time moment in the reconfiguration network into a fully connected layer, to obtain the backward-derived hidden state of the previous time moment.

10. The method according to claim 7 , wherein obtaining backward-derived hidden states through backward derivation comprises: obtaining, by using a back propagation algorithm, each of the backward-derived hidden states of the previous time moment through backward derivation from each of the hidden state in the time sequence of the hidden states outputted by the preliminarily-trained decoding network.

11. The method according to claim 7 , wherein preliminarily training the encoding network and the decoding network of the abstract description generation model based on the supervision of the first loss function comprises:

inputting the labeled training sample into the encoding network, and extracting a feature of the training sample, to obtain an eigenvector of the training sample;

inputting the eigenvector into the decoding network to obtain the time sequence of hidden states for the training sample;

for each time moment in the time sequence of hidden states, generating a word according to corresponding hidden state in the sequence of hidden states;

obtaining a value of the first loss function according to the generated word and an actual label word corresponding to each time moment; and

obtaining preliminary model parameters of the abstract description generation model by minimizing the first loss function to below a predetermined threshold value.

12. A computer device, comprising a memory and a processor, the memory storing a computer program, and the computer program, when executed by the processor, causing the processor to perform the following operations:

obtaining an input data item;

inputting the input data item into a pre-trained abstract description generation model, obtaining an eigenvector of the input data item through an encoding network of the pre-trained abstract description generation model;

decoding the eigenvector through a decoding network of the pre-trained abstract description generation model to generate a time sequence of hidden states; and

generating an abstract description of the input data item from the time sequence of hidden states,

wherein:

the encoding network and the decoding network are preliminarily-trained during a first training stage based on supervision of a first loss function representing an abstract description labeling loss to obtain preliminary model parameters for the encoding network and the decoding network; and

the preliminarily-trained encoding network and the decoding network are further trained during a second training stage by

obtaining a backward-derived hidden state of a previous time moment through backward derivation according to a hidden state of a current moment in a training time sequence of hidden state outputted by the encoding network; and

further training the abstract description generation model including the preliminarily-trained encoding network and the decoding network based on supervision of a second loss function representing a difference between the backward-derived hidden state of the previous time moment and actual hidden state of the previous time moment outputted by the decoding network.

13. A computer device, comprising: a memory and a processor, the memory storing a computer program, and the computer program, when executed by the processor, causing the processor to perform the following operations:

inputting a labeled training sample into an abstract description generation model;

preliminarily training an encoding network and a decoding network of the abstract description generation model based on a supervision of a first loss function, wherein the encoding network generates an eigenvector of the training sample, and the decoding network decodes the eigenvector using a recurrent neural network (RNN) to obtain a time sequence of hidden states for the training sample;

obtaining backward-derived hidden states of a previous time moment each through backward derivation from each of the hidden state in the time sequence of hidden states for the training sample outputted by the decoding network during the preliminary training;

obtaining a value of a second loss function according to difference between the backward-derived hidden states and actual hidden states outputted by the decoding network at same time moments; and

obtaining final model parameters of the abstract description generation model including the encoding network and the decoding network determined based on a supervision of the second loss function to reach a preset threshold value.

14. The computer device according to claim 13 , wherein the computer program, when executed by the processor, causes the processor to perform the following operation: back deriving each of the backward-derived hidden states using a reconfiguration network by processing the hidden states at each time moment outputted by the decoding network as input to the reconfiguration network at a corresponding time moment.

15. The computer device according to claim 14 , wherein the computer program, when executed by the processor, causes the processor to perform the following operations:

inputting a hidden state at current time moment outputted by the decoding network and a hidden state of a previous time moment of the current time moment in the reconfiguration network into a long short-term memory (LSTM) hidden layer at the current time moment in the reconfiguration network, to obtain a hidden state at the current time moment in the reconfiguration network; and

inputting the hidden state of the current time moment in the reconfiguration network into a fully connected layer, to obtain the backward-derived hidden state of the previous time moment.

16. The computer device according to claim 13 , wherein the computer program, when executed by the processor, causes the processor to perform the following operation: wherein obtaining backward-derived hidden states through backward derivation comprises: obtaining, by using a back propagation algorithm, each of the backward-derived hidden states of the previous time moment through backward derivation from each of the hidden state in the time sequence of the hidden states outputted by the preliminarily-trained decoding network.

17. The computer device according to claim 13 , wherein the computer program, when executed by the processor, causes the processor to perform the following operations:

inputting the labeled training sample into the encoding network, and extracting a feature of the training sample, to obtain an eigenvector of the training sample;

inputting the eigenvector into the decoding network to obtain the time sequence of hidden states for the training sample;

for each time moment in the time sequence of hidden states, generating a word according to corresponding hidden state in the sequence of hidden states;

obtaining a value of the first loss function according to the generated word and an actual label word corresponding to each time moment; and

obtaining preliminary model parameters of the abstract description generation model by minimizing the first loss function to below a predetermined threshold value.