Method and apparatus for creating dialogue, and storage medium
The present application discloses a method and an apparatus for generating a dialogue, and a storage medium. The method includes: obtaining problem information; inputting the problem information to a small sample learning model to generate a first feature; inputting the problem information to a deep learning (DL) model to generate a second feature; combining the first feature and the second feature to generate a feature sequence; and inputting the feature sequence to a fusion model to generate dialogue information corresponding to the problem information.
1. A method for generating a dialogue, performed by an electronic device, comprising:
obtaining problem information;
inputting the problem information to a small sample learning model to generate a first feature, wherein the first feature comprises a problem feature and a support set feature, the support set feature comprises a feature of a support set of the small sample learning model, and the support set comprises training data in the small sample learning model;
inputting the problem information to a deep learning (DL) model to generate a second feature, wherein the second feature comprises a low-order feature and a high-order feature, the low-order feature is adjacent to an input layer of the DL model, and the high-order feature is adjacent to a final output layer of the DI model;
combining the first feature and the second feature to generate a feature sequence;
inputting the feature sequence to a fusion model to generate dialogue information corresponding to the problem information; and
providing recommendation information for a user based on the dialogue information;
wherein combining the first feature and the second feature to generate the feature sequence comprises:
combining the first feature and the second feature in different layers by using a plurality of preset fusion operators to generate the feature sequence, wherein the feature operators comprise a splice operator, an inner product operator and a bilinear feature crossed product;
wherein inputting the problem information to the small sample learning model to generate the first feature comprises:
performing a feature extraction on the problem information by the small sample learning model to generate the problem feature; and
obtaining a support set corresponding to the problem information by the small sample learning model according to the problem feature, and obtaining the support set feature of the support set corresponding to the problem information,
wherein, obtaining the support set corresponding to the problem information by the small sample learning model according to the problem feature, comprises:
obtaining a plurality of candidate support sets by the small sample learning model, and obtaining an intention feature of each candidate support set;
generating a direct score of the problem feature relative to the plurality of the candidate support sets by the small sample learning model according to the problem feature and the intention feature of each candidate support set; and
selecting a support set corresponding to the problem information from the plurality of the candidates support sets according to the direct score of the problem feature relative to the plurality of candidate support sets.
2. The method of claim 1 , further comprising:
generating contextual features among the plurality of the candidate support sets according to the intention feature of each candidate support set; and
generating a comparison score according to the problem feature, the intention feature of each candidate support set and the contextual feature, and selecting the support set corresponding to the problem information from the plurality of the candidate support sets according to the direct score and the comparison score.
3. The method of claim 1 , wherein, the candidate support set comprises a plurality of problems, and the intention feature of the candidate support set is obtained by: obtaining problem features of a plurality of problems in the candidate support set; and generating the intention feature of the candidate support set according to the problem features of the plurality of problems.
4. The method of claim 1 , wherein, the small sample learning model, the DL model and the fusion model are obtained through joint training.
5. The method of claim 1 , wherein inputting the problem information to the DL model to generate the second feature comprises:
performing a feature extraction on the problem information by the DL model to generate feature information of the problem information; and
performing feature embedding processing on the feature information by the DL model to generate the low-order feature of the problem information;
performing feature fusion on the low-order feature by the DL model to generate the high-order feature.
6. The method of claim 5 , further comprising: receiving an external feature vector of the problem information by the DL model, wherein the external feature vector is generated by a preset vector representation model, and performing the feature fusion on the low-order feature by the DL model to generate the high-order feature comprises:
performing the feature fusion on the low-order feature and the external feature vector by the DL model to generate the high-order feature.
7. The method of claim 6 , further comprising: receiving an external knowledge vector of the problem information by the DL model, wherein the external knowledge vector is a vector representation corresponding to a target training sample matching the problem information in a preset training sample, and performing the feature fusion on the low-order feature by the DL model to generate the high-order feature comprises:
performing the feature fusion on the low-order feature, the external knowledge vector and the external feature vector to generate the high-order feature.
8. An apparatus for generating a dialogue, comprising:
one or more processors;
a memory storing instructions executable by the one or more processors;
wherein the one or more processors are configured to:
obtain problem information;
input the problem information to a small sample learning model to generate a first feature, wherein the first feature comprises a problem feature and a support set feature, the support set feature comprises a feature of a support set of the small sample learning model, and the support set comprises training data in the small sample learning model;
input the problem information to a deep learning (DL) model to generate a second feature, wherein the second feature comprises a low-order feature and a high-order feature, the low-order feature is adjacent to an input layer of the DL model, and the high-order feature is adjacent to a final output layer of the DI model;
combine the first feature and the second feature to generate a feature sequence;
input the feature sequence to a fusion model to generate dialogue information corresponding to the problem information; and
provide recommendation information for a user based on the dialogue information;
wherein the one or more processors are further configured to:
combine the first feature and the second feature in different layers by using a plurality of preset fusion operators to generate the feature sequence, wherein the feature operators comprise a splice operator, an inner product operator and a bilinear feature crossed product;
wherein the one or more processors are configured to:
perform a feature extraction on the problem information to generate the problem feature, and a support set corresponding to the problem information is obtained according to the problem feature; and
obtain a support set corresponding to the problem information by the small sample learning model according to the problem feature, and obtain the support set feature of the support set corresponding to the problem information,
wherein, the one or more processors are configured to:
obtain a plurality of candidate support sets, and obtain an intention feature of each candidate support set;
generate a direct score of the problem feature relative to the plurality of the candidate support sets according to the problem feature and the intention feature of each candidate support set; and
select a support set corresponding to the problem information from the plurality of the candidates support sets according to the direct score of the problem feature relative to the plurality of candidate support sets.
9. The apparatus of claim 8 , wherein the one or more processors are configured to:
generate contextual features among the plurality of the candidate support sets according to the intention feature of each candidate support set; and
generate a comparison score according to the problem feature, the intention feature of each candidate support set and the contextual feature, and selecting the support set corresponding to the problem information from the plurality of the candidate support sets according to the direct score and the comparison score.
10. The apparatus of claim 8 , wherein the candidate support set comprises a plurality of problems, and the one or more processors are configured to:
obtain problem features of a plurality of problems in the candidate support set; and
generate the intention feature of the candidate support set according to the problem features of the plurality of problems.
11. The method of claim 8 , wherein, the small sample learning model, the DL model and the fusion model are obtained through joint training.
12. The apparatus of claim 8 , wherein the one or more processors are configured to:
perform a feature extraction on the problem information by the DL model to generate feature information of the problem information; and
perform feature embedding processing on the feature information by the DL model to generate the low-order feature of the problem information;
perform feature fusion on the low-order feature by the DL model to generate the high-order feature.
13. The apparatus of claim 12 , wherein, the one or more processors are configured to: receive an external feature vector of the problem information, wherein the external feature vector is generated by a preset vector representation model; and perform the feature fusion on the low-order feature and the external feature vector to generate the high-order feature.
14. The apparatus of claim 13 , wherein, the one or more processors are configured to: receive an external knowledge vector of the problem information, the external knowledge vector is a vector representation corresponding to a target training sample matching the problem information in a preset training sample, and to perform the feature fusion on the low-order feature, the external knowledge vector and the external feature vector to generate the high-order feature.
15. A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are configured to cause a computer to execute a method for generating a dialogue, and the method comprises:
obtaining problem information;
inputting the problem information to a small sample learning model to generate a first feature, wherein the first feature comprises a problem feature and a support set feature, the support set feature comprises a feature of a support set of the small sample learning model, and the support set comprises training data in the small sample learning model;
inputting the problem information to a deep learning (DL) model to generate a second feature, wherein the second feature comprises a low-order feature and a high-order feature, the low-order feature is adjacent to an input layer of the DL model, and the high-order feature is adjacent to a final output layer of the DI model;
combining the first feature and the second feature to generate a feature sequence;
inputting the feature sequence to a fusion model to generate dialogue information corresponding to the problem information; and
providing recommendation information for a user based on the dialogue information;
wherein combining the first feature and the second feature to generate the feature sequence comprises:
combining the first feature and the second feature in different layers by using a plurality of preset fusion operators to generate the feature sequence, wherein the feature operators comprise a splice operator, an inner product operator and a bilinear feature crossed product;
wherein inputting the problem information to the small sample learning model to generate the first feature comprises:
performing a feature extraction on the problem information by the small sample learning model to generate the problem feature; and
obtaining a support set corresponding to the problem information by the small sample learning model according to the problem feature, and obtaining the support set feature of the support set corresponding to the problem information,
wherein, obtaining the support set corresponding to the problem information by the small sample learning model according to the problem feature, comprises:
obtaining a plurality of candidate support sets by the small sample learning model, and obtaining an intention feature of each candidate support set;
generating a direct score of the problem feature relative to the plurality of the candidate support sets by the small sample learning model according to the problem feature and the intention feature of each candidate support set; and
selecting a support set corresponding to the problem information from the plurality of the candidates support sets according to the direct score of the problem feature relative to the plurality of candidate support sets.