Method and apparatus for generating recommendation reason, and storage medium

Patent metadata

Field	Value
Publication number	US-12561724-B2
Application number	US-202118004008-A
Country	US
Kind code	B2
Filing date	Jun 25, 2021
Priority date	Jul 1, 2020
Publication date	Feb 24, 2026
Grant date	Feb 24, 2026

Abstract

The embodiment of the present application provides a method and apparatus for generating a recommendation reason, a device and a storage medium. Content information of an object to be recommended is acquired, and the recommendation reason of the object to be recommended is generated according to a pre-trained recommendation reason generation model and the content information, where the training data used by the recommendation reason generation model includes question-answer data and content information of multiple objects. In this technical solution, since the training of the recommendation reason generation model takes the question-answer data and the content information of multiple objects into account, commodities that users care about most are mined through the question-answer data, therefore, the recommendation reason generated by this solution can accurately target users' needs and improve user experience.

First claim

The invention claimed is:

1 . A method for generating a recommendation reason, applied to a server, comprising:
acquiring content information of an object to be recommended; and generating the recommendation reason of the object to be recommended according to a pre-trained recommendation reason generation model and the content information, training data used by the recommendation reason generation model comprising question-answer data and content information of multiple objects; wherein before the generating the recommendation reason of the object to be recommended according to the pre-trained recommendation reason generation model and the content information, the method further comprises: acquiring the training data; and training an initial model based on the training data and a target recommendation reason corresponding to the multiple objects to obtain the recommendation reason generation model; wherein the initial model comprises: a prior network module, configured to obtain a first recommendation reason according to a content information vector; a posterior network module, configured to obtain a second recommendation reason according to the content information vector and a question-answer data vector; and a decoding module, configured to obtain a predicted recommendation reason according to the first recommendation reason, the second recommendation reason and a target recommendation reason; wherein the initial model further comprises: a loss function calculating module, configured for at least one of the following purposes: obtaining a first loss function according to the question-answer data vector and the target recommendation reason, the first loss function being used to indicate to adjust a relevant parameter in the decoding module; obtaining a second loss function according to the predicted recommendation reason and the target recommendation reason, the second loss function being used to indicate to adjust a relevant parameter in the decoding module; and obtaining a third loss function according to the first recommendation reason and the second recommendation reason, the third loss function being used to indicate to adjust a relevant parameter in the prior network module and/or a relevant parameter in the posterior network module; wherein the first loss function is an REG loss function of the recommendation reason generated by the question-answer data, the second loss function is a KLL loss function of the recommendation reason generated by the posterior network module, and the third loss function is a KL distance of the prior network module and the posterior network module, wherein the KL distance is used to make the prior network module infinitely close to the posterior network module; wherein a final loss function is a sum of the first loss function, the second loss function and the third loss function, and the final loss function is used for training the initial model.

2 . The method according to claim 1 , wherein the generating the recommendation reason of the object to be recommended according to the pre-trained recommendation reason generation model and the content information comprises:
encoding the content information by using a preset encoder to obtain a content information vector; and generating the recommendation reason of the object to be recommended according to the pre-trained recommendation reason generation model and the content information vector.

3 . The method according to claim 2 , wherein the content information comprises a title and an attribute, the encoding the content information by using the preset encoder to obtain the content information vector comprises:
encoding the title and the attribute by using the preset encoder respectively to obtain a title vector and an attribute vector; and obtaining the content information vector according to the title vector, the attribute vector and a preset weight.

4 . The method according to claim 2 , wherein the preset encoder comprises at least one of the following encoders:
a bidirectional long short-term memory network (LSTM) encoder, a unidirectional LSTM encoder or a transformer model.

5 . The method according to claim 1 , wherein the acquiring the training data comprises:
acquiring the content information of the multiple objects; acquiring the question-answer data of the multiple objects; filtering the question-answer data and the content information to obtain valid data; and correspondingly, the training the initial model based on the training data and the target recommendation reason corresponding to the multiple objects to obtain the recommendation reason generation model comprises: training the initial model based on the content information of the multiple objects, the valid data and the target recommendation reason corresponding to the multiple objects to obtain the recommendation reason generation model.

6 . An apparatus for generating a recommendation reason, comprising a memory and a processor:
the memory is configured to store program instructions; and the processor is configured to invoke the program instructions in the memory to: acquire content information of an object to be recommended; and generate the recommendation reason of the object to be recommended according to a pre-trained recommendation reason generation model and the content information, the training data used by the recommendation reason generation model comprising question-answer data and content information of multiple objects; wherein the processor is further configured to: acquire the training data; and train an initial model based on the training data and a target recommendation reason corresponding to the multiple objects to obtain the recommendation reason generation model; wherein the initial model comprises: a prior network module, configured to obtain a first recommendation reason according to a content information vector; a posterior network module, configured to obtain a second recommendation reason according to the content information vector and a question-answer data vector; and a decoding module, configured to obtain a predicted recommendation reason according to the first recommendation reason, the second recommendation reason and a target recommendation reason; wherein the initial model further comprises: a loss function calculating module, configured for at least one of the following purposes: obtaining a first loss function according to the question-answer data vector and the target recommendation reason, the first loss function being used to indicate to adjust a relevant parameter in the decoding module; obtaining a second loss function according to the predicted recommendation reason and the target recommendation reason, the second loss function being used to indicate to adjust a relevant parameter in the decoding module; and obtaining a third loss function according to the first recommendation reason and the second recommendation reason, the third loss function being used to indicate to adjust a relevant parameter in the prior network module and/or a relevant parameter in the posterior network module; wherein the first loss function is an REG loss function of the recommendation reason generated by the question-answer data, the second loss function is a KLL loss function of the recommendation reason generated by the posterior network module, and the third loss function is a KL distance of the prior network module and the posterior network module, wherein the KL distance is used to make the prior network module infinitely close to the posterior network module; wherein a final loss function is a sum of the first loss function, the second loss function and the third loss function, and the final loss function is used for training the initial model.

Assignees

• Beijing Wodong Tianjun Information Technology Co Ltd
• Beijing Jingdong Century Trading Co Ltd

Inventors

• Zhang Hainan
• Chen Hongshen
• Ding Zhuoye
• Bao Yongjun
• Yan Weipeng

Classifications

• G06F16/335

  Filtering based on additional data, e.g. user or group profiles (filtering in web context G06F16/9535, G06F16/9536) · CPC title

• G06F16/3329

  Natural language query formulation · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N5/04

  Inference or reasoning models · CPC title