Method of training dialog generation model for recommendations, method of generating recommendations, and device

Patent metadata

Field	Value
Publication number	US-12430516-B2
Application number	US-202218056137-A
Country	US
Kind code	B2
Filing date	Nov 16, 2022
Priority date	Feb 21, 2022
Publication date	Sep 30, 2025
Grant date	Sep 30, 2025

Abstract

The present disclosure provides a method of training an information generation model, a method of generating an information, an electronic device, and a storage medium. A specific implementation solution of the method of training the information generation model includes: splitting a description information for a target object in an information pair into at least one description word, so as to obtain a description word sequence, wherein the information pair further includes a first recommendation information; inputting the description word sequence into a dialog generation model to obtain a probability vector sequence for the target object, wherein each probability vector in the probability vector sequence includes probability values for a plurality of predetermined words; and training the dialog generation model according to the probability vector sequence and the first recommendation information, so as to obtain the information generation model.

First claim

What is claimed is:

1. A method of training an information generation model that is implemented by a processor, the method comprising:
splitting a description information for a target object associated with a product to be promoted in an information pair into at least one description word, so as to obtain a description word sequence, wherein the information pair further comprises a first recommendation information;
inputting the description word sequence into a dialog generation model configured to generate diverse and contextually relevant recommendation information to obtain a probability vector sequence for the target object, wherein each probability vector in the probability vector sequence comprises probability values for a plurality of predetermined words; and
training the dialog generation model according to the probability vector sequence and the first recommendation information, to generate recommendation information exhibiting diversity and contextual alignment with the description information, as enabled by the discrete latent variable and multi-task learning configuration of the dialog generation model, so as to obtain the information generation model,
wherein the probability vector sequence indicates a second recommendation information for the target object;
wherein the training the dialog generation model according to the probability vector sequence and the first recommendation information comprises:
determining, according to the probability vector sequence, a prediction probability that the second recommendation information comprises a prompt information;
determining a first loss value of the dialog generation model according to the prediction probability;
determining a second loss value of the dialog generation model according to an association relationship between the second recommendation information and the description information; and
training the dialog generation model according to the first loss value, the second loss value, and the first recommendation information,
wherein the dialog generation model comprises a pre-trained dialog generation model with a discrete latent variable;
wherein the inputting the description word sequence into the dialog generation model to obtain the probability vector sequence for the target object comprises: inputting a random identification information and the description word sequence into the dialog generation model to enable multi-task learning by jointly training for association prediction and text generation tasks, so as to obtain the probability vector sequence and an association prediction value corresponding to the random identification information, and
wherein the association prediction value indicates the association relationship between the second recommendation information and the description information,
wherein the generated recommendation information is configured for presentation to a user via a human-computer interaction interface or used for display in connection with the promoted item.

2. The method according to claim 1 , further comprising:
splitting the prompt information for the target object into at least one prompt word, so as to obtain a prompt word sequence;
wherein the inputting the description word sequence into a dialog generation model to obtain a probability vector sequence for the target object further comprises: inputting the description word sequence and the prompt word sequence into the dialog generation model to obtain the probability vector sequence.

3. The method according to claim 1 , wherein the determining, according to the probability vector sequence, the prediction probability that the second recommendation information comprises the prompt information comprises:
determining, according to a probability value in the probability vector sequence for each prompt word in the prompt word sequence, a probability that the second recommendation information contains each prompt word; and
determining, according to at least one probability that the second recommendation information contains at least one prompt word, the prediction probability that the second recommendation information comprises the prompt information.

4. The method according to claim 1 , further comprising:
determining, in response to repeated words being contained in the second recommendation information, a probability vector for the repeated words in the probability vector sequence as a target probability vector according to a position information of the repeated words in the second recommendation information; and
determining a third loss value of the dialog generation model according to the target probability vector and the repeated words;
wherein the training the dialog generation model according to the probability vector sequence and the first recommendation information further comprises: training the dialog generation model according to the first loss value, the second loss value, and the third loss value.

5. The method according to claim 1 , further comprising:
splitting the first recommendation information into at least one recommendation word, so as to obtain a recommendation word sequence;
determining, in response to repeated words being contained in the recommendation word sequence, a probability vector for the repeated words in the probability vector sequence as a target probability vector according to a position information of the repeated words in the recommendation word sequence; and
determining a third loss value of the dialog generation model according to the target probability vector and the repeated words;
wherein the training the dialog generation model according to the probability vector sequence and the first recommendation information further comprises: training the dialog generation model according to the first loss value, the second loss value, and the third loss value.

6. A method of generating an information that is implemented by a processor, the method comprising:
splitting a description information for an object to be recommended into at least one description word, so as to obtain a description word sequence;
inputting the description word sequence into an information generation model to obtain a probability vector sequence for the object to be recommended, wherein each probability vector in the probability vector sequence comprises probability values for a plurality of predetermined words; and
determining a recommendation information for the object to be recommended, according to the probability vector sequence,
wherein the information generation model is trained using a method of training an information generation model, comprising:
splitting a description information for a target object associated with a product to be promoted in an information pair into at least one description word, so as to obtain a description word sequence, wherein the information pair further comprises a first recommendation information;
inputting the description word sequence into a dialog generation model configured to generate diverse and contextually relevant recommendation information to obtain a probability vector sequence for the target object, wherein each probability vector in the probability vector sequence comprises probability values for a plurality of predetermined words; and
training the dialog generation model according to the probability vector sequence and the first recommendation information, to generate recommendation information exhibiting diversity and contextual alignment with the description information, as enabled by the discrete latent variable and multi-task learning configuration of the dialog generation model, so as to obtain the information generation model,
wherein the probability vector sequence indicates a second recommendation information for the target object;
wherein the trai

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Lei Zeyang
• Xu Xinchao
• Wu Wenquan
• Niu Zhengyu

Classifications

• G06F40/284

  Lexical analysis, e.g. tokenisation or collocates · CPC title

• G06F40/56

  Natural language generation · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title