Method of generating text, method of training model, electronic device, and medium

Patent metadata

Field	Value
Publication number	US-12260186-B2
Application number	US-202217992436-A
Country	US
Kind code	B2
Filing date	Nov 22, 2022
Priority date	Jan 26, 2022
Publication date	Mar 25, 2025
Grant date	Mar 25, 2025

Abstract

A method of generating a text, a method of training a text generation model, an electronic device, and a storage medium, which relate to a field of a computer technology, in particular to fields of deep learning and natural language processing technologies. A specific implementation solution includes: determining a reference feature representation of a target semantic information; determining, based on the reference feature representation and at least one predetermined logical character, at least one sentence latent representation respectively corresponding to the at least one predetermined logical character; and generating a target text content based on the at least one sentence latent representation.

First claim

What is claimed is:

1. A method of generating a text, the method comprising:
determining a reference feature representation of a target semantic information;
determining, based on the reference feature representation and at least one predetermined logical character, at least one sentence latent representation respectively corresponding to the at least one predetermined logical character, wherein the determining the at least one sentence latent representation respectively corresponding to the at least one predetermined logical character comprises:
generating, for each predetermined logical character in the at least one predetermined logical character, an initial sentence latent representation corresponding to the predetermined logical character, by using the reference feature representation and the predetermined logical character, and
determining the at least one sentence latent representation based on the initial sentence latent representation; and

generating a target text content based on the at least one sentence latent representation.

2. The method according to claim 1 , wherein the generating a target text content based on the at least one sentence latent representation comprises:
determining, for each sentence latent representation in the at least one sentence latent representation, an (i+1) th text feature representation in a to-be-generated text sentence by using the sentence latent representation and i generated text feature representations in the to-be-generated text sentence, wherein the to-be-generated text sentence is a text sentence corresponding to the sentence latent representation, and i is an integer greater than or equal to 0; and
determining an (i+1) th text content based on the (i+1) th text feature representation.

3. The method according to claim 1 , wherein the generating a target text content based on the at least one sentence latent representation comprises:
determining, for each sentence latent representation in the at least one sentence latent representation, an (i+1) th text feature representation in a to-be-generated text sentence by using the sentence latent representation and i generated target text feature representations in the to-be-generated text sentence, wherein the to-be-generated text sentence is a text sentence corresponding to the sentence latent representation, and i is an integer greater than or equal to 0;
determining an (i+1) th target text feature representation based on the (i+1) th text feature representation and an auxiliary selection feature representation, wherein the auxiliary selection feature representation is obtained based on the reference feature representation; and
determining an (i+1) th text content based on the (i+1) th target text feature representation.

4. The method according to claim 1 , further comprising generating the reference feature representation based on the target semantic information, wherein the target semantic information comprises a title and at least one keyword.

5. The method according to claim 2 , further comprising generating the reference feature representation based on the target semantic information, wherein the target semantic information comprises a title and at least one keyword.

6. The method according to claim 3 , further comprising generating the reference feature representation based on the target semantic information, wherein the target semantic information comprises a title and at least one keyword.

7. A method of training a text generation model, the method comprising training the text generation model by using a training sample, so as to obtain a trained text generation model, wherein the training sample comprises a target sample semantic information and a sample text content, wherein the text generation model is configured to implement the method according to claim 1 .

8. The method according to claim 7 , wherein the training the text generation model by using a training sample, so as to obtain a trained text generation model comprises:
determining a sample reference feature representation of the target sample semantic information, wherein the target sample semantic information comprises a sample title and at least one sample keyword;
processing the sample reference feature representation and at least one predetermined logical character by using the text generation model, so as to obtain at least one sample sentence latent representation respectively corresponding to the at least one predetermined logical character;
obtaining a prediction keyword information for each sample sentence latent representation in the at least one sample sentence latent representation based on the sample sentence latent representation according to a bag-of-words prediction method; and
training the text generation model by using the prediction keyword information and a label, so as to obtain the trained text generation model, wherein the label is generated based on the sample text content, and the label indicates a keyword information in a sample text sentence corresponding to the sample sentence latent representation in the sample text content.

9. An electronic device, comprising:
at least one processor; and
a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions, when executed by the at least one processor, cause the at least one processor to at least:
determine a reference feature representation of a target semantic information;
determine, based on the reference feature representation and at least one predetermined logical character, at least one sentence latent representation respectively corresponding to the at least one predetermined logical character, wherein determination of the at least one sentence latent representation respectively corresponding to the at least predetermined logical character comprises:
generation, for each predetermined logical character in the at least one predetermined logical character, an initial sentence latent representation corresponding to the predetermined logical character, by using the reference feature representation and the predetermined logical character, and
determination of the at least one sentence latent representation based on the initial sentence latent representation; and

generate a target text content based on the at least one sentence latent representation.

10. The electronic device according to claim 9 , wherein the electronic device is further configured to:
determine, for each sentence latent representation in the at least one sentence latent representation, an (i+1) th text feature representation in a to-be-generated text sentence by using the sentence latent representation and i generated text feature representations in the to-be-generated text sentence, wherein the to-be-generated text sentence is a text sentence corresponding to the sentence latent representation, and i is an integer greater than or equal to 0; and
determine an (i+1) th text content based on the (i+1) th text feature representation.

11. The electronic device according to claim 9 , wherein the electronic device is further configured to:
determine, for each sentence latent representation in the at least one sentence latent representation, an (i+1) th text feature representation in a to-be-generated text sentence by using the sentence latent representation and i generated target text feature representations in the to-be-generated text sentence, wherein the to-be-generated text sentence is a text sentence corresponding to the sentence latent representation, and i is an integer greater than or equal to 0;
determine an (i+1) th target text feature representation

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Hu Zhe
• Liu Jiachen
• Xiao Xinyan

Classifications

• G06F40/30 Primary

  Semantic analysis · CPC title

• G06F40/56 Primary

  Natural language generation · CPC title

• G06N3/08

  Learning methods · CPC title

• G06F40/211

  Syntactic parsing, e.g. based on context-free grammar [CFG] or unification grammars · CPC title