Question answering method, method of training a question answering model, electronic device, and medium

Patent metadata

Field	Value
Publication number	US-12481683-B2
Application number	US-202318157452-A
Country	US
Kind code	B2
Filing date	Jan 20, 2023
Priority date	Jan 20, 2022
Publication date	Nov 25, 2025
Grant date	Nov 25, 2025

Abstract

A question answering method, a method of training a question answering model, a device, and a medium are provided, which relate to a field of artificial intelligence technology, in particular to fields of natural language processing technology, deep learning technology, and knowledge mapping technology. The question answering method includes: obtaining data to be processed, wherein the data to be processed includes a question and candidate answers; performing general semantic understanding on the data to be processed to obtain a general data feature; selecting a target question answering mode from candidate question answering modes based on the general data feature; and processing the general data feature by using the target question answering mode, to obtain a target answer for the question from the candidate answers.

First claim

What is claimed is:

1 . A method of training a question answering model, wherein the question answering model comprises:
a general understanding network, and a plurality of candidate question answering networks,
the general understanding network is a natural language processing model, and the candidate question answering network is a classification model, the method comprising:
obtaining a sample, wherein
the sample comprises a question, candidate answers and a sample label, and
the sample label represents an association between the candidate answers and the question;
performing general semantic understanding on the sample by using a general understanding network in the question answering model to be trained, to obtain a general data feature; selecting, for candidate question answering networks in the question answering model to be trained, a target question answering network from the plurality of candidate question answering networks based on the general data feature; processing the general data feature by using the target question answering network, to obtain the target answer for the question from the candidate answers; and comparing the target answer with the sample label to obtain a loss value, reversely adjusting a model parameter of the question answering model to be trained based on the loss value, the model parameter including a parameter of the general understanding network and parameters of the plurality of candidate question answering networks; wherein the performing of the general semantic understanding on the sample by using the general understanding network in the question answering model to be trained, to obtain a general data feature, comprises:
obtaining first knowledge data that comprises common sense data or professional data,
performing general semantic understanding on the sample based on the first knowledge data, to obtain a question feature, a candidate answer feature and a first association information,
wherein the first association information indicates an association between the question and the candidate answers, and
determining the question feature, the candidate answer feature and the first association information as the general data feature,
the candidate question answering network comprises a network label, and
the selecting a target question answering network from the plurality of candidate question answering networks based on the general data feature comprises:
selecting, based on a similarity between the general data feature and the network label, top N candidate questions answering networks ranked by similarity from the plurality of candidate question answering networks, as the target question answering network matched with a domain of the question, where N is an integer that is greater than or equal to 1, and
wherein the general data feature represents a domain information of the question, and the network label represents a domain targeted by the candidate question answering network.

2 . The method of claim 1 , wherein
the target question answering network comprises a plurality of target question answering networks; and the processing the general data feature by using the target question answering network, to obtain the target answer for the question from the candidate answers, comprises:
processing the general data feature by respectively using the plurality of target question answering networks, to obtain a plurality of target answers corresponding to the plurality of target question answering networks one by one;
determining, for each target question answering network, a weight of the target answer for each target question answering network based on the similarity; and
selecting the target answer for the question from the plurality of target answers based on the weight.

3 . The method of claim 1 , wherein
the candidate question answering networks comprise a plurality of candidate question answering networks; and the selecting a target question answering network from the candidate question answering networks based on the general data feature comprises:
determining, based on a computing resource, a number of target question answering network to be selected; and
selecting the number of target question answering networks from the plurality of candidate question answering networks based on the number and the general data feature.

4 . The method of claim 1 , wherein the processing the general data feature by using the target question answering network, to obtain a target answer for the question from the candidate answers, comprises:
obtaining second knowledge data; processing the general data feature by using the target question answering network, based on the second knowledge data, to obtain a second association information between the question and the candidate answers; and determining the target answer for the question from the candidate answers based on the second association information.

5 . The method of claim 1 , wherein the sample further comprises a description information for the question; and the general data feature further comprises a description feature for the description information.

6 . An electronic device, comprising:
at least one processor; and a memory communicatively coupled with 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 implement the method of claim 1 .

7 . The electronic device of claim 6 , wherein the target question answering network comprises a plurality of target question answering networks; and
the at least one processor is further configured to:
process the general data feature by using the plurality of target question answering networks respectively, to obtain a plurality of target answers corresponding to the plurality of target question answering networks one by one;
determine, for each target question answering network, a weight of the target answer for each target question answering network based on the similarity; and
select the target answer for the question from the plurality of target answers based on the weight.

8 . A non-transitory computer readable storage medium storing computer instructions, wherein the computer instructions are configured to cause a computer to implement the method of claim 1 .

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Jiang Wenbin
• Lv Yajuan
• Chai Chunguang
• Zhu Yong

Classifications

• G06F40/30

  Semantic analysis · CPC title

• Y02D10/00

  Energy efficient computing, e.g. low power processors, power management or thermal management · CPC title

• G06N3/084

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

• G06F16/36

  Creation of semantic tools, e.g. ontology or thesauri · CPC title

• G06F16/35

  Clustering; Classification · CPC title