Method and apparatus for training machine reading comprehension model, and non- transitory computer-readable recording medium

Patent metadata

Field	Value
Publication number	US-12566969-B2
Application number	US-202318316744-A
Country	US
Kind code	B2
Filing date	May 12, 2023
Priority date	May 19, 2022
Publication date	Mar 3, 2026
Grant date	Mar 3, 2026

Abstract

A method and an apparatus for training a machine reading comprehension model, and a non-transitory computer-readable recording medium are provided. A training process is repeatedly performed using a training sample set to obtain a machine reading comprehension model. The training process includes inputting a sample article and a sample question into the machine reading comprehension model, generating a first predicted answer, and calculating a first loss between the first predicted answer and a sample answer; replacing the sample question with a mask to obtain a mask question, inputting the sample article and the mask question into the machine reading comprehension model, generating a second predicted answer corresponding to the mask question, and calculating a second loss between the second predicted answer and the sample answer; and updating the machine reading comprehension model so as to minimize a total loss.

First claim

What is claimed is:

1 . A method for training a machine reading comprehension model, the method comprising:
obtaining a training sample set, the training sample set including a plurality of training samples, and each of the training samples including a sample article, a sample question, and a sample answer corresponding to the sample question; and repeatedly performing a training process using the training sample set until a preset training termination condition is met, so as to obtain a trained machine reading comprehension model, wherein the training process includes
inputting the sample article and the sample question into the machine reading comprehension model, generating a first predicted answer corresponding to the sample question using the machine reading comprehension model, and calculating a first loss between the first predicted answer and the sample answer corresponding to the sample question;
replacing at least a part of the sample question with at least one mask to obtain a mask question, inputting the sample article and the mask question into the machine reading comprehension model, generating a second predicted answer corresponding to the mask question using the machine reading comprehension model, and calculating a second loss between the second predicted answer and the sample answer corresponding to the sample question; and
calculating a total loss according to the first loss and the second loss, and updating the machine reading comprehension model so as to minimize the total loss,
wherein the replacing of at least the part of the sample question with the at least one mask to obtain the mask question includes at least one of
replacing all words of the sample question with the mask to obtain a first mask question; and
replacing a part of the words of the sample question with the mask to obtain a second mask question, and
wherein the replacing of the part of the words of the sample question with the mask to obtain the second mask question includes
replacing all words except a preset interrogative sentence in the sample question with the mask to obtain the second mask question.

2 . The method for training a machine reading comprehension model as claimed in claim 1 ,
wherein the total loss is positively correlated with the first loss, and the total loss is negatively correlated with the second loss.

3 . The method for training a machine reading comprehension model as claimed in claim 2 ,
wherein the calculating of the total loss according to the first loss and the second loss includes
calculating the total loss using a first formula,
the first formula being loss=α 1 ·loss 1 −β 1 ·loss 2,1 , in a case where the mask question includes the first mask question,
the first formula being loss=α 2 ·loss 2 −β 2 ·loss 2,2 , in a case where the mask question includes the second mask question, and
the first formula being loss=α 3 ·loss 1 −β 3 ·loss 2,1 −β 4 ·loss 2,2 , in a case where the mask question includes the first mask question and the second mask question,
where loss represents the total loss, lossi represents the first loss, loss 2 , 1 represents the second loss between the second predicted answer corresponding to the first mask question and the sample answer, loss 2,2 represents the second loss between the second predicted answer corresponding to the second mask question and the sample answer, and β 1 , β 1 , α 2 , β 2 , α 3 , β 3 and β 4 represent preset weights, respectively, and are positive values.

4 . The method for training a machine reading comprehension model as claimed in claim 1 , the method further comprising:
predicting an answer with respect to an input article and an input question using the trained machine reading comprehension model.

5 . An apparatus for training a machine reading comprehension model, the apparatus comprising:
a memory storing computer-executable instructions; and one or more processors configured to execute the computer-executable instructions such that the one or more processors are configured to
obtain a training sample set, the training sample set including a plurality of training samples, and each of the training samples including a sample article, a sample question, and a sample answer corresponding to the sample question; and
repeatedly perform a training process using the training sample set until a preset training termination condition is met, so as to obtain a trained machine reading comprehension model,
wherein the training process includes
inputting the sample article and the sample question into the machine reading comprehension model, generating a first predicted answer corresponding to the sample question using the machine reading comprehension model, and calculating a first loss between the first predicted answer and the sample answer corresponding to the sample question;
replacing at least a part of the sample question with at least one mask to obtain a mask question, inputting the sample article and the mask question into the machine reading comprehension model, generating a second predicted answer corresponding to the mask question using the machine reading comprehension model, and calculating a second loss between the second predicted answer and the sample answer corresponding to the sample question; and
calculating a total loss according to the first loss and the second loss, and updating the machine reading comprehension model so as to minimize the total loss,
wherein the replacing of at least the part of the sample question with the at least one mask to obtain the mask question includes at least one of
replacing all words of the sample question with the mask to obtain a first mask question; and
replacing a part of the words of the sample question with the mask to obtain a second mask question,
wherein the one or more processors are configured to
replace all words except a preset interrogative sentence in the sample question with the mask to obtain the second mask question.

6 . The apparatus for training a machine reading comprehension model as claimed in claim 5 ,
wherein the total loss is positively correlated with the first loss, and the total loss is negatively correlated with the second loss.

7 . The apparatus for training a machine reading comprehension model as claimed in claim 6 ,
wherein the one or more processors are configured to
calculate the total loss using a first formula,
the first formula being loss=α 1 ·loss 1 −β 1 ·loss 2,1 , in a case where the mask question includes the first mask question,
the first formula being loss=α 2 ·loss 1 −β 2 ·loss 2,2 , in a case where the mask question includes the second mask question, and
the first formula being loss=α 3 ·loss 1 −β 3 ·loss 2,1 −β 4 ·loss 2,2 , in a case where the mask question includes the first mask question and the second mask question,
where loss represents the total loss, lossi represents the first loss, loss 2 , 1 represents the second loss between the second predicted answer corresponding to the first mask question and the sample answer, loss 2,2 represents the second loss between the second predicted answer corresponding to the second mask question and the sample answer, and α 1 , β 1 , α 2 , β 2 , α 3 , β 3 and β 4 represent preset weights, respectively, and are positive values.

8 . The apparatus for training a machine reading comprehension model as claimed in claim 5 ,
wherein the one or more processors are further configured to
predict an answer with respect to an input article and an input question using the trained machine reading comprehension model.

9 . A non-transitory computer-readable recording medium having computer-executable instructions for execution by one or more processors, wherein, the comp

Assignees

• Li Hongyu
• Dong Bin
• Jiang Shanshan
• Ding Lei
• Ricoh Co Ltd

Inventors

• Li Hongyu
• Dong Bin
• Jiang Shanshan
• Ding Lei

Classifications

• G06F40/30 Primary

  Semantic analysis · CPC title

• G06F40/166

  Editing, e.g. inserting or deleting · CPC title

• G06F40/40

  Processing or translation of natural language (natural language analysis G06F40/20; semantic analysis G06F40/30) · CPC title

• G06N3/094

  Adversarial learning · CPC title

• G06N3/09 Primary

  Supervised learning · CPC title