Event argument extraction method, event argument extraction apparatus and electronic device

Patent metadata

Field	Value
Publication number	US-11880397-B2
Application number	US-202017036833-A
Country	US
Kind code	B2
Filing date	Sep 29, 2020
Priority date	Mar 20, 2020
Publication date	Jan 23, 2024
Grant date	Jan 23, 2024

Abstract

An event argument extraction (EAE) method, an EAE apparatus and an electronic device, relates to the technical field of knowledge graphs. A specific implementation scheme includes acquiring a to-be-extracted event content; and performing argument extraction on the to-be-extracted event content based on a trained EAE model, to obtain a target argument of the to-be-extracted event content; where the trained EAE model is obtained by training a pre-trained model with event news annotation data and a weight of each argument annotated in the event news annotation data.

First claim

What is claimed is:

1. An event argument extraction (EAE) method, comprising:
acquiring a to-be-extracted event content; and
performing argument extraction on the to-be-extracted event content based on a trained EAE model, to obtain a target argument of the to-be-extracted event content;
wherein the trained EAE model is obtained by training a pre-trained model with event news annotation data and a weight of each argument annotated in the event news annotation data,
wherein the trained EAE model is obtained at least in following manner:
training the pre-trained model in accordance with the event news annotation data and a loss function, to obtain the trained EAE model, wherein the loss function is associated with a predicted probability value, predicted by the pre-trained model, of each argument annotated in the event news annotation data and the weight of each argument annotated in the event news annotation data,
wherein the performing argument extraction on the to-be-extracted event content based on a trained EAE model, to obtain a target argument of the to-be-extracted event content comprises:
predicting an event type of the to-be-extracted event content using a trained event type categorization model;
predicting an event descriptive sentence of the to-be-extracted event content using a trained event sentence discriminator model;
constructing to-be-extracted question answering (QA) data corresponding to the to-be-extracted event content based on the event type, the event descriptive sentence, an argument role corresponding to the event type, and a preset QA format, wherein a format of the to-be-extracted QA data matches the preset QA format, and the to-be-extracted QA data comprises the event descriptive sentence of the to-be-extracted event content, the event type of the to-be-extracted event content, and the argument role corresponding to the event type; and
inputting the to-be-extracted QA data to the trained EAE model, and performing argument extraction using the trained EAE model, to obtain the target argument.

2. The EAE method according to claim 1 , wherein the loss function is a weighted sum of negative log-likelihoods of predicted probability values, predicted by the pre-trained model, of all arguments annotated in the event news annotation data.

3. The EAE method according to claim 1 , wherein the performing argument extraction on the to-be-extracted event content based on the trained EAE model, to obtain the target argument of the to-be-extracted event content comprises:
performing argument extraction on the to-be-extracted event content based on the trained EAE model, to obtain a predicted probability value of at least one argument in the to-be-extracted event content; and
determining an argument with a maximum predicted probability value and an argument with a predicted probability value greater than a probability threshold in the at least one argument as the target argument;
wherein the probability threshold is equal to the maximum predicted probability value multiplied by a preset coefficient, and the preset coefficient is a positive number less than or equal to 1.

4. The EAE method according to claim 1 , wherein the training the pre-trained model in accordance with the event news annotation data and the loss function, to obtain the trained EAE model comprises:
performing a format transformation on the event news annotation data in accordance with a preset question answering (QA) format, to obtain news QA data; and
training the pre-trained model in accordance with the news QA data and the loss function, to obtain the trained EAE model.

5. An electronic device, comprising:
at least one processor; and
a memory communicatively connected to the at least one processor,
wherein the memory stores therein instructions configured to be executed by the at least one processor, and the at least one processor is configured to:
acquire a to-be-extracted event content; and
perform argument extraction on the to-be-extracted event content based on a trained EAE model, to obtain a target argument of the to-be-extracted event content;
wherein the trained EAE model is obtained by training a pre-trained model with event news annotation data and a weight of each argument annotated in the event news annotation data,
wherein the trained EAE model is obtained at least in following manner:
training the pre-trained model in accordance with the event news annotation data and a loss function, to obtain the trained EAE model, wherein the loss function is associated with a predicted probability value, predicted by the pre-trained model, of each argument annotated in the event news annotation data and the weight of each argument annotated in the event news annotation data,
wherein the processor is further configured to:
predict an event type of the to-be-extracted event content using a trained event type categorization model;
predict an event descriptive sentence of the to-be-extracted event content using a trained event sentence discriminator model;
construct to-be-extracted question answering (QA) data corresponding to the to-be-extracted event content based on the event type, the event descriptive sentence, an argument role corresponding to the event type, and a preset QA format, wherein a format of the to-be-extracted QA data matches the preset QA format, and the to-be-extracted QA data comprises the event descriptive sentence of the to-be-extracted event content, the event type of the to-be-extracted event content, and the argument role corresponding to the event type; and
input the to-be-extracted QA data to the trained EAE model, and perform argument extraction using the trained EAE model, to obtain the target argument.

6. The electronic device according to claim 5 , wherein the loss function is a weighted sum of negative log-likelihoods of predicted probability values, predicted by the pre-trained model, of all arguments annotated in the event news annotation data.

7. The electronic device according to claim 5 , wherein the processor is further configured to perform argument extraction on the to-be-extracted event content based on the trained EAE model, to obtain a predicted probability value of at least one argument in the to-be-extracted event content; and
determine an argument with a maximum predicted probability value and an argument with a predicted probability value greater than a probability threshold in the at least one argument as the target argument;
wherein the probability threshold is equal to the maximum predicted probability value multiplied by a preset coefficient, and the preset coefficient is a positive number less than or equal to 1.

8. The electronic device according to claim 5 , wherein the training the pre-trained model in accordance with the event news annotation data and the loss function, to obtain the trained EAE model comprises:
performing a format transformation on the event news annotation data in accordance with a preset QA format, to obtain news QA data; and
training the pre-trained model in accordance with the news QA data and the loss function, to obtain the trained EAE model.

9. A non-transitory computer readable storage medium storing therein computer instructions, wherein the computer instructions are configured to cause a computer to implement the EAE method comprising:
acquiring a to-be-extracted event content; and
performing argument extraction on the to-be-extracted event content based on a trained EAE model, to obtain a target argument of the to-be-extracted event content;
wherein the trained EAE model is obtained by training a pre-trained model with event news annotation data and a weight of each argument annotated in the event news annotation data,
wherein the t

Assignees

• Beijing Baidu Netcom Sci & Tech Co Ltd
• Beijing Baidu Netcom Sci & Tech Co Ltd

Inventors

• Li Fayuan
• Chen Yuguang
• Pan Lu
• Liu Yuanzhen
• Han Cuiyun
• Shi Xi
• Huang Jiayan

Classifications

• G06F16/3344 Primary

  using natural language analysis · CPC title

• G06F18/214

  Generating training patterns; Bootstrap methods, e.g. bagging or boosting · CPC title

• G06F18/2415

  based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate · CPC title

• G06V20/20

  in augmented reality scenes · CPC title

• G06V20/43

  of news video content · CPC title