Event detection and classification method, apparatus, and device

Patent metadata

Field	Value
Publication number	US-12585686-B2
Application number	US-202318457107-A
Country	US
Kind code	B2
Filing date	Aug 28, 2023
Priority date	Aug 29, 2022
Publication date	Mar 24, 2026
Grant date	Mar 24, 2026

Abstract

Implementations of the present specification disclose an event risk detection method, apparatus, and device. The method includes: obtaining event description information provided by a plurality of different event initiators when the plurality of different event initiators each initiate a target event to a same event target party in a same event service; then converting, into a token sequence, a plurality of character sequences of the event description information provided by the plurality of different event initiators, the token sequence including a plurality of sub-token sequences each corresponding to a character sequence of event description information provided by an event initiator; setting a set of a first number of token positions for each sub-token sequence of the plurality of sub-token sequences, and sequentially placing characters in each sub-token sequence of the plurality of sub-token sequences at a corresponding set of the first number of token positions based on an order of each sub-token sequence; and determining, based on a corresponding sub-token sequence placed at each set of the first number of token positions, token information of an event initiator corresponding to each set of the first number of token positions, and a text classification model, whether the target event is at risk.

First claim

What is claimed is:

1 . A method, the method comprising:
obtaining event description information provided by a plurality of different event initiators, the plurality of different event initiators initiating a target event to a same event target party in a same event service, each event description information provided by an event initiator including a character sequence; converting, into a token sequence, a plurality of character sequences of the event description information provided by the plurality of different event initiators, the token sequence including a plurality of sub-token sequences each corresponding to a character sequence of event description information provided by an event initiator; setting a set of a first number of token positions for each sub-token sequence of the plurality of sub-token sequences, and sequentially placing characters in each sub-token sequence of the plurality of sub-token sequences at a corresponding set of the first number of token positions based on an order of each sub-token sequence; and classifying the target event, based on a sub-token sequence placed at a set of the first number of token positions, token information of an event initiator corresponding to the set of the first number of token positions, and a pre-trained text classification model, wherein the classifying includes:
for each event description information, obtaining an embedded feature corresponding to a character in the event description information;
obtaining, by using two layers of an encoder, an encoding result corresponding to the event description based on the embedded feature; and
classifying the target event based on comparing encoding results of the event description information provided by the plurality of different event initiators to determine whether the event description information provided by two or more of the plurality of event initiators is contradictory to one another.

2 . The method according to claim 1 , wherein the classifying the target event includes:
inputting, into the pre-trained text classification model, the corresponding sub-token sequence placed at each set of the first number of token positions and the token information of the event initiator corresponding to each set of the first number of token positions, to obtain a corresponding output result, and classifying, based on the output result, the target event.

3 . The method according to claim 1 , wherein the classifying the target event includes:
inputting, into the pre-trained text classification model, the corresponding sub-token sequence placed at each set of the first number of token positions and the token information of the event initiator corresponding to each set of the first number of token positions, to obtain a corresponding output result; and extracting, from the output result, an embedded feature corresponding to a first character in the event description information provided by each event initiator; inputting the embedded feature into the encoder to obtain a corresponding encoding result; and classifying, based on the encoding result, the target event.

4 . The method according to claim 3 , wherein the classifying, based on the encoding result, the target event includes:
determining similarities among a plurality of encoding results based on a similarity algorithm; and in response to that the similarities among the plurality of encoding results include a similarity value less than a similarity threshold, classifying the target event as a risk event.

5 . The method according to claim 3 , wherein the encoder is constructed by using a Transformer Block.

6 . The method according to claim 1 , further comprising:
obtaining historical event description information of a historical event in a plurality of different event services, wherein the historical event description information is provided by a plurality of different historical event initiators, the plurality of different historical event initiators having initiated the historical event to a same historical event target party in an event service of the plurality of different event services, each historical event description information provided by a historical event initiator including a character sequence; converting, into a token sequence sample, a plurality of character sequences of the historical event description information provided by the plurality of different historical event initiators, wherein the token sequence sample includes a plurality of sub-token sequence samples each corresponding to a character sequence of historical event description information provided by a historical event initiator; setting a set of a second number of token positions for each sub-token sequence sample of the plurality of sub-token sequence samples, and sequentially placing characters in each sub-token sequence sample of the plurality of sub-token sequence samples at a corresponding set of the second number of token positions based on an order of each sub-token sequence sample; and performing model training on the text classification model based on a corresponding sub-token sequence sample placed at each set of the second number of token positions and token information of a historical event initiator corresponding to each set of the second number of token positions, to obtain a trained text classification model.

7 . The method according to claim 1 , wherein the text classification model is constructed based on a Bidirectional Encoder Representations from Transformers (BERT) model.

8 . A computing system, comprising:
one or more processors; and one or more storage devices, the one or more storage devices storing computer-executable instructions, the executable instructions when executed by the one or more processors, causing the one or more processors to, individually or collectively, perform operations including: obtaining event description information provided by a plurality of different event initiators, the plurality of different event initiators initiating a target event to a same event target party in a same event service, each event description information provided by an event initiator including a character sequence; converting, into a token sequence, a plurality of character sequences of the event description information provided by the plurality of different event initiators, the token sequence including a plurality of sub-token sequences each corresponding to a character sequence of event description information provided by an event initiator; setting a set of a first number of token positions for each sub-token sequence of the plurality of sub-token sequences, and sequentially placing characters in each sub-token sequence of the plurality of sub-token sequences at a corresponding set of the first number of token positions based on an order of each sub-token sequence; and classifying the target event, based on a sub-token sequence placed at a set of the first number of token positions, token information of an event initiator corresponding to the set of the first number of token positions, and a pre-trained text classification model, wherein the classifying includes:
for each event description information, obtaining an embedded feature corresponding to a character in the event description information:
obtaining, by using two layers of an encoder, an encoding result corresponding to the event description based on the embedded feature; and
classifying the target event based on comparing encoding results of the event description information provided by the plurality of different event initiators to determine whether the event description information provided by two or more of the plurality of event initiators is contradictory to one another.

9 .

Assignees

• Alipay Hangzhou Inf Tech Co Ltd

Inventors

• Zhao Wenbiao
• Wang Haotian
• Wu Xiaofeng
• Wang Weiqiang

Classifications

• G06Q50/265 Primary

  Personal security, identity or safety · CPC title

• G06F40/284

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

• G06F40/216

  using statistical methods · CPC title

• G06F40/30

  Semantic analysis · CPC title

• G06F16/35 Primary

  Clustering; Classification · CPC title