CN-121996785-A - Method, device, medium and equipment for detecting text abstract generated by large model

CN121996785ACN 121996785 ACN121996785 ACN 121996785ACN-121996785-A

Abstract

The embodiment of the disclosure provides a method, a device, a medium and equipment for detecting text summaries generated by a large model. The method comprises the steps of obtaining target data, determining a data mode of the target data, extracting target texts corresponding to the target data if the data mode is a video or audio mode, generating a first text abstract of the target texts through a first large model, wherein the first text abstract comprises a plurality of time intervals divided according to the target texts and content abstracts of the time intervals, extracting text contents of the time intervals from the target texts according to the time intervals, determining whether the text contents of the time intervals are matched with the content abstracts of the time intervals through a second large model, and determining whether unreal contents exist in the first text abstract according to whether the text contents of the time intervals are matched with the content abstracts. By the method, whether the unreal content exists in the text abstract generated by the large model can be determined and detected efficiently and accurately.

Inventors

SHEN SITING
HUANG SHENWEI
LU SHIXIAN
LIAO HAIZHEN
SUN JINGWEI

Assignees

北京字跳网络技术有限公司

Dates

Publication Date: 20260508
Application Date: 20241106

Claims (11)

1. A method of detecting a text excerpt generated by a large model, comprising: If the data mode is a video mode or an audio mode, extracting a target text corresponding to the target data, and generating a first text abstract corresponding to the target text through a first large model, wherein the first text abstract comprises a plurality of time intervals divided according to the target text and content abstracts corresponding to the time intervals; Determining whether text content of each time interval in the plurality of time intervals is matched with the content abstract of the time interval through a second large model, and determining whether unreal content exists in the first text abstract according to whether text content of each time interval is matched with the content abstract of the time interval.
2. The method of claim 1, wherein determining whether there is unreal content in the first text excerpt based on whether text content for each time interval matches a content excerpt for the time interval comprises: If a target interval exists in the time intervals, determining that unreal content exists in the text abstract if the text content of the target interval is not matched with the content abstract of the target interval; And if the target interval does not exist in the time intervals, determining that the unreal content does not exist in the first text abstract.
3. The method of claim 2, wherein determining whether text content for each of the plurality of time intervals matches a content digest for the time interval comprises: for each time interval of time it is possible, If the length of the text abstract in the time interval is larger than a preset threshold value, extracting a plurality of first events in the text abstract through a third large model; If the first events which are not deduced according to the text content of the time interval exist in the first events, determining that the text content of the time interval is not matched with the content abstract of the time interval, and if the first events which are not deduced according to the text content of the time interval do not exist in the first events, determining that the text content of the time interval is matched with the content abstract of the time interval.
4. The method of claim 3, wherein determining whether text content for each of the plurality of time intervals matches a content digest for the time interval further comprises: for each time interval of time it is possible, If the length of the text abstract of the time interval is not greater than a preset threshold value, determining whether to infer the content abstract of the time interval according to the text content of the time interval through a second large model; if the content abstract of the time interval cannot be inferred according to the text content of the time interval, the text content of the time interval is determined to be not matched with the content abstract of the time interval.
5. The method of claim 1, wherein if the data modality is a video modality or an audio modality, extracting the target text corresponding to the target data comprises: and if the data mode is a video mode, extracting a first text corresponding to the target data through an optical character recognition tool, and extracting a second text corresponding to the target data through an automatic voice recognition tool, wherein the first text and the second text are used as the target text.
6. The method of claim 1, wherein if the data modality is a video modality or an audio modality, extracting the target text corresponding to the target data comprises: And if the data mode is an audio mode, extracting a second text corresponding to the target data through an automatic voice recognition tool, and taking the second text as the target text.
7. The method of claim 1, further comprising: If the data mode is a text mode, generating a second text abstract corresponding to the target data through a first large model, and extracting a plurality of second events in the second text abstract through a third large model; and determining whether each second event is inferred according to the target data or not through a second large model, and determining whether unreal content exists in the second text abstract or not according to whether each second event is inferred according to the target data or not.
8. The method of claim 7, wherein determining whether there is unreal content in the second text excerpt based on whether the respective second event is inferred from the target data comprises: And if the second events which are not inferred according to the target data do not exist in the second events, determining that the non-real content does not exist in the second text abstract.
9. An apparatus for detecting a text excerpt generated by a large model, comprising: The system comprises an acquisition unit, an extraction unit, a text extraction unit and a text extraction unit, wherein the acquisition unit is configured to acquire target data and determine a data mode of the target data, and if the data mode is a video mode or an audio mode, the target text corresponding to the target data is extracted, a first text abstract corresponding to the target text is generated through a first large model, and the first text abstract comprises a plurality of time intervals divided according to the target text and content summaries corresponding to the time intervals respectively; A judging unit configured to determine, through a second large model, whether text content of each of the plurality of time intervals matches a content digest of the time interval; and determining whether unreal content exists in the first text abstract according to whether the text content of each time interval is matched with the content abstract of the time interval.
10. A computer readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method of any of claims 1-8.
11. An electronic device comprising a memory having executable code stored therein and a processor, which when executing the executable code, implements the method of any of claims 1-8.

Description

Method, device, medium and equipment for detecting text abstract generated by large model Technical Field The embodiment of the disclosure relates to the technical field of computers, in particular to a method, a device, a medium and equipment for detecting text summaries generated by a large model. Background Large models refer to artificial intelligence models that have a large number of parameters on the order of one hundred million or more and are pre-trained on large-scale data. Text summaries generally refer to a brief, accurate general representation formed by processing a long text to extract key information therefrom. With the development of large model technology, generating a text abstract of a longer text through a large model is a common way to obtain the text abstract. However, due to the "illusion" problem of the large model, i.e. that content generated by the large model may appear as non-authentic content, it is often necessary to detect whether or not non-authentic content is present in the text excerpt generated by the large model. However, the existing scheme for detecting the text abstract generated by the large model still has the problems of low detection accuracy or overlarge labor cost. Disclosure of Invention The embodiment of the disclosure describes a method, a device, a medium and equipment for detecting text abstracts generated by large models. According to a first aspect, there is provided a method of detecting a text excerpt generated by a large model, comprising: If the data mode is a video mode or an audio mode, extracting a target text corresponding to the target data, and generating a first text abstract corresponding to the target text through a first large model, wherein the first text abstract comprises a plurality of time intervals divided according to the target text and content abstracts corresponding to the time intervals; Determining whether text content of each time interval in the plurality of time intervals is matched with the content abstract of the time interval through a second large model, and determining whether unreal content exists in the first text abstract according to whether text content of each time interval is matched with the content abstract of the time interval. According to a second aspect, there is provided an apparatus for detecting a text excerpt generated by a large model, comprising: The system comprises an acquisition unit, an extraction unit, a text extraction unit and a text extraction unit, wherein the acquisition unit is configured to acquire target data and determine a data mode of the target data, and if the data mode is a video mode or an audio mode, the target text corresponding to the target data is extracted, a first text abstract corresponding to the target text is generated through a first large model, and the first text abstract comprises a plurality of time intervals divided according to the target text and content summaries corresponding to the time intervals respectively; A judging unit configured to determine, through a second large model, whether text content of each of the plurality of time intervals matches a content digest of the time interval; and determining whether unreal content exists in the first text abstract according to whether the text content of each time interval is matched with the content abstract of the time interval. According to a third aspect, there is provided a computer readable storage medium having stored thereon a computer program which, when executed in a computer, causes the computer to perform the method of the first aspect. According to a fourth aspect, there is provided an electronic device comprising a memory having executable code stored therein and a processor which when executing the executable code implements the method of the first aspect. Methods, apparatuses, devices, and media for detecting text summaries generated by large models are provided in accordance with embodiments of the present disclosure. The method comprises the steps of obtaining target data, determining a data mode of the target data, extracting target text corresponding to the target data if the data mode is a video mode or an audio mode, generating a first text abstract corresponding to the target text through a first large model, wherein the first text abstract comprises a plurality of time intervals divided according to the target text and content abstracts of the time intervals, and extracting text contents of the time intervals from the target text according to the time intervals. And determining whether unreal content exists in the first text abstract according to whether the text content of each time interval is matched with the content abstract. By the method, whether unreal content exists in the text abstract generated by the large model can be determined and detected efficiently and accurately in an automatic mode. Drawings FIG. 1 shows a schematic diagram of a text excerpt generated by manually detecting