CN-120705321-B - System and method for identifying large model generated text

CN120705321BCN 120705321 BCN120705321 BCN 120705321BCN-120705321-B

Abstract

The invention discloses a system and a method for identifying a large model generation text, wherein the system comprises an input module, a neural network module and an output module, the input module is used for acquiring a history input text, the history input text comprises a history human input text and a history large model input text, the neural network module is used for extracting probability differences and linguistic features in the history input text, dynamically fusing the probability differences and the linguistic features to obtain fused feature vectors, updating edge weights at the same time, the output module acquires a real-time input text, and meanwhile, calculating the large model generation probability of the real-time input text based on the updated edge weights, and judging whether the real-time input text is large model input. The invention not only can remarkably improve the detection precision, but also can keep lower calculation cost, is suitable for rapid screening of large-scale text data, provides a feasible solution for solving the current challenges, and is expected to promote the technical progress and development of the related fields.

Inventors

FANG SHUN
YU FANG
Fang Yingfeng
ZHANG ZHIHENG

Assignees

北京广安渲光科技有限公司

Dates

Publication Date: 20260505
Application Date: 20250617

Claims (4)

1. A system for identifying large model generation text is characterized by comprising an input module, a neural network module and an output module; the input module is used for acquiring a history input text, wherein the history input text comprises a history human input text and a history large model input text; the neural network module is used for extracting probability differences and linguistic features in the historical input text, dynamically fusing the probability differences and the linguistic features to obtain fused feature vectors, and updating edge weights; The output module acquires a real-time input text, calculates a large model generation probability of the real-time input text based on the updated edge weight, and judges whether the real-time input text is large model input or not; the neural network module comprises a multidimensional feature extraction layer, a binocular telescope core layer, a dynamic feature fusion layer and a graph neural network; The multidimensional feature extraction layer is used for extracting text features, grammar features, semantic consistency features and emotion consistency features of the historical input text and fusing the text features, grammar features, semantic consistency features and emotion consistency features into multidimensional features; The binocular telescope core layer calculates the phase difference confusion degree and the cross confusion degree based on the performer model and the observer model, and performs confusion degree fusion to obtain the confusion degree after fusion; the dynamic feature fusion layer is used for fusing the multidimensional features and the fused confusion degree to obtain the fused feature vector; The graph neural network is used for updating the edge weight by taking each dimension feature in the fused feature vector as a node through the graph neural network to obtain the updated edge weight; The workflow of the binocular core layer comprises: Obtaining confusion of the history input text by using the observer model and the performer model respectively, and calculating the confusion by using the confusion: Wherein, the Representing the confusion of the viewer model over the historically entered text, Representing the confusion of the performer model to the history input text, s representing the text sequence of the history input text, L representing the length of the text sequence, xi representing the ith Token in the text sequence, Representing the first i-1 Token of the text sequence s, Representing a model of an observer at a given point The probability of the i-th Token xi is predicted, Representing a model of a performer at a given time The probability of the i-th Token xi is predicted, Indicating a degree of confusion; respectively acquiring probability distribution of vocabulary in the history input text by using the observer model and the performer model, and calculating the cross confusion degree based on the probability distribution: Wherein, the Representing a model of an observer at a given point In the case of the vocabulary V, the probability distribution for the vocabulary V, Representing a model of a performer at a given time In the case of a logarithmic probability distribution for the vocabulary V, Representing the inner product between the two distributions, Representing a degree of cross confusion; And carrying out confusion degree fusion on the phase difference confusion degree and the cross confusion degree to obtain the confusion degree after fusion: Wherein, the Indicating the degree of confusion after fusion.
2. The system for identifying large model generated text according to claim 1, wherein the multidimensional feature extraction layer comprises a text feature extraction unit, a grammar feature extraction unit, a semantic consistency extraction unit, an emotion consistency extraction unit and a fusion unit; The text feature extraction unit is used for extracting text features in the history input text, wherein the text features comprise punctuation mark distribution, sentence length statistics, vocabulary richness and sentence head word distribution; The grammar feature extraction unit is used for extracting grammar tree depths in the history input text as the grammar features; The semantic consistency extraction unit is used for calculating similarity matrix variances in the historical input text to serve as the semantic consistency characteristics; the emotion consistency extraction unit is used for calculating emotion fluctuation variance and emotion conflict times in the historical input text as emotion consistency characteristics; the fusion unit is used for fusing the text feature, the grammar feature, the semantic consistency feature and the emotion consistency feature into the multidimensional feature.
3. The system for identifying large model generated text of claim 1, wherein the workflow of the dynamic feature fusion layer comprises: And performing gating weight calculation based on the multidimensional features and the confusion degree after fusion to obtain fusion weights: Wherein, the Representing a learnable weight matrix, sigma representing an activation function, The degree of confusion after fusion is indicated, The representation of the multi-dimensional features is performed, Representing the fusion weight; and fusing based on the fusion weight, the phase difference confusion degree and the multidimensional features to obtain the fused feature vector: Wherein, the Representing the fused feature vector.
4. A method for identifying large model generated text, the method being applied to the system of any of claims 1-3, comprising the steps of: s1, acquiring a history input text, wherein the history input text comprises a history human input text and a history large model input text; S2, extracting probability differences and linguistic features in the historical input text, dynamically fusing the probability differences and the linguistic features to obtain fused feature vectors, and updating edge weights; S3, acquiring a real-time input text, simultaneously calculating the large model generation probability of the real-time input text based on the updated edge weight, and judging whether the real-time input text is large model input or not.

Description

System and method for identifying large model generated text Technical Field The invention belongs to the technical field of text recognition, and particularly relates to a system and a method for recognizing a large model to generate text. Background As the application range of large language models is wider and wider, the potential risk is also increasing, and especially in the case of widely spread AI-generated content, it is important to be able to identify whether text is generated by a large model. In the information age, problems of false information, misleading content, and copyright disputes are layered, and these problems can be alleviated to a large extent by determining the source of the content. Confirmation of the true origin of text is particularly critical when dealing with areas requiring high accuracy and reliability, such as news stories, academic research, legal documents, etc. Further, on social media and online platforms, ensuring the authenticity of the posting content helps prevent the spread of rumors and the occurrence of social panic. Therefore, an efficient and accurate method is developed to detect whether the text is generated by a large language model, and the method has important significance for guaranteeing information safety and maintaining social order. The existing identification methods are mainly divided into two types, each of which has its limitations. The first is a method based on the variance of the confusion (Perplexity) which evaluates the complexity of the text by calculating its probability distribution, theoretically, the text generated by the language model should have a low confusion. However, this approach is susceptible to resistance attacks and performs poorly in handling high quality manually written text, as such text may also have a low degree of confusion. The second category of methods relies on specific linguistic features such as syntax tree depth or semantic consistency. While such methods can distinguish, to some extent, manually written text from machine-generated text, they tend to ignore the overall structure and style characteristics of the text, resulting in limited accuracy. Disclosure of Invention The invention aims to solve the defects in the prior art, and provides the following scheme: a system for identifying large model generation text includes an input module, a neural network module, and an output module; the input module is used for acquiring a history input text, wherein the history input text comprises a history human input text and a history large model input text; the neural network module is used for extracting probability differences and linguistic features in the historical input text, dynamically fusing the probability differences and the linguistic features to obtain fused feature vectors, and updating edge weights; And the output module acquires the real-time input text, calculates the large model generation probability of the real-time input text based on the updated edge weight, and judges whether the real-time input text is large model input or not. Preferably, the neural network module comprises a multidimensional feature extraction layer, a binoculars core layer, a dynamic feature fusion layer and a graph neural network; The multidimensional feature extraction layer is used for extracting text features, grammar features, semantic consistency features and emotion consistency features of the historical input text and fusing the text features, grammar features, semantic consistency features and emotion consistency features into multidimensional features; The binocular telescope core layer calculates the phase difference confusion degree and the cross confusion degree based on the performer model and the observer model, and performs confusion degree fusion to obtain the confusion degree after fusion; the dynamic feature fusion layer is used for fusing the multidimensional features and the fused confusion degree to obtain the fused feature vector; and the graph neural network is used for updating the edge weight by taking each dimension characteristic in the fused characteristic vector as a node through the graph neural network to obtain the updated edge weight. Preferably, the multidimensional feature extraction layer comprises a text feature extraction unit, a grammar feature extraction unit, a semantic consistency extraction unit, an emotion consistency extraction unit and a fusion unit; The text feature extraction unit is used for extracting text features in the history input text, wherein the text features comprise punctuation mark distribution, sentence length statistics, vocabulary richness and sentence head word distribution; The grammar feature extraction unit is used for extracting grammar input depth in the history input text as the grammar feature; The semantic consistency extraction unit is used for calculating similarity matrix variances in the historical input text to serve as the semantic consistency char