AI generated text detection method and system based on multi-granularity hidden space feature enhancement
By extracting multi-granular features and injecting noise into AI-generated text detection, and utilizing reconstruction networks and contrastive learning, the robustness problem in text perturbation scenarios is solved, achieving highly accurate and robust text detection.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING INST OF TECH
- Filing Date
- 2025-12-01
- Publication Date
- 2026-05-12
AI Technical Summary
Existing AI-generated text detection methods lack robustness in character perturbation scenarios and struggle to effectively distinguish between AI-generated text and human text.
By extracting coarse-grained document-level features and fine-grained fragment-level features from the text, injecting noise and using a reconstruction network for denoising, and combining contrastive learning, the text's discriminative power in the latent space is enhanced, thus constructing a detection method with multi-granular latent space feature enhancement.
It improves the accuracy and robustness of AI-generated text detection, and can maintain high detection performance in scenarios with text perturbation.
Smart Images
Figure CN122019769A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to an AI-generated text detection method and system based on multi-granularity latent space feature enhancement, belonging to the field of AI-generated text detection technology, and applied to AI-generated text fragment-level recognition scenarios. Background Technology
[0002] In recent years, the rise of large language models has greatly promoted the development of the field of artificial intelligence (AI). An increasing number of conversational AI products have significantly improved productivity thanks to their powerful text generation capabilities. However, because AI can generate text that is highly similar to human expression, while assisting people in writing, it also increases the risk of inappropriate use of AI-generated text. Against this backdrop, AI-generated text detection has emerged, primarily utilizing machine learning and deep learning algorithms to efficiently distinguish between AI-generated text and human text.
[0003] Current AI-generated text detection methods mainly fall into three categories: zero-shot detection methods, training-based detection methods, and watermark detection methods. Zero-shot detection methods extract inherent statistical features from the text and compare them with a pre-defined threshold to determine whether the text is AI-generated. These methods are easy to deploy, but their detection performance is difficult to guarantee. Training-based detection methods use labeled data to fine-tune a pre-trained model to complete the detection. These methods have high detection performance, but typically require a large amount of labeled data and computational resources. Watermark detection methods first divide the candidate word list into red and green lists. During text generation, the probability of words in the green list being selected is increased, thereby expanding the proportion of green list words in the final generated text. During detection, if the proportion of green list words in the text exceeds a specified threshold, it is considered AI-generated text. This method also has good detection performance, but it usually requires the AI service provider to implement it, lacking flexibility.
[0004] While current detection methods each have their strengths, they all suffer from insufficient robustness. Research has found that even minor perturbations to the text, such as deleting or adding a letter, swapping the order of two letters, or changing a punctuation mark, can interfere with the detection model's judgment. This significantly impacts the reliability of AI-generated text detection methods in real-world applications.
[0005] Therefore, improving the robustness of AI-generated text recognition models in character perturbation scenarios has become an urgent problem to be solved. Summary of the Invention
[0006] The purpose of this invention is to address the technical problem of improving the robustness of AI-generated text recognition models in character perturbation scenarios, and to propose an AI-generated text detection method and system based on multi-granularity latent space feature enhancement.
[0007] The objective of this invention is achieved through the following technical solution:
[0008] The AI-generated text detection method based on multi-granularity latent space feature enhancement disclosed in this invention includes the following steps:
[0009] Step 1: Obtain AI-generated text and human text data, perform lexicalization on the text data and convert it into word vectors;
[0010] Step 1.1: Divide the AI-generated text and human text into training and test sets;
[0011] Step 1.2: Use the word segmenter of the pre-trained language model to perform lexicalization on the text;
[0012] Step 1.3: Use the word embedding layer of the pre-trained language model to convert the word-element data into word vectors;
[0013] Step 2: Extract coarse-grained document-level features and fine-grained fragment-level features from the text;
[0014] Step 2.1: Input the word vectors into the pre-trained language model. Through the calculation of the pre-trained language model, the hidden state vector corresponding to each word in each layer of the pre-trained language model is output. The hidden state vector corresponding to the first word in the last layer of the pre-trained language model is used as the coarse-grained document-level feature of the corresponding text data.
[0015] Step 2.2: Input the word vectors into the Disconnected Recurrent Neural Network (DNN) and output the hidden state vectors corresponding to the text segments within a sliding window of a specified size.
[0016] Step 2.3: The hidden state vectors corresponding to the multiple text segments obtained in Step 2.2 are fused into a single feature vector through pooling operations, which serves as the fine-grained segment-level feature of the corresponding text data;
[0017] Step 3: Inject noise into the latent space of document-level and fragment-level features;
[0018] Step 3.1: Input the document-level and fragment-level feature vectors into a multilayer perceptron, and encode the two feature vectors into the low-dimensional latent space as shown in Equation (1):
[0019]
[0020] Among them, z doc and z seg Let these represent the document-level and fragment-level feature vectors in the low-dimensional latent space, respectively. and H represents a multilayer perceptron used to encode document-level and fragment-level feature vectors into a low-dimensional latent space, respectively. doc and h seg These represent the document-level and fragment-level feature vectors in the original feature space, respectively.
[0021] Step 3.2: Extract the document-level and fragment-level feature vectors z from the low-dimensional latent space. view Decompose into semantic terms using two fully connected layers respectively. and disturbance terms
[0022] Step 3.3: Add to the disturbance term Inject Gaussian noise and combine semantic terms. The noisy low-dimensional latent space feature representation z′ is composed as shown in equation (2). view ;
[0023]
[0024] in, Indicates Gaussian noise. Represents a vector in which all elements are 1;
[0025] Step 3.4: Using regular expressions For semantic items and disturbance terms The scale and noise level are constrained as shown in equation (3);
[0026]
[0027] Here, α is a hyperparameter used to adjust the degree of constraint on noise;
[0028] Step 4: Use the reconstruction network to denoise the document-level and fragment-level features and restore the original features;
[0029] Step 4.1: Represent the noisy low-dimensional latent space features z′ view The input is fed into the reconstruction network for decoding. The structure of the reconstruction network is a multilayer perceptron, and the output is the noise-removed reconstruction feature as shown in Equation (4).
[0030]
[0031] in, The reconstructed feature vector after removing noise. To reconstruct the network;
[0032] Step 4.2: Calculate the reconstructed feature vectors With the original feature vector h view Mean square error between As shown in equation (5);
[0033]
[0034] Step 4.3: Calculate the average reconstruction loss of the reconstruction network over the entire training set. As shown in equation (6);
[0035]
[0036] in, For the test set data, β is a hyperparameter used to adjust the effect of the regularization term.
[0037] Step 5: Perform contrastive learning using the document-level and fragment-level features restored from the positive and negative samples respectively; positive samples refer to samples of the same category being positive samples of each other, and negative samples refer to samples of different categories being negative samples of each other. Calculate the contrastive loss as shown in Equation (7).
[0038]
[0039] Where, n + Indicates the number of positive samples. Represents the reconstructed features of positive samples. This represents the reconstructed features of all positive and negative samples except for the sample itself; sim represents the similarity calculation; and τ is the temperature coefficient.
[0040] Step 6: Integrate document-level and fragment-level text features for text classification and learn from the AI-generated text detection task;
[0041] Step 6.1: Refactor document-level features and fragment-level reconstruction features The data is fused, and the corresponding class probability is input into the classifier. The binary cross-entropy loss is then calculated as the classification loss.
[0042] Step 6.2: Calculate the overall loss function of the model. As shown in equation (8);
[0043]
[0044] Among them, γ1 and γ2 are two hyperparameters used to balance the importance of reconstruction loss and contrast loss;
[0045] Step 6.3: Update the model parameters using gradient descent to obtain the minimum loss function. Using the minimum loss function Train the model to obtain the trained model;
[0046] Step 7: Input the test set into the trained model to obtain the recognition results of the AI-generated text;
[0047] The AI-generated text detection system based on multi-granularity latent space feature enhancement disclosed in this invention is used to implement the above method, including a data preprocessing module, a document-level feature extraction module, a fragment-level feature extraction module, a feature latent space noise enhancement module, a contrastive learning module, a feature fusion module, and a classification and recognition module;
[0048] The data preprocessing module is used to clean the input text and perform word embedding processing, which will serve as the input to the document-level feature extraction module and the fragment-level feature extraction module.
[0049] The document-level feature extraction module is used to extract global semantic features from the preprocessed text through a pre-trained language model, which will be used as input to the feature latent space noise enhancement module.
[0050] The fragment-level feature extraction module is used to extract local semantic features from the disconnected recurrent neural network, which will serve as the input to the contrastive learning module.
[0051] The feature latent space noise enhancement module is used to process noise in the global and local semantic features of the variational autoencoder and will be used as the input of the feature fusion module.
[0052] The contrastive learning module, which uses contrastive loss to enhance the feature discrimination between positive and negative samples, will be used as the input to the feature fusion module.
[0053] The feature fusion module is used to fuse document-level features and fragment-level features, which will be used as input to the classification and recognition module.
[0054] The classification and recognition module is used to generate recognition results for text using AI.
[0055] Compared with existing technologies, it has the following beneficial effects:
[0056] 1. In this invention, by simultaneously mining coarse-grained document-level features and fine-grained fragment-level features, both AI-generated text and human text are comprehensively characterized, effectively modeling the features of different types of text in the hidden space and improving the accuracy of AI-generated text detection.
[0057] 2. In this invention, on the one hand, noise is injected into the hidden space to simulate real text perturbation. This method is more flexible and avoids the tedious process of manually constructing perturbation text. On the other hand, a reconstruction network is used to denoise the feature vectors, enhancing the model's ability to resist noise. Simultaneously, contrastive learning enhances the distinguishability between AI-generated text and human text in the hidden space. Combining these techniques, this invention can construct more robust feature representations, greatly enhancing the ability of AI-generated text detection methods to cope with text perturbation. Attached Figure Description
[0058] Figure 1 This is a framework diagram of the method provided in the embodiments of the present invention;
[0059] Figure 2 This is a Disconnected RNN framework. Detailed Implementation
[0060] To better illustrate the purpose and advantages of this invention, the invention will be further described below with reference to the accompanying drawings and examples. It should be noted that the implementation of this invention is not limited to the following embodiments, and any modifications or alterations made to this invention will fall within the scope of protection of this invention.
[0061] Example
[0062] like Figure 1 As shown in the figure, the AI-generated text detection method based on multi-granularity latent space feature enhancement in this embodiment has the following specific implementation steps:
[0063] Step 1: Obtain AI-generated text and human text data, perform lexicalization on the text data and convert it into word vectors;
[0064] Step 1.1: Divide the AI-generated text and human text into training and test sets;
[0065] Step 1.2: Use the word segmenter of the pre-trained language model to perform lexicalization on the text;
[0066] Step 1.3: Use the word embedding layer of the pre-trained language model to convert the word-element data into word vectors;
[0067] In this embodiment, AI-generated text and human text are divided into training and test sets in a 7:3 ratio; the text is lexicalized using the tokenizer of the pre-trained language model RoBERTa-base; and the lexicalized data is converted into word vectors using the word embedding layer of the pre-trained language model RoBERTa-base.
[0068] Step 2: Extract coarse-grained document-level features and fine-grained fragment-level features from the text;
[0069] Step 2.1: Input the word vectors into the pre-trained language model. Through the calculation of the pre-trained language model, the hidden state vector corresponding to each word in each layer of the pre-trained language model is output. The hidden state vector corresponding to the first word in the last layer of the pre-trained language model is used as the coarse-grained document-level feature of the corresponding text data.
[0070] Step 2.2: Input the word vectors into the Disconnected Recurrent Neural Network (DNN) and output the hidden state vectors corresponding to the text segments within a sliding window of a specified size.
[0071] Step 2.3: The hidden state vectors corresponding to the multiple text segments obtained in Step 2.2 are fused into a single feature vector through pooling operations, which serves as the fine-grained segment-level feature of the corresponding text data;
[0072] In this embodiment, word vectors are input into a pre-trained language model, RoBERTa-base. Through the calculations of the pre-trained language model, the hidden state vector corresponding to each word in each layer of the pre-trained language model is output. The hidden state vector corresponding to the first word in the last layer of the pre-trained language model is used as the coarse-grained document-level feature of the corresponding text data. The word vectors are then input into a system such as... Figure 2 In the Disconnected RNN shown, the sliding window size is set to 10, and the output is the hidden state vector corresponding to the text segment within the sliding window. The hidden state vectors corresponding to multiple text segments are fused into a feature vector by max pooling, which serves as the fine-grained segment-level feature of the corresponding text data.
[0073] Step 3: Inject noise into the latent space of document-level and fragment-level features;
[0074] Step 3.1: Input the document-level and fragment-level feature vectors into a multilayer perceptron, and encode the two feature vectors into the low-dimensional latent space as shown in Equation (1):
[0075]
[0076] Among them, z doc and z seg Let these represent the document-level and fragment-level feature vectors in the low-dimensional latent space, respectively. and H represents a multilayer perceptron used to encode document-level and fragment-level feature vectors into a low-dimensional latent space, respectively. doc and h seg These represent the document-level and fragment-level feature vectors in the original feature space, respectively.
[0077] Step 3.2: Extract the document-level and fragment-level feature vectors z from the low-dimensional latent space. view Decompose into semantic terms using two fully connected layers respectively. and disturbance terms
[0078] Step 3.3: Add to the disturbance term Inject Gaussian noise and combine semantic terms. The noisy low-dimensional latent space feature representation z′ is composed as shown in equation (2). view ;
[0079]
[0080] in, Indicates Gaussian noise. Represents a vector in which all elements are 1;
[0081] Step 3.4: Using regular expressions For semantic items and disturbance terms The scale and noise level are constrained as shown in equation (3);
[0082]
[0083] Here, α is a hyperparameter used to adjust the degree of constraint on noise;
[0084] In this embodiment, α is set to 1;
[0085] Step 4: Use the reconstruction network to denoise the document-level and fragment-level features and restore the original features;
[0086] Step 4.1: Represent the noisy low-dimensional latent space features z′ view The input is fed into the reconstruction network for decoding. The structure of the reconstruction network is a multilayer perceptron, and the output is the noise-removed reconstruction feature as shown in Equation (4).
[0087]
[0088] in, The reconstructed feature vector after removing noise. To reconstruct the network;
[0089] Step 4.2: Calculate the reconstructed feature vectors With the original feature vector h view Mean square error between As shown in equation (5);
[0090]
[0091] Step 4.3: Calculate the average reconstruction loss of the reconstruction network over the entire training set. As shown in equation (6);
[0092]
[0093] in, For the test set data, β is a hyperparameter used to adjust the effect of the regularization term;
[0094] In this embodiment, β is set to 0.5;
[0095] Step 5: Perform contrastive learning using the document-level and fragment-level features restored from the positive and negative samples respectively; positive samples refer to samples of the same category being positive samples of each other, and negative samples refer to samples of different categories being negative samples of each other. Calculate the contrastive loss as shown in Equation (7).
[0096]
[0097] Where, n + Indicates the number of positive samples. Represents the reconstructed features of positive samples. This represents the reconstructed features of all positive and negative samples except for the sample itself; sim represents the similarity calculation; and τ is the temperature coefficient.
[0098] In this embodiment, τ is a temperature coefficient, set to 0.1;
[0099] Step 6: Integrate document-level and fragment-level text features for text classification and learn from the AI-generated text detection task;
[0100] Step 6.1: Refactor document-level features and fragment-level reconstruction features The data is fused, and the corresponding class probability is input into the classifier. The binary cross-entropy loss is then calculated as the classification loss.
[0101] Step 6.2: Calculate the overall loss function of the model. As shown in equation (8);
[0102]
[0103] Among them, γ1 and γ2 are two hyperparameters used to balance the importance of reconstruction loss and contrast loss;
[0104] Step 6.3: Update the model parameters using gradient descent to obtain the minimum loss function. Using the minimum loss function Train the model to obtain the trained model;
[0105] In this embodiment, both γ1 and γ2 are set to 0.01;
[0106] Step 7: Input the test set into the trained model to obtain the recognition results of the AI-generated text;
[0107] The AI-generated text detection system based on multi-granularity latent space feature enhancement disclosed in this invention is used to implement the above method, including a data preprocessing module, a document-level feature extraction module, a fragment-level feature extraction module, a feature latent space noise enhancement module, a contrastive learning module, a feature fusion module, and a classification and recognition module;
[0108] The data preprocessing module is used to clean the input text and perform word embedding processing, which will serve as the input to the document-level feature extraction module and the fragment-level feature extraction module.
[0109] The document-level feature extraction module is used to extract global semantic features from the preprocessed text through a pre-trained language model, which will be used as input to the feature latent space noise enhancement module.
[0110] The fragment-level feature extraction module is used to extract local semantic features from the disconnected recurrent neural network, which will serve as the input to the contrastive learning module.
[0111] The feature latent space noise enhancement module is used to process noise in the global and local semantic features of the variational autoencoder and will be used as the input of the feature fusion module.
[0112] The contrastive learning module, which uses contrastive loss to enhance the feature discrimination between positive and negative samples, will be used as the input to the feature fusion module.
[0113] The feature fusion module is used to fuse document-level features and fragment-level features, which will be used as input to the classification and recognition module.
[0114] The classification and recognition module is used to generate recognition results for text using AI.
[0115] This invention simultaneously mines both coarse-grained document-level features and fine-grained fragment-level features of text to perform comprehensive text modeling. It simulates text perturbation by injecting noise into the latent space and uses a reconstruction network for denoising to improve the method's resistance to noise. Furthermore, it enhances the discriminative power of different text categories in the latent space by utilizing contrastive learning. Ultimately, it improves detection performance while enhancing the robustness of the method, enabling it to maintain high detection accuracy even on texts with character and word-level perturbations.
[0116] The above detailed description further illustrates the purpose, technical solution, and beneficial effects of the invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An AI-generated text detection method based on multi-granularity latent space feature enhancement, characterized in that: Includes the following steps, Step 1: Obtain AI-generated text and human text data, perform lexicalization on the text data and convert it into word vectors; Step 2: Extract coarse-grained document-level features and fine-grained fragment-level features from the text; Step 2.1: Input the word vectors into the pre-trained language model. Through the calculation of the pre-trained language model, the hidden state vector corresponding to each word in each layer of the pre-trained language model is output. The hidden state vector corresponding to the first word in the last layer of the pre-trained language model is used as the coarse-grained document-level feature of the corresponding text data. Step 2.2: Input the word vectors into the disconnected recurrent neural network, and output the hidden state vectors corresponding to the text segments within a sliding window of a specified size; Step 2.3: The hidden state vectors corresponding to the multiple text segments obtained in Step 2.2 are fused into a single feature vector through pooling operations, which serves as the fine-grained segment-level feature of the corresponding text data; Step 3: Inject noise into the latent space of document-level and fragment-level features; Step 4: Use the reconstruction network to denoise the document-level and fragment-level features and restore the original features; Step 4.1: Represent the noisy low-dimensional latent space features z′ view The input is fed into the reconstruction network for decoding. The structure of the reconstruction network is a multilayer perceptron, and the output is the noise-removed reconstruction feature as shown in Equation (4). in, The reconstructed feature vector after removing noise. To reconstruct the network; Step 4.2: Calculate the reconstructed feature vectors With the original feature vector h view Mean square error between As shown in equation (5); Step 4.3: Calculate the average reconstruction loss of the reconstruction network over the entire training set. As shown in equation (6); in, For the test set data, β is a hyperparameter used to adjust the effect of the regularization term; Step 5: Perform contrastive learning using the document-level and fragment-level features restored from the positive and negative samples respectively; positive samples refer to samples of the same category being positive samples of each other, and negative samples refer to samples of different categories being negative samples of each other. Calculate the contrastive loss as shown in Equation (7). Where, n + Indicates the number of positive samples. Represents the reconstructed features of positive samples. This represents the reconstructed features of all positive and negative samples except for the sample itself; sim represents the similarity calculation; and τ is the temperature coefficient. Step 6: Integrate document-level and fragment-level text features for text classification and learn from the AI-generated text detection task; Step 7: Input the test set into the trained model to obtain the recognition results of the AI-generated text.
2. The AI-generated text detection method based on multi-granularity latent space feature enhancement as described in claim 1, characterized in that: Step 1 is implemented as follows: Step 1.1: Divide the AI-generated text and human-generated text into training and test sets; Step 1.2: Use the word segmenter of the pre-trained language model to perform lexicalization on the text; Step 1.3: Use the word embedding layer of the pre-trained language model to convert the lexicalized data into word vectors.
3. The AI-generated text detection method based on multi-granularity latent space feature enhancement as described in claim 1, characterized in that: Step 3 is implemented as follows: Step 3.1: Input the document-level and fragment-level feature vectors into a multilayer perceptron, and encode the two feature vectors into the low-dimensional latent space as shown in Equation (1): Among them, z doc and z seg Let these represent the document-level and fragment-level feature vectors in the low-dimensional latent space, respectively. and H represents a multilayer perceptron used to encode document-level and fragment-level feature vectors into a low-dimensional latent space, respectively. doc and h seg These represent the document-level and fragment-level feature vectors in the original feature space, respectively. Step 3.2: Extract the document-level and fragment-level feature vectors z from the low-dimensional latent space. view Decompose into semantic terms using two fully connected layers respectively. and disturbance terms Step 3.3: Add to the disturbance term Inject Gaussian noise and combine semantic terms. The noisy low-dimensional latent space feature representation z′ is composed as shown in equation (2). view ; in, Indicates Gaussian noise. Represents a vector in which all elements are 1; Step 3.4: Using regular expressions For semantic items and disturbance terms The scale and noise level are constrained as shown in equation (3); Here, α is a hyperparameter used to adjust the degree of constraint on noise.
4. The AI-generated text detection method based on multi-granularity latent space feature enhancement as described in claim 1, characterized in that: Step 6 is implemented as follows: Step 6.1: Refactor document-level features and fragment-level reconstruction features The data is fused, and the corresponding class probability is input into the classifier. The binary cross-entropy loss is then calculated as the classification loss. Step 6.2: Calculate the overall loss function of the model. As shown in equation (8); Among them, γ1 and γ2 are two hyperparameters used to balance the importance of reconstruction loss and contrast loss; Step 6.3: Update the model parameters using gradient descent to obtain the minimum loss function. Using the minimum loss function Train the model to obtain the trained model.
5. An AI-generated text detection system based on multi-granularity latent space feature enhancement, as described in claim 1, characterized in that: It includes a data preprocessing module, a document-level feature extraction module, a fragment-level feature extraction module, a feature latent space noise enhancement module, a contrastive learning module, a feature fusion module, and a classification and recognition module; The data preprocessing module is used to clean the input text and perform word embedding processing, which will serve as the input to the document-level feature extraction module and the fragment-level feature extraction module. The document-level feature extraction module is used to extract global semantic features from the preprocessed text through a pre-trained language model, which will be used as input to the feature latent space noise enhancement module. The fragment-level feature extraction module is used to extract local semantic features from the disconnected recurrent neural network, which will serve as the input to the contrastive learning module. The feature latent space noise enhancement module is used to process noise in the global and local semantic features of the variational autoencoder and will be used as the input of the feature fusion module. The contrastive learning module, which uses contrastive loss to enhance the feature discrimination between positive and negative samples, will be used as the input to the feature fusion module. The feature fusion module is used to fuse document-level features and fragment-level features, which will be used as input to the classification and recognition module. The classification and recognition module is used to generate recognition results for text using AI.