An AIGC text detection method and system based on multi-dimensional feature fusion
By employing a multi-dimensional feature fusion method, combined with attention networks and adaptive thresholding, this approach addresses the performance bottleneck of existing AIGC text detection technologies in multi-model hybrid content generation, achieving efficient and accurate AIGC text detection and providing interpretable judgment results and high-risk sentence/segment annotation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HANGZHOU ZHULONG ZHIYUAN TECHNOLOGY CO LTD
- Filing Date
- 2026-02-27
- Publication Date
- 2026-06-02
AI Technical Summary
Existing AIGC text detection technologies struggle to simultaneously achieve accuracy, robustness against circumvention, cross-scenario/cross-model generalization ability, and interpretability of results when dealing with text rewriting/polishing and multi-model hybrid content. This results in a significant decline in detection performance, failing to meet the needs of complex real-world scenarios.
A multi-dimensional feature fusion method is adopted to extract feature vectors in parallel from four orthogonal dimensions: statistical linguistics, semantics and fluency, syntactic structure, and information theory. These vectors are then weighted and fused through an attention network. Combined with adaptive threshold determination and interpretability report generation, this method achieves efficient and accurate detection of AIGC text.
It effectively reduces the risk of detection being circumvented, is applicable to different generative models and application scenarios, provides clear judgment criteria and high-risk segment annotations, improves the credibility and usability of detection, and supports large-scale deployment and real-time response.
Smart Images
Figure CN122132553A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of natural language processing, and in particular to an AIGC text detection method and system based on multi-dimensional feature fusion. Background Technology
[0002] With the rapid development of large language model technologies (such as ChatGPT and Wenxin Yiyan), the quality of AIGC (AI Generated Content) generated text is constantly improving, increasingly approaching the level of human writing. However, the misuse of AIGC text has also brought a series of serious security and governance risks, including academic misconduct, the spread of misinformation, and phishing. Therefore, there is an urgent need for an effective technical means to accurately determine the source of text (human writing / AI generation).
[0003] Existing AIGC text detection technologies mainly fall into four categories, and each category has significant shortcomings: The method based on text feature engineering and traditional machine learning classifiers treats the detection problem as a binary classification task. After text preprocessing, statistical features such as vocabulary, syntax, and semantics are extracted and input into classifiers such as SVM and logistic regression for discrimination. This method relies on expert experience to design features, has insufficient cross-domain / cross-scenario generalization ability, and features with single or few dimensions are easily circumvented by adversarial operations such as rewriting, polishing, and synonym substitution, resulting in poor robustness. Methods based on language model probability / perplexity utilize the statistical difference that AI text often conforms to the model distribution and has lower perplexity for discrimination. However, this type of method is highly dependent on the selected language model, lacks stability when switching models or crossing models, and the perplexity features change significantly after simple rewriting or polishing of the text, leading to a decrease in detection reliability. Deep learning end-to-end detection methods (such as BERT / RoBERTa fine-tuning): Fine-tuning pre-trained models using large-scale labeled data to directly classify the original text end-to-end. This method can achieve high accuracy on specific datasets, but it does not generalize well to unfamiliar generative model texts or new application domains. It also has high computational complexity and inference costs, and is black-box in nature, making it difficult to interpret the judgment criteria. Digital watermark detection technology embeds detectable watermark information into AIGC text at the generation stage, and the detection stage identifies the source of the text through watermark recognition. Although this method is accurate, it requires the generation stage to embed the watermark in advance, and it cannot cover third-party models, historical existing text, or generated content without watermarks, thus limiting its versatility.
[0004] In summary, existing technologies struggle to simultaneously achieve accuracy, robustness against circumvention, cross-scenario / cross-model generalization ability, and interpretability of results. When faced with text rewriting / polishing or content generated by multiple models, detection performance drops significantly, and the technology cannot effectively explain "why it was judged as AI-generated," making it difficult to meet the application requirements of complex real-world scenarios.
[0005] Therefore, there is a need for an AIGC text detection method and system based on multi-dimensional feature fusion to solve the performance bottlenecks and the difficulty in interpreting judgment results in existing technologies for text rewriting / polishing and multi-model mixed content detection. Summary of the Invention
[0006] Purpose of the invention: This invention provides an AIGC text detection method and system based on multi-dimensional feature fusion, which achieves efficient, accurate and reliable detection of AIGC text, while taking into account anti-circumvention robustness, cross-scenario generalization ability and result interpretability. It solves the performance bottleneck and difficulty in interpreting judgment results in existing technologies in text rewriting / polishing and multi-model mixed content detection.
[0007] Technical Solution: To solve the above-mentioned technical problems, according to one aspect of the present invention, more specifically, an AIGC text detection method based on multi-dimensional feature fusion, characterized by comprising the following steps: S1. Multi-dimensional feature extraction: Receive the text to be detected and extract feature vectors in parallel from four mutually orthogonal dimensions: statistical linguistics, semantics and fluency, syntactic structure, and information theory to form a multi-dimensional evidence set. The features of the four dimensions respectively cover the differences between "human writing vs. AI generation" in terms of macro statistics, semantic fluency, syntactic structure, and information randomness. S2, Attention-weighted feature fusion: The multi-dimensional feature vector is Z-score standardized to eliminate the influence of dimensions. The standardized feature vector is then input into the attention network to learn the importance of each feature dimension to the discrimination task and assign dynamic weights. A fusion discrimination score is generated by weighted summation to achieve evidence selection and fusion that varies according to the text. S3. Adaptive Threshold Determination and Confidence Assessment: An adaptive threshold is dynamically generated based on the metadata of the text to be detected. The fusion discrimination score is compared with the adaptive threshold, and the determination result of "AI generated" or "human writing" is output. At the same time, the confidence score is calculated based on the difference between the fusion discrimination score and the threshold, providing a basis for graded handling for the business side. S4. Explainability Report Generation: Outputs an explanatory report that includes feature contribution analysis and highlights high-risk segments. The high-risk segments are identified by comparing the deviation between the sentence's local feature vector and global feature vector, making the judgment results auditable and verifiable.
[0008] Furthermore, the statistical linguistic features mentioned in step S1 include lexical diversity (TTR type-to-symbol ratio), average sentence length and standard deviation, goodness of fit of high-frequency word Zipf's law distribution, and frequency and pattern of punctuation usage; the semantic and fluency features include text perplexity, cosine similarity distribution of adjacent word vectors, semantic coherence score between sentences, and semantic repetition rate (overuse of synonyms); the syntactic structure features include part-of-speech tag distribution entropy, average / maximum depth of dependency syntax trees, and diversity of phrase structure rules (noun phrases, verb phrases, etc.); the information theory features include Shannon entropy, relative entropy with a standard human corpus (KL divergence), and point mutual information (PMI) distribution of adjacent words.
[0009] Furthermore, in step S1, the text perplexity is calculated using lightweight GPT-2, the cosine similarity of adjacent word vectors is calculated using Word2Vec, and the average depth of the syntax tree is obtained through dependency analysis.
[0010] Furthermore, the metadata mentioned in step S3 includes text length, domain, and style. The adaptive threshold is dynamically generated through a preset "metadata → threshold" mapping rule or function, so that different types of text, such as long technical documents and short social media comments, are judged by different criteria.
[0011] Furthermore, the feature contribution analysis described in step S4 quantifies and displays the contribution of each feature dimension and specific feature item to the final judgment, clearly informing the user of the main basis for the judgment.
[0012] According to another aspect of the present invention, and more specifically an AIGC text detection system based on multi-dimensional feature fusion, comprising: Feature extraction module: used to extract feature vectors in parallel from four orthogonal dimensions of the text to be detected: statistical linguistics, semantics and fluency, syntactic structure and information theory, to construct a complementary multidimensional evidence set; Feature fusion module: It is used to perform Z-score standardization on feature vectors, learn the dynamic weights of each feature dimension through attention network and sum them up to generate a fusion discrimination score, avoiding scene adaptation problems caused by fixed weighting or simple concatenation. The discrimination module is used to generate an adaptive threshold based on text metadata, compare the fused discrimination score with the adaptive threshold, output the judgment result of "AI generated" or "human writing", and calculate the confidence level based on the difference between the score and the threshold. Report generation module: Used to output interpretable reports that include feature contribution analysis and highlight high-risk segments, improving the transparency and usability of the results.
[0013] Furthermore, it also includes an offline construction unit, which is used to determine / update the human corpus benchmark (for calculating relative features such as KL divergence), train or calibrate the attention network with labeled data (so that it can stably output fusion scores), form an adaptive rule / mapping of "metadata → threshold" and perform threshold calibration in combination with the validation set.
[0014] Furthermore, it also includes an online inference unit, which is used to perform a real-time detection process for multi-dimensional feature extraction of the text to be detected, attention-weighted feature fusion, adaptive threshold determination and confidence assessment, and interpretability report generation, supporting large-scale deployment and real-time response.
[0015] Beneficial effects: (1) The present invention constructs a complementary evidence chain through four mutually orthogonal feature dimensions. Adversarial operations of a single dimension (such as rewriting, polishing, and synonym substitution) are difficult to deceive all feature dimensions at the same time, effectively reducing the risk of detection being breached. It is especially suitable for detection scenarios of text rewriting / polishing and multi-model hybrid content generation. (2) This invention does not rely on specific AIGC model fingerprints. It dynamically selects effective features through an attention mechanism and combines them with adaptive thresholds based on text metadata, so that the discrimination benchmark can be dynamically adjusted according to text type (length, domain, style) to adapt to the detection needs of different generation models and different application scenarios. (3) This invention upgrades from "giving conclusions" to "giving evidence reports". Through feature contribution analysis and high-risk sentence highlighting, it clearly presents the judgment basis, improves the credibility and usability of the results, supports auditing and manual review, and solves the core pain point of the black box nature of the existing technology; (4) The present invention adopts a modular structure design, with clear functions and low coupling of each module, which is easy to integrate into existing systems such as content risk control, academic plagiarism detection, and platform review. At the same time, feature extraction is based on efficient statistics and lightweight model calculation, avoiding heavy dependence on large language models, with low inference cost, which is conducive to large-scale deployment and real-time detection. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation
[0017] To make the technical solution of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0018] Reference Figure 1This paper proposes an AIGC text detection method based on multi-dimensional feature fusion. It constructs a complete closed loop of "evidence acquisition → evidence fusion → adaptive judgment → interpretation output," replacing "single fingerprint" with "multi-dimensional evidence chain," "fixed threshold" with "dynamic decision-making," and "black box conclusion" with "interpretable report." Specifically, it includes the following steps: Multi-dimensional feature extraction: After receiving the text to be detected, it does not rely on the probabilistic features of a single model or a certain type of shallow statistical features. Instead, it extracts feature vectors in parallel from four mutually orthogonal dimensions: statistical linguistics, semantics and fluency, syntactic structure, and information theory. This forms a multi-dimensional and complementary set of evidence, which comprehensively covers the essential differences between "human writing vs. AI generation" in terms of macro-statistical laws, semantic coherence and generation traces, structural complexity and pattern diversity, information content and distribution differences. This makes it difficult for attackers to deceive all feature dimensions at the same time through a single dimension.
[0019] Statistical linguistic features (macro-statistical profiling): used to characterize the overall statistical patterns of text, including lexical diversity (TTR type-word-symbol ratio), average sentence length and standard deviation, Zipf's law distribution fit of high-frequency words, punctuation usage frequency and patterns, etc. These features are more sensitive to generated text that is "template-based, homogenized, and overly regularized". Semantic and fluency features (semantic coherence and “generation traces”): These features are extracted using lightweight language models and word embedding models, including text perplexity, cosine similarity distribution of adjacent word vectors, semantic coherence scores between sentences, and semantic repetition rate (overuse of synonyms). These features can capture “fluent but potentially lacking in real information increments” of generation features. Specifically, perplexity can be calculated using lightweight GPT-2 and cosine similarity of adjacent word vectors can be calculated using Word2Vec. Syntactic structural features (structural complexity and pattern diversity): used to measure the complexity and variability of syntactic structures, including part-of-speech tag distribution entropy, average / maximum depth of dependency syntax trees, and diversity of phrase structure rules (noun phrases, verb phrases, etc.). Relevant indicators are obtained through dependency analysis, which can identify AI generation suspicions of "single structure and repetitive patterns". Information theory features (randomness, information content, and distribution differences): The complexity and "randomness" of a text are measured from the perspective of information content, including Shannon entropy, relative entropy (KL divergence) with a standard human corpus, and point mutual information (PMI) distribution of adjacent words. Among them, KL divergence is suitable for "relative discrimination" across domains by comparing it with the "benchmark distribution of human corpus".
[0020] Attention-weighted feature fusion: To address the issue of varying effective evidence dimensions across different text genres and lengths, this method avoids simple concatenation or fixed weighting. Instead, it achieves dynamic feature fusion through a combination of standardization and attention networks. Feature standardization: All extracted features are Z-score standardized to eliminate the influence of different dimensions between features and avoid the "natural dominance" of some features with larger numerical ranges. Attention weight learning: The standardized multidimensional feature vector is input into the attention network. The network is trained with labeled data and can automatically learn the importance of each feature dimension / feature item to the discrimination task, and dynamically allocate weights according to the characteristics of the input text. For example, in technical documents, the weights of features such as perplexity and syntax tree depth are automatically increased. Fusion score calculation: Each feature value is multiplied by its corresponding dynamic weight and summed to obtain a unified fusion score, enabling evidence selection and aggregation that varies according to the text.
[0021] Adaptive threshold determination and confidence assessment: To improve cross-scenario generalization ability and avoid false positives and false negatives caused by a fixed threshold, this method adopts an adaptive threshold determination mechanism based on text metadata. Adaptive threshold generation: The threshold is dynamically adjusted based on the metadata of the text to be detected. The metadata includes text length, domain, style, etc. For example, long technical documents and short social media comments can use different discrimination criteria. Specifically, it is generated through a preset "metadata → threshold" mapping rule or function, and can be combined with a validation set for threshold calibration. Final judgment rule: The fusion discrimination score is compared with an adaptive threshold. If the score is higher than the threshold, it is judged as "AI generated"; otherwise, it is judged as "human writing". Confidence assessment: The confidence level is calculated based on the difference between the fusion discrimination score and the threshold, providing a quantitative reference for the "stability of the conclusion" for the business side, which facilitates graded handling.
[0022] Explainability Report Generation: To address the difficulty in interpreting the "black box conclusions" of existing technologies, this method generates a structured, explainable report along with the judgment results, containing at least two core categories of content: Feature contribution analysis: Quantitatively display the contribution of each feature dimension and specific feature item to the final judgment, and clearly inform users "what is the main basis", for example, mark "low perplexity and simple syntactic structure" as the core reasons for judging AI generation; High-risk sentence segment highlighting: By comparing the deviation between the local feature vector of each sentence and the global feature vector of the text, abnormal sentence segments are identified, and the areas most likely to be generated by AI are highlighted in the original text, helping reviewers to quickly locate suspected generated segments and achieve auditable and manually reviewed judgment results.
[0023] Reference Figure 1An AIGC text detection system based on multi-dimensional feature fusion is described. The system corresponds to the steps of the aforementioned method and adopts a modular pipeline design for easy engineering deployment and expansion. Specifically, it includes: Feature extraction module: Used to perform multi-dimensional feature extraction steps. It has built-in statistical linguistics feature extraction sub-module, semantic and fluency feature extraction sub-module, syntactic structure feature extraction sub-module, and information theory feature extraction sub-module. It can extract feature vectors from the four orthogonal dimensions of the text to be detected in parallel to form a multi-dimensional complementary evidence set.
[0024] Feature fusion module: Used to perform attention-weighted feature fusion steps, including Z-score normalization unit and attention network unit. The normalization unit eliminates the influence of feature dimensions, and the attention network unit learns dynamic weights and calculates the fusion discrimination score, avoiding the limitations of fixed weighting or simple splicing.
[0025] The discrimination module is used to perform adaptive threshold determination and confidence assessment steps. It includes a metadata recognition unit, a threshold generation unit, and a confidence calculation unit. The metadata recognition unit parses information such as text length, domain, and style. The threshold generation unit dynamically generates an adaptive threshold. The confidence calculation unit outputs the confidence score based on the difference between the score and the threshold.
[0026] Report generation module: This module is used to perform the interpretable report generation steps. It integrates the feature contribution analysis unit and the high-risk sentence segment location unit to generate a structured report that includes the judgment results, confidence level, feature contribution analysis, and highlighted original text.
[0027] Offline building blocks: These are used to support the stable operation of the system, including determining / updating the human corpus benchmark (for calculating relative features such as KL divergence), training or calibrating the attention network with labeled data (so that it can stably output fusion scores), forming adaptive rules / mappings from "metadata to thresholds" and performing threshold calibration, which can be updated once or periodically.
[0028] Online inference unit: Used to execute the real-time detection process, it receives the text to be detected by the user, and sequentially calls the feature extraction module, feature fusion module, discrimination module, and report generation module to complete the end-to-end processing from text input to explanatory report output, supporting large-scale real-time detection.
[0029] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.
Claims
1. An AIGC text detection method based on multi-dimensional feature fusion, characterized in that, Includes the following steps: S1. Multi-dimensional feature extraction: Receive the text to be detected and extract feature vectors in parallel from four mutually orthogonal dimensions: statistical linguistics, semantics and fluency, syntactic structure, and information theory to form a multi-dimensional evidence set. The features of the four dimensions respectively cover the differences between "human writing vs. AI generation" in terms of macro statistics, semantic fluency, syntactic structure, and information randomness. S2, Attention-weighted feature fusion: The multi-dimensional feature vector is Z-score standardized to eliminate the influence of dimensions. The standardized feature vector is then input into the attention network to learn the importance of each feature dimension to the discrimination task and assign dynamic weights. A fusion discrimination score is generated by weighted summation to achieve evidence selection and fusion that varies according to the text. S3. Adaptive Threshold Determination and Confidence Assessment: An adaptive threshold is dynamically generated based on the metadata of the text to be detected. The fusion discrimination score is compared with the adaptive threshold, and the determination result of "AI generated" or "human writing" is output. At the same time, the confidence score is calculated based on the difference between the fusion discrimination score and the threshold, providing a basis for graded handling for the business side. S4. Explainability Report Generation: Outputs an explanatory report that includes feature contribution analysis and highlights high-risk segments. The high-risk segments are identified by comparing the deviation between the sentence's local feature vector and global feature vector, making the judgment results auditable and verifiable.
2. The AIGC text detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, The statistical linguistic features mentioned in step S1 include lexical diversity, average sentence length and standard deviation, Zipf's law distribution fit of high-frequency words, and punctuation usage frequency and pattern. The semantic and fluency features include text perplexity, cosine similarity distribution of adjacent word vectors, semantic coherence score between sentences, and semantic repetition rate; The syntactic structural features include part-of-speech tag distribution entropy, average / maximum depth of dependency syntax tree, and diversity of phrase structure rules; The information-theoretic features include Shannon entropy, relative entropy to a standard human corpus (KL divergence), and point mutual information (PMI) distribution of adjacent words.
3. The AIGC text detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, In step S1, the text perplexity is calculated using lightweight GPT-2, the cosine similarity of adjacent word vectors is calculated using Word2Vec, and the average depth index of the syntactic tree is obtained through dependency analysis.
4. The AIGC text detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, The metadata mentioned in step S3 includes text length, domain, and style. The adaptive threshold is dynamically generated through a preset "metadata → threshold" mapping rule or function, so that different types of text, such as long technical documents and short social media comments, are judged by different criteria.
5. The AIGC text detection method based on multi-dimensional feature fusion according to claim 1, characterized in that, The feature contribution analysis described in step S4 quantifies and displays the contribution of each feature dimension and specific feature item to the final judgment, clearly informing the user of the main basis for the judgment.
6. An AIGC text detection system based on multi-dimensional feature fusion, characterized in that, include: Feature extraction module: used to extract feature vectors in parallel from four orthogonal dimensions of the text to be detected: statistical linguistics, semantics and fluency, syntactic structure and information theory, to construct a complementary multidimensional evidence set; Feature fusion module: It is used to perform Z-score standardization on feature vectors, learn the dynamic weights of each feature dimension through attention network and sum them up to generate a fusion discrimination score, avoiding scene adaptation problems caused by fixed weighting or simple concatenation. The discrimination module is used to generate an adaptive threshold based on text metadata, compare the fused discrimination score with the adaptive threshold, output the judgment result of "AI generated" or "human writing", and calculate the confidence level based on the difference between the score and the threshold. Report generation module: Used to output interpretable reports that include feature contribution analysis and highlight high-risk segments, improving the transparency and usability of the results.
7. The AIGC text detection system based on multi-dimensional feature fusion according to claim 6, characterized in that, It also includes an offline construction unit, which is used to determine / update human corpus benchmarks, train or calibrate attention networks with labeled data, form adaptive rules / mappings of "metadata → threshold" and perform threshold calibration in combination with a validation set.
8. The AIGC text detection system based on multi-dimensional feature fusion according to claim 6, characterized in that, It also includes an online inference unit, which is used to perform a real-time detection process for the text to be detected, including multi-dimensional feature extraction, attention-weighted feature fusion, adaptive threshold determination and confidence assessment, and interpretability report generation, supporting large-scale deployment and real-time response.