Man-machine mixed text detection method based on multilevel style preference optimization

Abstract

The invention particularly relates to a man-machine mixed text detection method based on multilevel style preference optimization, which comprises the following steps of: firstly, training a machine style scoring model through preference learning, carrying out joint modeling on three granularities of sequences, phrases and vocabularies by the model, and dynamically adjusting and optimizing parameters based on text complexity; therefore, style features of the machine text can be comprehensively captured and amplified. In a detection stage, a probability curvature index D-score of a to-be-detected text is calculated by utilizing the trained model, and a text source is judged according to comparison between the probability curvature index D-score and a threshold value. According to the method, the problems that an existing detection method is difficult to cope with man-machine mixed texts and the feature granularity is single are solved, and the detection precision and generalization ability of various man-machine mixed texts such as pure machine generation, machine rewriting and machine polishing are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of text detection technology, and specifically to a human-computer hybrid text detection method based on multi-level style preference optimization. Background Technology

[0002] Before the advent of large language models (LMs), traditional machine text generation methods (such as those based on recurrent neural networks) had limited capabilities. The text they generated was significantly inferior to human writing in terms of fluency, logic, and professionalism, and the machine traces were obvious and easy to identify. Therefore, the problem of machine text detection did not attract widespread attention.

[0003] Recent research on LLMs has yielded significant results, demonstrating remarkable performance and textual creativity across numerous tasks. Text generated by large language models has approached or even rivaled human levels in fluency, coherence, and knowledge content, greatly advancing content creation, academic research, and news reporting. However, this progress has also spurred new technological demands: Machine-Generated Text Detection (MGTD). MGTD aims to determine whether a given text was generated or modified (e.g., rewritten or polished) by large language models (LLMs), and to distinguish between purely human creation and varying degrees of human-machine collaboration. This technology has significant applications in areas such as academic integrity protection, content authenticity verification, copyright attribution determination, and automated content review. In particular, in broader practical applications, users do not solely rely on large language models for pure text generation; rather, they increasingly use them as auxiliary tools to "machine rewrite" (i.e., retaining the original meaning but significantly modifying the wording) or "machine polish" (i.e., locally optimizing sentences to improve fluency) existing human drafts, resulting in a large volume of "human-machine hybrid text."

[0004] The emergence of human-machine hybrid text detection presents a significant challenge to existing detection technologies, specifically: 1) Existing detection methods struggle to handle hybrid scenarios: Most existing methods are designed to distinguish between "purely human text" and "purely machine text," heavily relying on global statistical features of the text (such as word probability distribution or entropy). However, in human-machine hybrid text, local machine traces render these global features insignificant, drastically reducing performance. 2) Single-granularity feature extraction: Even some emerging detection methods extract features only at a single granularity (such as overall sequence or local semantics), lacking a multi-level comprehensive analysis framework. 3) Failure to capture multi-level style traces: Existing methods fail to effectively capture and integrate the multi-level style "fingerprints" left by machines at the word, phrase, and sequence levels, resulting in insufficient detection accuracy. 4) Lack of adaptability to text complexity: Existing methods often employ static strategies, lacking the ability to perceive the complexity of different texts (such as academic and news texts) and unable to dynamically adjust strategies based on differences in human and machine contributions, leading to poor generalization and insufficient robustness. Therefore, how to design a detection framework that can capture machine style preferences from multiple granularities (vocabulary, phrase, sequence) and dynamically adapt to text complexity in order to effectively identify human-machine mixed text of various degrees is a key technical problem that urgently needs to be solved in the field of text authenticity and security.

[0005] It should be noted that the information disclosed in the background section above is only used to enhance the understanding of the background of the present invention, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] To address the shortcomings of existing machine-generated text detection methods, such as single-granularity modeling and fixed parameter settings, this invention provides a human-machine hybrid text detection method based on multi-level style preference optimization.

[0007] Other features and advantages of the invention will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0008] According to a first aspect of the present invention, a human-computer hybrid text detection method based on multi-level style preference optimization is provided, the method comprising: Construct a training dataset containing human-machine text pairs, which are used to model machine style preference relationships in the text; A scoring model is constructed based on the Direct Preference Optimization (DPO) framework. The scoring model is based on the language model to be optimized and the reference model. The scoring function quantifies the machine generation probability of the text through a scoring function that satisfies the following condition: the score of machine text is higher than the score of the corresponding human text. Based on the training dataset, a multi-level style preference optimization training scoring model is performed, which jointly optimizes style features at three granular levels: sequence, phrase, and word. The overall optimization objective function of the training includes a sequence-level loss function, a phrase-level loss function, and a word-level loss function. The relative contribution of features at each level is controlled by hyperparameters. Parameters are dynamically adjusted and optimized based on text complexity. The text complexity is calculated by weighting four indicators: type-to-tag ratio, average sentence length, average word length, and punctuation ratio, and is adaptively adjusted based on the complexity differences between human and machine text pairs. parameter; The detection is performed based on the trained scoring model. The text type is determined by calculating the style probability curvature D-score of the text to be tested. When the D-score is greater than the judgment threshold, it is judged as machine-generated text; otherwise, it is judged as human text.

[0009] In some exemplary embodiments, the scoring function is expressed as:

[0010] in, For the language model to be optimized, As a reference model, Parameters used to control the magnitude of rewards.

[0011] In some exemplary embodiments, the overall optimization objective function is expressed as:

[0012] in, , , It captures style features at the sequence level, phrase level, and lexical level, respectively. and This is a hyperparameter that controls the relative contribution of features at each level to the overall optimization objective.

[0013] In some exemplary embodiments, the sequence-level loss function Represented as:

[0014] in, Represents machine-human text pairs. This is the sigmoid function.

[0015] In some exemplary embodiments, the phrase-level loss function Represented as:

[0016] in, This indicates the first [item] extracted from the text. i n-gram sequences, and These represent the joint probabilities of the n-gram under the optimized model and the reference model, respectively. For boundary parameters, This represents the total number of n-grams extracted from the text.

[0017] In some exemplary embodiments, the vocabulary-level loss function Represented as:

[0018] in, and Representing the positions of human text and machine text respectively. i prefix context, and This represents the probability distribution of the optimized model over the entire vocabulary under the given context. This represents the KL divergence.

[0019] In some exemplary embodiments, the adaptive adjustment based on the complexity differences of human-machine text pairs... The parameters are as follows:

[0020] in, As the baseline value, Control the maximum adjustment range. Control and adjust sensitivity.

[0021] According to a second aspect of the present invention, a storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the human-computer hybrid text detection method based on multi-level style preference optimization described in the first aspect.

[0022] According to a third aspect of the present invention, a computer program product is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, the human-computer hybrid text detection method based on multi-level style preference optimization described in the first aspect is implemented.

[0023] According to a fourth aspect of the present invention, an electronic device is provided, comprising: Processor; and Memory for storing the executable instructions of the processor; The processor is configured to implement the human-computer hybrid text detection method based on multi-level style preference optimization described in the first aspect above by executing the executable instructions.

[0024] The human-machine hybrid text detection method based on multi-level style preference optimization provided by the embodiments of the present invention addresses the shortcomings of existing machine-generated text detection methods, which rely on single-granularity modeling and fixed parameter settings. It proposes a detection framework based on multi-level style preference optimization. This framework jointly models the style features of machine text at three granularity levels: sequence, phrase, and vocabulary. It also dynamically adjusts optimization parameters based on text complexity, achieving accurate detection of various types of text, including purely machine-generated, machine-rewritten, and machine-polished text. Compared with traditional methods using single sequence-level modeling, the present invention can more comprehensively capture the style differences in machine text, significantly improving the detection performance of human-machine hybrid text. Compared with the prior art, the present invention has the following beneficial effects: 1. This invention innovatively proposes a multi-level style preference optimization framework, which captures global probability distribution at the sequence level, n-gram combination patterns at the phrase level, and word selection differences at the lexical level. By jointly optimizing style features at these three levels, this invention comprehensively identifies machine text traces from macro to micro perspectives, thereby enhancing the model's detection capabilities in human-machine hybrid scenarios. 2. This invention designs a dynamic parameter adjustment mechanism based on text complexity. This mechanism does not use the traditional fixed parameter method, but adaptively adjusts and optimizes parameters by comprehensively considering indicators such as type-to-tag ratio, average sentence length, average word length and punctuation ratio, which can help the model maintain stable performance on diverse text types. 3. This invention models the detection task as a preference learning problem, trains a scoring model to "preference" machine text features, and combines probability curvature detection to achieve unsupervised judgment, thereby enhancing the generalization ability and scalability of the method. 4. Extensive experiments have been conducted on multiple representative fields and mainstream large-scale language models. The experimental results confirm the effectiveness and superiority of the invention, especially demonstrating significant advantages in human-machine hybrid text detection scenarios. It should be understood that the above general description and the following detailed description are merely exemplary and explanatory, and do not limit the invention. Attached Figure Description

[0025] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0026] Figure 1 This is the architecture diagram of the present invention - the multi-level style preference optimization stage; Figure 2This is the architecture diagram of the present invention - style probability curvature detection stage; Figure 3 This is a flowchart of the method of the present invention. Detailed Implementation

[0027] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the invention will be more comprehensive and complete, and will fully convey the concept of the exemplary embodiments to those skilled in the art. The described features, structures, or characteristics may be combined in any suitable manner in one or more embodiments.

[0028] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0029] This invention proposes a human-machine hybrid text detection method based on multi-level style preference optimization, called the MSPO (Multi-level Style Preference Optimization) framework. This method models the style features of machine text at three granular levels: sequence, phrase, and vocabulary, and dynamically adjusts optimization parameters based on text complexity to achieve accurate detection of various levels of human-machine collaborative output, including purely machine-generated, machine-rewritten, and machine-polished text. Figure 1 and Figure 2 As shown, the method of this invention provides a complete end-to-end human-machine hybrid text detection technology solution. Logically, this solution constitutes a complete technical closed loop from data construction and model optimization to inference detection, mainly including three core implementation stages: 1) Data preparation stage, constructing human-machine text pairs containing multiple types (such as rewriting and polishing) to provide a data foundation for preference learning; 2) Multi-level style preference optimization training stage (such as... Figure 1 As shown), this stage abandons the traditional binary classification paradigm and, through joint modeling at three granularities—sequence, phrase, and word—and combined with a dynamic parameter adjustment mechanism based on text complexity, collaboratively optimizes a machine style scoring model; 3) the detection stage based on style probability curvature (such as... Figure 2 As shown in the figure, the trained scoring model is used to calculate the style probability curvature of the text under test to determine its origin. The core innovation of the entire framework lies in the synergistic optimization of multi-granularity style modeling and dynamic parameter adjustment.

[0030] like Figure 3 As shown, the specific steps may include: Step S11: Construct a training dataset containing human-machine text pairs, which are used to model the machine style preference relationship of the text; Step S12: Construct a scoring model based on the Direct Preference Optimization (DPO) framework. The scoring model is based on the language model to be optimized and the reference model. It quantifies the machine generation probability of the text through a scoring function. The scoring function satisfies the following condition: the score of machine text is higher than the score of the corresponding human text. Step S13: Based on the training dataset, perform multi-level style preference optimization training of the scoring model, jointly optimize style features at three granular levels: sequence, phrase, and word. The overall optimization objective function of the training includes a sequence-level loss function, a phrase-level loss function, and a word-level loss function. The relative contribution of each level of features is controlled by hyperparameters. Step S14: Dynamically adjust and optimize parameters based on text complexity. The text complexity is calculated by weighting four indicators: type-to-tag ratio, average sentence length, average word length, and punctuation ratio, and is adaptively adjusted based on the complexity differences between human and machine text pairs. parameter; Step S15: Detection is performed based on the trained scoring model. The text type is determined by calculating the style probability curvature D-score of the text to be tested. When the D-score is greater than the determination threshold, it is determined to be machine-generated text; otherwise, it is determined to be human text.

[0031] The steps in this exemplary embodiment will now be described in more detail with reference to the accompanying drawings and embodiments.

[0032] In step S11, a training dataset D containing human-machine text pairs is constructed, which are used to model the machine style preference relationship of the text.

[0033] In step S12, a scoring model is constructed based on the Direct Preference Optimization (DPO) framework. This scoring model is based on the language model to be optimized and a reference model, and quantifies the machine-generated probability of the text through a scoring function. This scoring function satisfies the condition that machine-generated text receives a higher score than corresponding human-generated text. Specifically: Given a text sequence ,in This represents the i-th word in the text, and the detection task aims to determine whether it is human text. Machine text This invention remodels this problem as a preference learning task and trains a scoring model. The probability of machine-generated text is quantified. The model optimization objective is to satisfy the following preference relationship: (1) A higher score indicates stronger machine style features. Based on the above preference relationship, the optimization objective is to maximize the expected score difference: (2) Here, D represents the training dataset containing human-machine text pairs. The objective function ensures that the machine text consistently scores higher than the corresponding human text on all training samples, thereby learning stable and distinguishable style feature representations.

[0034] Following the Direct Preference Optimization (DPO) framework, this invention is based on a language model. Constructing a scoring function : (3) in For the language model to be optimized, Use the reference model (usually the initial model before optimization). This parameter controls the magnitude of the reward. The scoring function measures the improvement in the log probability of the text by the optimized model relative to the reference model. When the model's ability to model machine text is enhanced after optimization, the machine text will naturally score higher than the human text.

[0035] In step S13, multi-level style preference optimization training is performed, jointly optimizing style features at three granular levels: sequence, phrase, and vocabulary. The overall optimization objective function of the training includes a sequence-level loss function, a phrase-level loss function, and a vocabulary-level loss function, and the relative contribution of features at each level is controlled by hyperparameters; specifically as follows: Traditional style detection methods primarily perform overall probability modeling at the sequence level, neglecting fine-grained style patterns at the local unit level (such as phrase combinations and word selection). To address this issue, this invention proposes a multi-level style preference optimization method that captures style differences at three complementary granularities: sequence, phrase, and word. The overall optimization objective function is expressed as: (4) in , , It captures style features at the sequence level, phrase level, and lexical level, respectively. and The hyperparameter controls the relative contribution of features at each level to the overall optimization objective. By jointly optimizing the loss functions at the three levels, the model can comprehensively learn the style features of machine text from macro to micro levels.

[0036] For example, sequence-level style optimization specifically involves: Sequence-level optimization focuses on the overall probability distribution and global style consistency of the text. Using the Bradley-Terry model, rating differences are converted into preference probabilities, resulting in the sequence-level loss function: (5) in, Represents machine-human text pairs. This is the sigmoid function. The meaning of this loss function is: optimizing the model. machine text The log-probability improvement (relative to the reference model) should be significantly greater than that for human text. The improvement is achieved by minimizing this loss, allowing the model to learn overall style patterns of machine text at the sequence level, such as higher fluency and more regular syntactic structures. Specifically, for text... Its logarithmic probability is: (6) in, This represents the context preceding position i. This sequence-level modeling can capture global probabilistic features of the text, such as overall fluency and syntactic regularity. For example, phrase-level style optimization specifically involves: Phrase-level optimization focuses on n-gram combination patterns, capturing the consistency differences in local phrase structures. Machine-generated text often exhibits stronger regularity and repetitive patterns in local phrase combinations. This invention optimizes phrase-level features through the following hinge loss: (7) in, This represents the i-th n-gram sequence extracted from the text. and These represent the joint probabilities of the n-gram under the optimized model and the reference model, respectively. Here, is the boundary parameter, representing the total number of n-grams extracted from the text. Specifically, for an n-gram sequence... The joint probability is calculated as follows: (8) This hinge loss encourages the optimization model to assign a higher probability to common n-gram patterns relative to the reference model when processing machine text, thereby amplifying the stylistic differences between human and machine text at the phrase level. For example, machine-edited text tends to use certain fixed phrase combinations (such as "in order to" instead of "to", "make useof" instead of "use", etc.), and these local patterns can be effectively captured under phrase-level constraints.

[0037] For example, lexical-level style optimization specifically involves: Lexical-level optimization measures the distributional differences in word selection using the KL divergence at each token position: (9) in, and Represent the prefix context of human text and machine text at position i, respectively. and This represents the probability distribution of the optimized model over the entire vocabulary under the given context. The formula for calculating the KL divergence is: (10) Here, V represents the vocabulary. This constraint prevents the model from overfitting to style patterns specific to a particular context, ensuring that the optimized model maintains a relatively consistent output distribution under different contextual conditions, thereby improving generalization ability. For example, machine-rewritten text exhibits stronger consistency in word selection (such as a preference for high-frequency words), and vocabulary-level constraints can capture this subtle difference.

[0038] In step S14, the optimization parameters are dynamically adjusted based on the text complexity. The text complexity is calculated by weighting four indicators: type-to-tag ratio, average sentence length, average word length, and punctuation ratio, and is adaptively adjusted based on the complexity differences between human and machine text pairs. Parameters; details are as follows: In the DPO framework, parameters The deviation of the optimization model from the reference model is controlled. Traditional methods use static... The values ​​do not take into account the differences in complexity between human and machine text across different samples. This invention introduces dynamic... An adjustment mechanism adaptively adjusts the optimization intensity based on changes in text complexity.

[0039] This invention uses four metrics to measure text complexity: (1) Type-to-Tagged Ratio (TTR): measures lexical diversity, TTR(X) = number of unique lexical units / total number of lexical units; (2) Average Sentence Length (ASL): measures syntactic complexity, ASL(X) = total number of lexical units / number of sentences; (3) Average Word Length (AWL): measures lexical complexity. (4) Punctuation Ratio (PR): measures the degree of structuring, PR(X) = number of punctuation marks / total number of morphemes.

[0040] The overall complexity score is calculated by weighted summation: (11) in, , , , These are the weighting coefficients for each indicator. The β parameter is dynamically adjusted based on the complexity differences between human and machine text pairs. (12) in, As the baseline value, Control the maximum adjustment range. Controlling the sensitivity of the adjustment. This formula is adjusted through differences in relative complexity. Increase when human text becomes more complex When the two are close, the decrease This allows it to adapt to the characteristics of different samples.

[0041] In step S15, detection is performed based on the trained scoring model. The text type is determined by calculating the style probability curvature D-score of the text to be tested. When the D-score is greater than the determination threshold, it is determined to be machine-generated text; otherwise, it is determined to be human-generated text. The details are as follows: Complete testing process as follows Figure 2 As shown. After multi-level style preference optimization training, a machine style scoring model is obtained. Used for detection. This invention employs the D-score metric based on probability curvature for judgment. For the text to be detected... The D-score measures the degree of deviation of the observed log-likelihood from the expected log-likelihood. (13) The expected mean and standard deviation are calculated as follows: , Due to the optimization model Machine-generated patterns are assigned higher probabilities; the actual log-likelihood of machine text tends to be higher than the expected value (positive D-score), while human text tends to be lower than or close to the expected value (negative or small D-score). Based on the D-score, classification is performed as follows: (14) Where f(x)=1 indicates that the text is determined to be machine-generated. The threshold can be determined through validation set optimization. This is achieved through multi-level style modeling (sequence-level, phrase-level, lexical-level) and dynamic parameter optimization (adaptive based on text complexity). By combining with probability curvature detection (D-score), this invention achieves accurate detection of various human-computer hybrid texts, including pure generation, rewriting, and polishing, while significantly improving detection performance while maintaining cross-domain generalization capabilities.

[0042] The effects of the present invention are further illustrated by the following simulation experiments: Please provide a brief explanation based on Tables 1 and 2.

[0043] To verify the effectiveness of this invention, we conducted comparative experiments in computer science (CS), health and social sciences (HSS), and physics (PHX) (as shown in Table 1). The results show that this invention maintains a high accuracy of over 98% on pure machine text; it demonstrates significant advantages in more challenging human-machine hybrid scenarios, particularly in machine rewriting tasks, where the AUROC value reaches 80%–82%, significantly outperforming the current best baseline model (approximately 71%), and the average AUROC in machine polishing tasks also exceeds 96%.

[0044] Furthermore, ablation experiments on CS domain datasets (as shown in Table 2) further confirm the necessity of each module. Compared to the full model (97.44%), removing vocabulary-level, phrase-level constraints, or dynamic parameter adjustment mechanisms all led to performance degradation, while retaining only sequence-level optimization yielded the worst performance at only 91.75%. This fully demonstrates that the synergistic effect of multi-level style modeling and dynamic parameter adjustment is key to achieving high-precision detection.

[0045] Table 1. Performance comparison of the present invention and the baseline model on the XSUM dataset.

[0046] Table 2. Performance comparison of different modules for text ablation on the CS dataset using this invention.

[0047] Furthermore, the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Additionally, it is readily understood that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0048] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims.

[0049] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is defined only by the appended claims.

Claims

1. A human-computer hybrid text detection method based on multi-level style preference optimization, characterized in that, The method includes: Construct a training dataset containing human-machine text pairs, which are used to model machine style preference relationships in the text; A scoring model is constructed based on the Direct Preference Optimization (DPO) framework. The scoring model is based on the language model to be optimized and the reference model. The scoring function quantifies the machine generation probability of the text through a scoring function that satisfies the following condition: the score of machine text is higher than the score of the corresponding human text. Based on the training dataset, a multi-level style preference optimization training scoring model is performed, which jointly optimizes style features at three granular levels: sequence, phrase, and word. The overall optimization objective function of the training includes a sequence-level loss function, a phrase-level loss function, and a word-level loss function. The relative contribution of features at each level is controlled by hyperparameters. Parameters are dynamically adjusted and optimized based on text complexity. The text complexity is calculated by weighting four indicators: type-to-tag ratio, average sentence length, average word length, and punctuation ratio, and is adaptively adjusted based on the complexity differences between human and machine text pairs. parameter; The detection is performed based on the trained scoring model. The text type is determined by calculating the style probability curvature D-score of the text to be tested. When the D-score is greater than the judgment threshold, it is judged as machine-generated text; otherwise, it is judged as human text.

2. The method according to claim 1, characterized in that, The scoring function is expressed as follows: in, For the language model to be optimized, As a reference model, Parameters used to control the magnitude of rewards.

3. The method according to claim 1, characterized in that, The overall optimization objective function is expressed as: in, , , It captures style features at the sequence level, phrase level, and lexical level, respectively. and This is a hyperparameter that controls the relative contribution of features at each level to the overall optimization objective.

4. The method according to claim 3, characterized in that, The sequence-level loss function Represented as: in, Represents machine-human text pairs. This is the sigmoid function.

5. The method according to claim 3, characterized in that, The phrase-level loss function Represented as: in, This indicates the first [item] extracted from the text. i n-gram sequences, and These represent the joint probabilities of the n-gram under the optimized model and the reference model, respectively. For boundary parameters, This represents the total number of n-grams extracted from the text.

6. The method according to claim 3, characterized in that, The vocabulary-level loss function Represented as: in, and Representing the positions of human text and machine text respectively. i prefix context, and This represents the probability distribution of the optimized model over the entire vocabulary under the given context. This represents the KL divergence.

7. The method according to claim 1, characterized in that, The adaptive adjustment based on the complexity difference between human and machine text pairs The parameters are as follows: in, As the baseline value, Control the maximum adjustment range. Control and adjust sensitivity.

8. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the human-computer hybrid text detection method based on multi-level style preference optimization as described in any one of claims 1 to 7.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the human-computer hybrid text detection method based on multi-level style preference optimization as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, include: processor; as well as Memory for storing the executable instructions of the processor; The processor is configured to execute the human-computer hybrid text detection method based on multi-level style preference optimization as described in any one of claims 1 to 7 by executing the executable instructions.

Images