An ai-generated text detection method and system based on a deep learning model

Abstract

The application relates to the technical field of natural language processing, in particular to an AI generated text detection method and system based on a deep learning model. The method comprises the following steps: performing character normalization and sentence segmentation on a to-be-detected text, segmenting the text by a sliding window to generate a segmented sequence; constructing deep semantic discrimination features, logarithmic probability difference trace features, style structure features and entity consistency features; inputting a gate fusion discrimination network, outputting a generated probability and confidence degree after probability calibration, and outputting a segmented contribution degree; performing threshold adaptive determination according to a risk level, a text length and the confidence degree, outputting a review state and generating a traceable record when the confidence degree is insufficient; and triggering incremental updating based on drift monitoring and a feedback sample pool. The application improves cross-domain robustness and interpretability, reduces false positives and false negatives, and supports long-term stable operation.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology. More specifically, this application relates to an AI-generated text detection method and system based on a deep learning model. Background Technology

[0002] AI-generated text is widely used in content creation, customer service Q&A, search enhancement, and office writing. Detecting this type of text is a crucial foundation for content security and academic integrity. By determining whether input text is AI-generated, a basis can be provided for manual review, automatic interception, annotation and prompting, and risk management.

[0003] Existing detection methods typically employ deep learning text classification models, encoding the input text into vector representations and outputting binary classification results. To improve performance, some methods fuse statistical features such as sentence length, punctuation distribution, word distribution, and repetition. These methods are effective within the same domain and with similar generative models, but they are prone to fluctuations in false positives and false negatives in cross-domain and cross-model scenarios.

[0004] Another approach utilizes language models to construct generation trace indicators based on the probability distribution features of text, such as statistical log probability, the proportion of repeated segments, and information density. Other approaches enhance the ability to distinguish machine-generated text from human-generated text through training strategies such as contrastive learning. However, because AI-generated text can circumvent fixed features through synonym rewriting, translation rewriting, structural perturbation, and prompt-based writing constraints, the aforementioned approaches still have shortcomings in adversarial evasion and long text structure discrimination.

[0005] Meanwhile, existing solutions often only output scores or labels, lacking confidence level explanations and evidence fragments, making the results difficult to audit and verify. Fixed thresholds are difficult to adapt to business scenarios with different risk levels and text lengths, easily leading to high online false positive costs or high false negative risks. The lack of feedback loops and drift monitoring mechanisms also results in insufficient long-term model stability. Therefore, there is a need for a detection method and system that can remain stable under cross-domain and adversarial rewriting conditions, output confidence levels and evidence, and support threshold adaptation and continuous updates. Summary of the Invention

[0006] The purpose of this application is to propose an AI-generated text detection method and system based on a deep learning model, in order to solve the problems in existing technologies that are difficult to guarantee detection accuracy and robustness, and lack confidence calibration and evidence output.

[0007] In a first aspect, this application provides an AI-generated text detection method based on a deep learning model, comprising: acquiring a text to be detected; performing character normalization, sentence segmentation, and sliding window segmentation on the text to be detected to generate a segmented text sequence, and generating index information corresponding one-to-one with the segmented text sequence; constructing multi-view detection features for the segmented text sequence, the multi-view detection features including at least: deep semantic discriminative features output by a pre-trained text encoder, generation trace features obtained from the log probability difference of a reference language model, style structure features obtained from text statistical indicators, and entity consistency features obtained from entity transfer relations; inputting the multi-view detection features into a gated fusion discriminative network, outputting a probability-calibrated generation probability and confidence level, and outputting a segment contribution; acquiring the risk level of the text to be detected, determining an adaptive discrimination threshold based on the generation probability, the confidence level, the text length, and the risk level, outputting a detection result indicating whether the text to be detected is AI-generated text or human-generated text, and generating evidence fragments based on the segment contribution and index information.

[0008] This application improves detection accuracy and robustness under adversarial rewriting and cross-domain conditions by fusing multi-perspective features and combining probability calibration and threshold adaptive strategies, while outputting evidence fragments for auditing and review.

[0009] Optionally, the character normalization, sentence segmentation, and sliding window segmentation include: performing full-width / half-width conversion, whitespace normalization, abnormal control character filtering, and repeated line breaks on the text to be detected to obtain normalized text; segmenting the normalized text into sentences and segments based on paragraph boundaries and punctuation boundaries; when the length of a certain segment exceeds a preset upper limit, segmenting the segment using an overlapping sliding window with an overlap ratio of 10% to 30% to generate the segmented text sequence.

[0010] Optionally, the construction of the generation trace features includes: calculating the log probability of each segment of text in the segmented text sequence under the reference language model; generating K perturbed texts for the segmented text, where K is an integer greater than or equal to three, and the perturbing method includes random mask replacement or synonym replacement, with the masking ratio of random mask replacement being 10% to 20%; calculating the log probability of each perturbed text under the reference language model, and calculating the difference in log probability between the segmented text and the corresponding perturbed text, where the reference language model adopts an open-source language model; and statistically analyzing the difference in log probability to obtain the mean, variance, and quantiles as the generation trace features.

[0011] By constructing trace features through perturbation-sensitive statistics, the detection becomes more stable in cross-model scenarios and reduces false positives caused by relying solely on semantic discriminators.

[0012] Optionally, the construction of the style structure features includes: statistically analyzing the mean sentence length, standard deviation of sentence length, distribution of punctuation marks, and lexical richness; statistically analyzing the proportion of repeated segments, which includes the proportion of overlapping phrases between adjacent sentences and the proportion of high-frequency segments within the text; statistically analyzing paragraph structure consistency, which includes at least the distribution of paragraph length and the distribution of inter-sentence connecting words; and combining the above statistical results into a style feature vector as the style structure features.

[0013] Optionally, the construction of the entity consistency feature includes: performing entity recognition and referential clue extraction on the segmented text sequence to obtain an entity set and entity category; constructing an entity transition graph based on the occurrence order and co-occurrence relationship of entities in adjacent sentences or adjacent paragraphs; performing graph convolutional encoding or transition probability statistics on the entity transition graph to obtain an entity consistency vector as the entity consistency feature.

[0014] Optionally, the gated fusion discriminant network includes a multi-branch encoding layer, a gated weight layer, and a classification output layer. The multi-branch encoding layer performs nonlinear mapping on deep semantic discriminant features, generation trace features, style structure features, and entity consistency features to obtain branch representations with alignment dimensions. The gated weight layer calculates fusion weights based on text length and domain indicator features, whereby the domain indicator features are determined by the source type of the text to be detected, the title keyword matching results, or the domain labels output by the topic classification model. The classification output layer classifies the fused representations and outputs the generation probabilities.

[0015] By adaptively allocating the contribution of each branch through gating weights, the performance fluctuations caused by the dominance of a single branch during domain offset are reduced.

[0016] Optionally, the probability calibration and confidence calculation include: learning probability calibration parameters on a calibration dataset so that the calibrated generation probability matches the actual occurrence frequency; performing temperature scaling or segmented calibration on the generation probability to obtain the calibrated generation probability; and calculating the confidence based on the calibrated generation probability and historical reliability statistics.

[0017] Optionally, the adaptive discrimination threshold determination includes: selecting a basic threshold based on the risk level, wherein the risk level is determined in a preset mapping table based on the business scenario or text source of the text to be detected, and the risk level includes at least low risk, medium risk, and high risk; dynamically correcting the basic threshold based on the text length and confidence level to obtain the adaptive discrimination threshold; and when the confidence level is lower than a preset lower limit, outputting a need for review status and generating a prompt message to supplement the context or increase the text length.

[0018] By using risk level-driven threshold adaptive judgment, the cost of false alarms and the risk of false alarms are balanced across different business scenarios.

[0019] Optionally, the method further includes incremental update and drift monitoring steps: generating a traceable record for each detection, the traceable record including at least a text hash value, segment index, risk level, and evidence fragment; receiving manual review results or business feedback tags and forming a feedback sample pool; calculating the difference between the new data distribution and the training data distribution on preset statistical features, the difference including a weighted sum of Jensen-Shannon divergence and multi-quantile offset measure; triggering a drift alarm when the difference exceeds a preset threshold, and using the feedback sample pool to perform small-step fine-tuning or comparative learning updates on the parameters of the gating fusion discrimination network.

[0020] By using drift monitoring and incremental update closed loop, the model can be kept running stably for a long time, reducing failures caused by decay over time.

[0021] In the second aspect, an AI-generated text detection system based on a deep learning model includes:

[0022] processor;

[0023] The memory stores computer instructions for AI-generated text detection, which, when executed by the processor, cause the system to perform the aforementioned AI-generated text detection method based on a deep learning model.

[0024] The beneficial effects of this application are: by multi-view deep feature fusion, probability calibration, threshold adaptation, evidence output and closed-loop update, cross-domain robustness and interpretability are improved, false positives and false negatives are reduced and long-term stable operation is supported. Attached Figure Description

[0025] Figure 1 This is a flowchart of an AI-generated text detection method based on a deep learning model, according to an embodiment of this application.

[0026] Figure 2 This is a multi-view detection feature response heatmap of an AI-generated text detection method based on a deep learning model, according to an embodiment of this application.

[0027] Figure 3 This is a distribution difference measurement and drift alarm diagram of an AI-generated text detection method based on a deep learning model according to an embodiment of this application.

[0028] Figure 4 This is a performance comparison chart of an AI-generated text detection method based on a deep learning model, according to an embodiment of this application.

[0029] Figure 5 This is a structural block diagram of an AI-generated text detection system based on a deep learning model, according to an embodiment of this application. Detailed Implementation

[0030] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Figure 1 The diagram shown is a flowchart of an AI-generated text detection method based on a deep learning model, according to an embodiment of this application.

[0031] S1: Obtain the text to be detected and perform normalization and segmentation modeling.

[0032] In this embodiment, the text to be detected can be obtained from the content acquisition terminal, the business interface, or user input. After receiving the text to be detected, character normalization processing is first performed to eliminate the impact of encoding differences from different sources on subsequent feature statistics. Character normalization processing includes full-width and half-width conversion, continuous whitespace normalization, invisible control character filtering, cleaning of common web page tags and footnote symbols, and merging of repeated line breaks to obtain normalized text.

[0033] To ensure stable sentence boundaries, a set of punctuation marks is constructed, and symbols such as ellipses, book titles, and parentheses are merged according to rules to avoid splitting consecutive punctuation marks into multiple sentence boundaries. At the same time, patterns such as numerical serial numbers and citation numbers are protected to prevent structural symbols such as Chapter X and a dot from being misidentified as the end of a sentence.

[0034] When segmenting standardized text into sentences and paragraphs, paragraph line breaks are used as coarse-grained boundaries first, and then punctuation marks such as periods, question marks, exclamation marks, semicolons, and colons are used as fine-grained boundaries within paragraphs to generate sentence sequences and paragraph sequences. For colloquial text lacking punctuation, maximum length constraints and conjunction heuristics are used to supplement the segmentation, reducing the risk of truncation of model input due to excessively long single paragraphs.

[0035] Considering the upper limit of the input length of the deep semantic encoder and the reference language model, this embodiment uses an overlapping sliding window to segment each paragraph to generate segmented text. The upper limit of the segment length is set to 512 symbols, the overlap ratio is set to 20%, and the sliding window step size is adaptively adjusted according to the actual length. When the paragraph length is less than the upper limit, it is directly treated as a segmented text. When the paragraph length significantly exceeds the upper limit, the number of segments increases linearly with the length, thereby ensuring that the overall information of the long text is covered.

[0036] To support evidence fragment tracing and audit trail recording, index information is generated for each segment of text. The index information includes at least the start and end positions of the original text, the paragraph number, the sentence number, and the segment number. At the same time, a text hash value is calculated for the normalized text to achieve record consistency verification and deduplication statistics without saving the original text.

[0037] For very short texts, texts containing only titles or only citation lists, this embodiment marks them as low-information texts and performs a verification strategy in conjunction with confidence in the subsequent discrimination stage; for texts containing mixed languages, language recognition is performed first and modeling is done separately for each language segment, and then the segmentation results are merged to reduce the error caused by the interference of mixed languages ​​in the probability statistics of the language model.

[0038] S2: Construct deep semantic discriminative features and multi-view detection features such as generated traces.

[0039] After obtaining the segmented text sequence, the multi-view detection feature construction stage begins. To reduce the risk of adversarial evasion of a single feature, this embodiment simultaneously constructs deep semantic discriminative features, generation trace features, style structure features, and entity consistency features, and aligns them at the same segment granularity to form a fusionable feature tensor.

[0040] The construction process of deep semantic discriminative features is as follows: Segmented text is input into a pre-trained text encoder to obtain segment-level and sentence-level vectors for each segment; attention aggregation is applied to the sentence-level vectors to obtain segmented semantic representations; and weighted pooling is then applied to each segmented semantic representation to obtain a text-level semantic representation. The text-level semantic representation is used to characterize differences in logical coherence, information density, and semantic expansion methods between sentences.

[0041] The process of constructing generation trace features is as follows: A reference language model is selected to calculate the log probability for each segment of text, and K perturbation texts are generated for that segment. The generation of perturbation texts uses random mask replacement, with a mask ratio set to 10% to 20%. The preferred replacement strategy is to replace only content words while keeping punctuation and numbers unchanged to reduce the degree to which the perturbation disrupts semantics. The log probability difference between the original segment and the perturbation text is calculated, and the mean, variance, and quantiles of the difference are statistically analyzed as generation trace features.

[0042] To enhance resistance to rewriting avoidance, the degree of repetition and template-based structure within each segment can be further analyzed. Repetition indicators include the phrase overlap rate between adjacent sentences, the proportion of high-frequency segments within a segment, and the repetition rate of sentences beginning with the same structure; the degree of template-based structure is measured by the concentration of conjunction distribution and the excessive concentration of sentence length distribution.

[0043] The process of constructing stylistic structural features involves: statistically analyzing the mean sentence length, standard deviation of sentence length, distribution of punctuation categories, lexical richness, and paragraph structural consistency. Specifically, the distribution of punctuation categories reflects differences in writing structures such as narration, enumeration, and transition; lexical richness is characterized by the ratio of the number of different word forms to the total number of word forms; and paragraph structural consistency is characterized by the distribution of paragraph length and the distribution of inter-sentence connectors, thereby capturing the stability differences in the structure of long texts.

[0044] The construction process of entity consistency features is as follows: Entity recognition and referential clue extraction are performed on segmented text to obtain entities and entity categories; an entity transition graph is constructed based on the order of occurrence and co-occurrence relationships of entities in adjacent sentences and paragraphs; graph convolutional encoding or transition probability statistics are performed on the entity transition graph to obtain entity consistency vectors. Entity consistency vectors can reflect the structural patterns of entity introduction, expansion, and anaphora, and can form complementary evidence with style structure features.

[0045] Optionally, to reduce the risk of adversarial attacks against the reference language model, multiple reference language models are used to compute and generate trace features in parallel, and the results of multiple models are subject to consistency constraints. When there are significant conflicts in the generated trace indicators of different reference models, the generation probability output confidence is reduced and the level of review required is increased.

[0046] Thus, we obtain the aligned multi-view detection feature set. Each segment of text corresponds to a set of feature vectors, and the text-level features are obtained through segment aggregation, providing input for subsequent gating fusion discrimination.

[0047] like Figure 2 The image shows a heatmap of multi-view detection feature responses for an AI-generated text detection method based on a deep learning model, according to an embodiment of this application. The AI-generated samples and manually written samples exhibit significant complementary differences in deep semantic features, generation trace features, style and structural features, and entity consistency features, verifying the necessity of multi-view feature fusion.

[0048] S3: Perform gated fusion discrimination and probability calibration and generate evidence fragments.

[0049] In the fusion discrimination stage, multi-view detection features are input into the gated fusion discrimination network. The gated fusion discrimination network includes a multi-branch coding layer, a gated weight layer, and a classification output layer. The multi-branch coding layer performs nonlinear mapping on each type of feature, making them comparable and weighted in the same latent space.

[0050] The gated weight layer calculates the fusion weights based on text length, domain indicator features, and the internal confidence information of each branch. Domain indicator features are feature vectors representing the domain to which the text to be detected belongs. In this embodiment, the domain label is preferably encoded as a one-hot vector or a probability vector, wherein the domain label is determined by at least one of the following: parsing the source type metadata of the text to be detected, matching the title keywords with a preset domain keyword dictionary, and inputting the text to be detected into a topic classification model to obtain the domain category; when the results from multiple sources are inconsistent, the final domain label is obtained by fusing according to a preset priority or confidence level; when the text belongs to a specific domain and the style structure features are relatively weak in that domain, the gated weight layer reduces the weight of the style branch and increases the weight of the semantic branch and the generation trace branch to reduce false positives caused by domain shift.

[0051] The classification output layer classifies the fused latent representation to obtain the initial generation probability. To ensure that the generation probability has interpretable confidence meaning, this embodiment learns probability calibration parameters on the calibration dataset and performs temperature scaling or segmented calibration on the initial generation probability to obtain the calibrated generation probability; at the same time, confidence scores are generated based on calibration error statistics for subsequent threshold adaptive strategies.

[0052] During the training phase, the gated fusion discriminative network employs cross-entropy loss for supervised learning and combines it with contrastive learning loss to enhance cross-model transferability. Contrastive learning samples consist of positive sample pairs made up of rewritten texts with the same semantics and negative sample pairs made up of texts with different semantics, thereby encouraging the model to learn more robust discriminative representations to writing style perturbations.

[0053] To generate evidence fragments, contribution is calculated at the segment level, determined jointly by attention weights and occlusion sensitivity analysis. The occlusion sensitivity analysis process involves partially occluding a segment and recalculating its generation probability. The greater the change in generation probability, the more critical the segment is to the judgment.

[0054] The segments with the highest contribution are mapped back to the index information, outputting a set of candidate evidence segments. Each evidence segment is accompanied by a corresponding statistical summary of generation traces and a style structure summary, forming an auditable basis for judgment and supporting rapid manual location of abnormal segments.

[0055] To control inference costs, this embodiment caches and reuses the calculation of the reference language model: when multiple texts share the same segmentation or the same perturbation strategy, the intermediate probability results are reused; at the same time, a segmentation priority strategy is adopted for extremely long texts, and only segments with high contribution are calculated to generate complete trace features, thereby reducing the overall latency while ensuring accuracy.

[0056] S4: Perform threshold adaptive determination and drift monitoring and incremental updates.

[0057] During the strategy determination phase, an adaptive threshold determination is performed based on the risk level, text length, and confidence level. The risk level is configured by the business scenario and determined by a preset mapping table, which maps the source channel or business scenario of the text to be detected to low-risk, medium-risk, and high-risk levels. The base threshold is monotonically adjusted with the risk level, with a lower base threshold used in high-risk scenarios to reduce false negatives and a higher base threshold used in low-risk scenarios to reduce false positives.

[0058] A dynamic correction term is introduced above the base threshold. The dynamic correction term is determined by both the text length and the confidence level. The shorter the text, the more conservative the correction term. The lower the confidence level, the higher the threshold or the direct switch to a state requiring review, in order to avoid giving incorrect definitive conclusions under conditions of insufficient information.

[0059] When the calibrated generation probability is not lower than the adaptive discrimination threshold, the output is a detection result of suspected AI-generated text; when the calibrated generation probability is lower than the adaptive discrimination threshold, the output is a detection result of suspected human-generated text; when the confidence level is lower than the preset lower limit or the input is low-information text, the output is a status requiring review, and a prompt is made to supplement the context or merge more paragraphs before detection.

[0060] A traceable record is generated for each detection. This traceable record includes at least the text hash value, segment index, calibrated generation probability, confidence level, risk level, base threshold, dynamically adjusted parameters, and a set of evidence fragments. The traceable record supports auditing and statistics without storing the original text and can be used for feedback sample retrieval and model iteration.

[0061] To achieve long-term stable operation, drift monitoring is performed periodically. Drift monitoring is based on the feature distribution difference D, whereby the feature includes at least the calibrated generation probability distribution, generation trace statistical features, or the intermediate confidence distribution of the fused discriminant network.

[0062] Difference The calculation formula is as follows:

[0063] ;

[0064] in, For the degree of difference, The Jensen-Shannon divergence between the reference distribution P and the current distribution Q is used to measure the similarity between the two probability distributions. It is a multi-quantile offset metric used to measure the degree of numerical shift at specific quantiles; and The weight coefficients are non-negative and satisfy the following conditions: .

[0065] The Jensen-Shannon divergence The calculation formula is as follows:

[0066] ;

[0067] in, For Jensen-Shannon divergence, and represents the feature histogram probability distributions of the reference distribution and the current distribution obtained under the same binning rule; i is the binning index, and ln is the natural logarithm. Let be the average probability of two distributions located at the i-th bin, i.e. .

[0068] The calculation of the multi-quantile offset metric includes: calculating the quantile values ​​of the current distribution and the reference distribution on a preset quantile set, and taking the average of the absolute differences of each quantile as the multi-quantile offset metric; the quantile set includes at least tenths, fiftyths, and ninetieths.

[0069] When the difference D exceeds a preset threshold, a drift alarm is triggered, and more recent samples are prioritized for inclusion in the review and annotation queue. Figure 3 The diagram shown is a distribution difference measurement and drift warning diagram of an AI-generated text detection method based on a deep learning model according to an embodiment of this application.

[0070] The incremental update process uses the feedback sample pool as input, employs small-step fine-tuning to update the parameters of the gated fusion discriminator network, and retains historical representative samples for replay training to avoid catastrophic forgetting. The update cycle is configured daily or weekly, and after the update, the probability calibration parameters are relearned using a calibration dataset to maintain the reliability of the generated probabilities.

[0071] like Figure 4 The figure shows a performance comparison of an AI-generated text detection method based on a deep learning model according to an embodiment of this application. It demonstrates the performance comparison between the method of this application and existing technologies in terms of accuracy, recall, F1 score, and cross-domain robustness. In particular, regarding cross-domain robustness, the method of this application, benefiting from gated fusion and adaptive threshold mechanisms, achieves an improvement of over 15% compared to single semantic classification or rule matching methods.

[0072] According to a second aspect of this application, this application also provides an AI-generated text detection system based on a deep learning model. Figure 5 This is a structural block diagram of an AI-generated text detection system based on a deep learning model, according to an embodiment of this application. For example... Figure 5 As shown, the system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement an AI-generated text detection method based on a deep learning model according to the first aspect of this application. The system also includes other components well-known to those skilled in the art, such as a communication bus and a communication interface. Their configurations and functions are known in the art and will not be described further here.

[0073] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in this application, based on the technical solution and inventive concept of this application, should be within the scope of protection of this application.

Claims

1. An AI-generated text detection method based on a deep learning model, characterized in that, The detection methods include: The system acquires the text to be inspected, performs character normalization, sentence segmentation, and sliding window segmentation on the text to be inspected, generates a segmented text sequence, and generates index information that corresponds one-to-one with the segmented text sequence. For segmented text sequences, multi-view detection features are constructed. The multi-view detection features include at least: deep semantic discriminative features output by a pre-trained text encoder, generation trace features obtained from the log probability difference of a reference language model, style structure features obtained from text statistical indicators, and entity consistency features obtained from entity transfer relations. The construction of generation trace features includes: generating K perturbation texts for the segmented text sequence, where K is an integer greater than or equal to three, and the perturbation method includes random mask replacement or synonym replacement, with the mask ratio of random mask replacement being between 10% and 20%; calculating the log probability of the perturbation texts under the reference language model, and calculating the difference in log probability between the segmented texts and the corresponding perturbation texts, with the reference language model being an open-source language model; and statistically analyzing the log probability differences to obtain the mean, variance, and quantiles as generation trace features. The construction of style structure features includes: statistically analyzing the mean sentence length, standard deviation of sentence length, distribution of punctuation marks, and lexical richness; the statistical results are then used to form a style feature vector as style structure features. The construction of entity consistency features includes: performing entity recognition and referential clue extraction on segmented text sequences to obtain entity sets and entity categories; constructing entity transition graphs based on the occurrence order and co-occurrence relationship of entities in adjacent sentences or paragraphs; and performing graph convolutional encoding or transition probability statistics on entity transition graphs to obtain entity consistency vectors as entity consistency features. The multi-view detection features are input into the gated fusion discriminant network, which outputs the generation probability and confidence after probability calibration, and also outputs the segmented contribution. The risk level of the text to be detected is obtained. An adaptive discrimination threshold is determined based on the generation probability, confidence level, text length and risk level. The detection result is output as AI-generated text or human-generated text. Evidence fragments are generated based on segment contribution and index information.

2. The AI-generated text detection method based on a deep learning model according to claim 1, characterized in that, The character normalization, sentence segmentation, and sliding window segmentation include: The text to be detected is subjected to full-width / half-width conversion, whitespace normalization, abnormal control character filtering, and repeated line breaks merging to obtain normalized text. The standardized text is segmented into sentences and paragraphs based on paragraph boundaries and punctuation boundaries; When the length of a certain text segment exceeds a preset upper limit, the text segment is divided using an overlapping sliding window with an overlap ratio of 10% to 30% to generate the segmented text sequence.

3. The AI-generated text detection method based on a deep learning model according to claim 1, characterized in that, The construction of the style structure features also includes: The proportion of repeated segments is statistically analyzed, including the proportion of overlapping phrases between adjacent sentences and the proportion of high-frequency segments within the text. The consistency of paragraph structure is statistically analyzed, and the consistency of paragraph structure includes at least the distribution of paragraph length and the distribution of inter-sentence connecting words.

4. The AI-generated text detection method based on a deep learning model according to claim 1, characterized in that, The gated fusion discriminant network includes a multi-branch coding layer, a gated weight layer, and a classification output layer; The multi-branch coding layer performs nonlinear mapping on deep semantic discriminative features, generation trace features, style structure features and entity consistency features to obtain the branch representation of the alignment dimension; The gated weight layer calculates the fusion weight based on the text length and the domain indicator features. The domain indicator features are determined by the source type of the text to be detected, the title keyword matching results, or the domain labels output by the topic classification model. The classification output layer classifies the fused representation and outputs the generation probability.

5. The AI-generated text detection method based on a deep learning model according to claim 1, characterized in that, The probability calibration and confidence calculation include: Learn probability calibration parameters on the calibration dataset so that the calibrated generation probability matches the actual occurrence frequency; Perform temperature scaling or segmented calibration on the generation probability to obtain the calibrated generation probability; The confidence level is calculated based on the calibrated generation probability and historical reliability statistics.

6. The AI-generated text detection method based on a deep learning model according to claim 1, characterized in that, The determination of the adaptive discrimination threshold includes: A basic threshold is selected based on the risk level. The risk level is determined in a preset mapping table based on the business scenario or text source of the text to be detected, and the risk level includes at least low risk, medium risk and high risk. The basic threshold is dynamically adjusted based on the text length and confidence level to obtain an adaptive discrimination threshold. When the confidence level is lower than the preset lower limit, a verification status is output and a prompt message is generated to supplement the context or increase the text length.

7. The AI-generated text detection method based on a deep learning model according to claim 1, characterized in that, The method also includes incremental update and drift monitoring steps: A traceable record is generated for each detection, and the traceable record includes at least a text hash value, a segment index, a risk level, and a fragment of evidence; Receive manual review results or business feedback tags and form a feedback sample pool; Calculate the degree of difference between the new data distribution and the training data distribution on a preset statistical feature, wherein the degree of difference includes a weighted sum of Jensen-Shannon divergence and multi-quantile offset measure; When the difference exceeds a preset threshold, a drift alarm is triggered, and the parameters of the gating fusion discrimination network are fine-tuned or updated by comparison learning using the feedback sample pool.

8. An AI-generated text detection system based on a deep learning model, characterized in that, include: processor; A memory, wherein a computer program is stored; When the processor is configured to execute the computer program, it implements an AI-generated text detection method based on a deep learning model as described in any one of claims 1 to 7.

Images