Method and apparatus for detecting ai-generated content based on enhanced interaction intent

Abstract

The application discloses an AI generated content detection method and device based on interactive intention enhancement, and belongs to the technical field of AIGC detection. The method comprises the following steps: obtaining a first text, and performing vectorization coding on the first text to obtain a first word unit feature sequence of the first text; performing word unit level interactive intention enhancement on the first word unit feature sequence based on an interactive intention dictionary to obtain a first intention feature vector; obtaining a category intention word set corresponding to each intention category in the interactive intention dictionary; performing category level interactive intention enhancement on the first word unit feature sequence according to the category intention word set of each category to obtain a second intention feature vector; fusing the first intention feature vector and the second intention feature vector to obtain a third intention feature vector; and predicting a prediction category of the first text based on the third intention feature vector. The method can realize AIGC detection with high accuracy, good generalization and strong robustness.

Description

Technical Field

[0001] This invention relates to the field of AIGC detection technology, and in particular to a method and apparatus for detecting AI-generated content based on enhanced interactive intent. Background Technology

[0002] With the development of generative artificial intelligence technology, text generation techniques, represented by large language models, have been able to generate text content that closely resembles human expression in terms of syntax, semantics, and discourse structure. However, the widespread application of AI-generated content (AIGC) has also brought about a series of problems: on the one hand, AI-generated text is highly similar to human text in form, making it difficult to identify the source of the content; on the other hand, its large-scale dissemination may affect the authenticity of information and even exacerbate the risk of spreading misleading information. Therefore, it is necessary to accurately detect AIGC and human-generated text to reduce the adverse effects of AIGC.

[0003] AIGC detection methods in related technologies mainly start from the internal features of the text, focusing on language statistical features and generation probability distribution.

[0004] The principle behind AIGC detection methods using linguistic statistical features lies in artificially designing certain linguistic statistical features (such as word frequency distribution, TF-IDF values, part-of-speech tagging ratios, and punctuation frequency). Because there are quantifiable, systematic differences in language habits between human writing and AI-generated text—for example, human text typically exhibits greater vocabulary diversity and more natural sentence length variations—AI-generated text often displays higher statistical regularity. However, while such methods offer a degree of interpretability, their generalization ability is limited in complex contexts and cross-domain scenarios.

[0005] The principle behind AIGC detection methods based on generated probability distributions lies in the systematic differences between AIGC and human text in terms of probability distribution within language models. Specifically, AIGC is generated by a target language model through word-by-word prediction, such as the GPT series models. Therefore, some texts often have a higher generation probability under a large language model. Human writing, however, does not rely on this probability maximization mechanism, and the generation probability distribution of human text under any language model is usually more dispersed. Based on this principle, AIGC can be identified by calculating metrics such as log-likelihood, perplexity, or entropy. The technical advantage of this method is that it can achieve zero-shot detection without training data. However, its detection performance is highly sensitive to the type of generating model and the decoding strategy. When the text to be detected is generated by an unknown model or the decoding parameters change, the detection effect will significantly decrease, resulting in poor generalization. Summary of the Invention

[0006] This invention provides a method, apparatus, device, medium, and product for detecting AI-generated content based on enhanced interactive intent, which can achieve high accuracy, good generalization, and strong robustness in AIGC detection. The technical solution includes at least the following: Firstly, a method for detecting AI-generated content based on enhanced interactive intent is provided, comprising: acquiring a first text and vectorizing the first text to obtain a first word feature sequence of the first text, wherein the first text is the text to be detected as AI-generated content; enhancing the first word feature sequence with word-level interactive intent based on an interactive intent dictionary to obtain a first intent feature vector; acquiring a set of category intent words corresponding to each intent category in the interactive intent dictionary; enhancing the first word feature sequence with category-level interactive intent based on the set of category intent words for each category to obtain a second intent feature vector; fusing the first intent feature vector and the second intent feature vector to obtain a third intent feature vector; and predicting the predicted category of the first text based on the third intent feature vector; wherein the predicted category includes AI-generated text, human-generated text, text originating from humans but rewritten by AI, and text originating from AI but rewritten by humans.

[0007] Optionally, after vectorizing the first text, a text-level feature vector and a self-attention weight vector of the first text are also obtained. The step of performing word-level intent enhancement on the first word feature sequence based on the interaction intent dictionary to obtain a first intent feature vector includes: calculating the interaction intent score of each word feature in the first word feature sequence according to the interaction intent dictionary to obtain the interaction intent weight vector of the first word feature sequence; fusing the interaction intent weight vector and the self-attention weight vector using a learnable gating mechanism to obtain a fused weight vector; enhancing the first word feature sequence using the fused weight vector to obtain a second word sequence; and fusing the text-level feature vector with the weighted pooled second word sequence to obtain the first intent feature vector.

[0008] Optionally, the step of calculating the interaction intent score of each word feature in the first word feature sequence according to the interaction intent dictionary to obtain the interaction intent weight vector of the first word feature sequence includes: for any word feature in the first word feature sequence, calculating the similarity between the word feature and the semantic vector of each interaction intent word in the interaction intent dictionary, and taking the largest similarity as the interaction intent score of the word feature.

[0009] Optionally, the step of performing category-level interactive intent enhancement on the first lexical feature sequence based on the category intent word set of each category to obtain a second intent feature vector includes: for any lexical feature in the first lexical feature sequence, calculating the category intent weight of the lexical feature in the c-th intent category based on the category intent word set of the c-th intent category; performing a weighted summation on the first lexical feature sequence based on the category intent weight of each lexical feature in the c-th intent category to obtain the global intent feature of the c-th intent category; and concatenating the global intent features of each intent category and processing them through a linear layer mapping to obtain the second intent feature vector.

[0010] Optionally, the category intent weight of the j-th word feature in the first word feature sequence for the c-th intent category is calculated using the following formula:

[0011] in, Let be the category intent weight of the j-th word in the first word feature sequence for the c-th intent category. Let j be the lexical feature of the j-th lexical in the first lexical feature sequence. The set of category intent words for the c-th intent category Interactive intent words in yes semantic vectors, Indicates calculation and Similarity; The global intent feature of the c-th intent category is calculated using the following formula:

[0012] in, The global intent feature of the c-th intent category is given, where j ranges from 1 to... , The total number of lexical units in the first lexical feature sequence.

[0013] Optionally, the interaction intent dictionary includes five intent categories: assertion, commitment, instruction, declaration, and expression. The interaction intent dictionary is obtained by: obtaining a seed word set for each intent category; expanding the seed word set for each intent category to obtain an expanded candidate word set; and filtering the expanded candidate word set to obtain the interaction intent dictionary.

[0014] Secondly, an AI-generated content detection device based on interactive intent enhancement is also provided, comprising: a text acquisition module for acquiring first text and vectorizing the first text to obtain a first lexical feature sequence of the first text, wherein the first text is the text to be detected as AI-generated content; a first intent enhancement module for performing lexical-level interactive intent enhancement on the first lexical feature sequence based on an interactive intent dictionary to obtain a first intent feature vector; an intent dictionary acquisition module for acquiring a set of category intent words corresponding to each intent category in the interactive intent dictionary; a second intent enhancement module for performing category-level interactive intent enhancement on the first lexical feature sequence according to the set of category intent words for each category to obtain a second intent feature vector; a fusion module for fusing the first intent feature vector and the second intent feature vector to obtain a third intent feature vector; and a prediction module for predicting the predicted category of the first text based on the third intent feature vector; wherein the predicted category includes AI-generated text, human-generated text, text originating from humans but rewritten by AI, and text originating from AI but rewritten by humans.

[0015] Optionally, after vectorizing the first text, a text-level feature vector and a self-attention weight vector of the first text are obtained. The first intent enhancement module is further configured to calculate the interaction intent score of each word feature in the first word feature sequence according to the interaction intent dictionary, and obtain the interaction intent weight vector of the first word feature sequence; use a learnable gating mechanism to fuse the interaction intent weight vector and the self-attention weight vector to obtain a fused weight vector; use the fused weight vector to enhance the first word feature sequence to obtain a second word sequence; and fuse the text-level feature vector with the weighted pooled second word sequence to obtain the first intent feature vector.

[0016] Optionally, the first intent enhancement module is further configured to calculate the similarity between any one of the word features in the first word feature sequence and the semantic vector of each interactive intent word in the interactive intent dictionary, and take the maximum similarity as the interactive intent score of that word.

[0017] Optionally, the second intent enhancement module is further configured to, for any one word feature in the first word feature sequence, calculate the category intent weight of the word feature in the c-th intent category based on the category intent word set of the c-th intent category; perform a weighted summation on the first word feature sequence based on the category intent weight of each word feature in the c-th intent category to obtain the global intent feature of the c-th intent category; and concatenate the global intent features of each intent category and process them through a linear layer mapping to obtain the second intent feature vector.

[0018] Optionally, in the second intent enhancement module, the category intent weight of the j-th word feature in the first word feature sequence in the c-th intent category is calculated using the following formula:

[0019] in, Let be the category intent weight of the j-th word in the first word feature sequence for the c-th intent category. Let j be the lexical feature of the j-th lexical in the first lexical feature sequence. The set of category intent words for the c-th intent category Interactive intent words in yes semantic vectors, Indicates calculation and Similarity; The global intent feature of the c-th intent category is calculated using the following formula:

[0020] in, The global intent feature of the c-th intent category is given, where j ranges from 1 to... , The total number of lexical units in the first lexical feature sequence.

[0021] Optionally, the interaction intent dictionary includes five intent categories: assertion, commitment, instruction, declaration, and expression. The intent dictionary acquisition module is further configured to acquire the interaction intent dictionary in the following manner: acquire a seed word set for each intent category; expand the seed word set for each intent category to obtain an expanded candidate word set; and filter the expanded candidate word set to obtain the interaction intent dictionary.

[0022] Thirdly, a computer device is also provided, comprising: a memory and a processor, wherein the memory stores at least one computer program, the at least one computer program being loaded and executed by the processor to perform the AI-generated content detection method based on enhanced interactive intent described in the above embodiments.

[0023] Fourthly, a computer-readable storage medium is also provided, wherein at least one computer program is stored in the computer-readable storage medium, the at least one computer program being loaded and executed by a processor to perform the AI-generated content detection method based on enhanced interactive intent described in the above embodiments.

[0024] Fifthly, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the method described in the first aspect.

[0025] The beneficial effects of the technical solution provided by this invention include at least the following: In this embodiment, the first lexical feature sequence is enhanced with lexical-level interaction intent through an interaction intent dictionary, introducing interaction intent information at the lexical level to obtain a first intent feature vector. Then, the first lexical feature sequence is enhanced with category-level interaction intent based on the category intent word set for each category, introducing interaction intent information at the global category level to obtain a second intent feature vector. Finally, the first and second intent feature vectors are fused to obtain a third intent feature vector. AIGC detection is then performed on the first text based on this third intent feature vector, which is equivalent to performing AIGC detection after fusing interaction intent information and semantic information. Since there are significant differences in interaction intent between AI-generated text, human-generated text, text rewritten by AI but originating from humans, and text rewritten by humans but originating from AI, fusing interaction intent information and semantic information before performing AIGC detection can effectively improve the accuracy of AIGC detection. Furthermore, since interaction intent information is deep-level prior information independent of text statistical features, AIGC detection based on this interaction intent information has good generalization and robustness. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in this embodiment, the accompanying drawings used in the description of the embodiment will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 A flowchart of an AI-generated content detection method based on enhanced interactive intent, provided by an exemplary embodiment of the present invention, is shown. Figure 2 This is a schematic diagram of the IAD (A Interactional Intent-Aware Deep Learning Approach for Detecting AI-Generated Content) model. Figure 3 A schematic diagram of the structure of an AI-generated content detection device based on enhanced interactive intent provided in an exemplary embodiment of the present invention is shown. Figure 4This is a schematic diagram of the structure of a computer device provided in an exemplary embodiment of the present invention. Detailed Implementation

[0028] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The terms “connected” or “linked” and similar terms are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect.

[0029] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0030] Existing research typically formalizes AIGC detection as a binary classification task to identify whether text is generated by artificial intelligence or human users. Its core lies in understanding and extracting the differentiated features between AIGC and User-Generated Content (UGC). Current research primarily models the differences between AIGC and UGC at the levels of linguistic statistical patterns, linguistic features, and semantic representations. These methods can be broadly categorized into two types: one is based on detection methods using human feature engineering and discrimination criteria. This type measures the differences between AIGC and UGC by constructing text statistical and linguistic features such as text length, language style, lexical diversity, syntactic complexity, sentiment expression, perplexity, lexical probability distribution, probability curvature, and attention distribution, and then constructs detection algorithms accordingly. The other type is based on representation learning detection methods. This type uses pre-trained language models such as BERT and RoBERTa, as well as deep neural networks, to automatically learn deep semantic representations of text, thereby acquiring richer textual expressive features and completing classification predictions.

[0031] Although the two approaches differ in their specific implementation mechanisms and detection performance, they essentially treat text as a linguistic statistical object, relying on the relatively stable language distribution, statistical patterns, or semantic pattern differences between AIGC and UGC for discrimination. However, this research paradigm faces new challenges in online platform scenarios. Unlike traditional long-text scenarios such as news reports and academic writing, texts on online platforms typically exist in the form of comments, replies, Q&A, and interactive communications, generally characterized by short texts, weak context, and fragmented interactions. Their language expression is highly dependent on specific interactive contexts and immediate interactive goals, making stable language patterns and contextual cues sparser, thus significantly increasing the difficulty of detection. More importantly, with the continuous improvement of the generation capabilities of Large Language Models (LLMs), AIGC is increasingly approaching human text in terms of language fluency, semantic coherence, and naturalness of expression, further exacerbating the convergence of surface language styles and deep semantics between AIGC and UGC, making traditional detection methods that rely on linguistic statistical patterns and semantic pattern differences face increasingly severe challenges.

[0032] In fact, text on online platforms is not merely an expression of information, but also a manifestation of social interaction directed at specific individuals. Unlike traditional texts such as news reports and academic writing, which are primarily focused on information delivery, texts in online interactions typically serve multiple communicative functions, including responding, offering advice, providing comfort, making requests, making promises, and expressing emotions. Their language organization is often highly dependent on the specific interactive context and objectives.

[0033] This invention, based on speech act theory, introduces a new analytical perspective for AI-generated content detection from the interaction and pragmatic levels. In speech act theory, language is not only used to convey information but also as a crucial means of performing communicative acts. Its functions are typically categorized into assertive, expressive, directive, commissive, and declarative types. Compared to traditional language style or content features, speech act theory focuses more on the communicative function of text within specific interactive contexts and the underlying interactive intentions. In interpersonal interactions, human users typically form interactive intentions based on their social relationships with the interacting object, situational feedback, community norms, and interaction goals, and further organize their language expression. Therefore, their texts often possess strong situational adaptability, pragmatic responsiveness, and communicative consistency. In contrast, while AIGC can effectively simulate language forms and semantic structures, its generation process essentially still relies primarily on statistical patterns and probabilistic predictions in the training corpus. Its generation goals are more about satisfying cue word constraints, maintaining language fluency and semantic rationality, rather than achieving communicative behavior in real social interactions.

[0034] Therefore, the interactive intent in AIGC is more of a "mimicry" human-computer interaction behavior induced by prompts and statistical generation mechanisms, rather than a subjective participation stemming from real interpersonal interaction scenarios. This difference is not necessarily directly manifested as explicit language errors or stylistic abnormalities, but is more likely to reflect the different coupling relationships between interactive intent and context.

[0035] Specifically, user-generated text (UGC) typically revolves around specific interactive objects, real-world experiences, and immediate feedback, with its intent cues often highly coupled with the context. While AIGC can superficially simulate communicative behaviors such as requests, explanations, reassurance, or suggestions, its generation process relies more on the combination and reconstruction of high-frequency language patterns in the training corpus, thus exhibiting characteristics such as balance, generalization, and weak context dependence. Therefore, the difference between AIGC and UGC lies not only in surface-level linguistic statistical regularities and semantic patterns but also potentially in deeper social interaction mechanisms such as the expression of interactive intent and the realization of communicative functions. This perspective provides a new theoretical foundation for understanding the differences between AIGC and UGC and for constructing AIGC detection methods for online platforms.

[0036] More importantly, in human-machine hybrid text generation, this difference in interactive intent may still be preserved and further manifest as a difference in directional intent realization patterns. Specifically, although online text generated by human users and rewritten by large models (denoted as UGCtAI) and text generated by large models and rewritten by human users (denoted as AItUGC) both involve the joint participation of human users and LLMs, their interactive intent formation and language realization processes are fundamentally different due to the different generation order.

[0037] For UGCtAI, the text is first generated by humans and then rewritten by LLMs. In this process, the interaction purpose, contextual response, and experiential stance in the original human text usually still constitute the pragmatic core of the text, while AI mainly adjusts the wording, sentence structure, and fluency. Therefore, this type of text often retains strong characteristics of human interaction goals and contextualized responses, but at the same time, it is accompanied by traces of smoothing, standardization, and template-based implementation in local language.

[0038] Unlike UGCtAI, for AItUGC, the text is first generated by a large model and then revised by humans. At this point, the interactive intent framework in the original text has often been constructed by LLMs based on cue words and common pragmatic patterns in the training corpus, such as a relatively balanced structure of explanation, suggestion, reassurance, or summary. Subsequent human modifications are more about supplementing local experiences, using colloquial expressions, enhancing emotions, or refining contextual details. Therefore, although local language may exhibit strong human involvement, the overall organization of interactive intent categories and discourse structure may still retain the generalization characteristics of AIGC.

[0039] Therefore, the difference between UGCtAI and AItUGC is not merely a matter of "the degree of human involvement," but essentially a problem of directional mixed sources. Since interactive intent usually precedes the specific language implementation, the generation order will further influence which pragmatic information is retained as the core of the text and which information is only manifested as subsequent modification traces, thus causing the two to ultimately present different interactive intents and semantic feature patterns.

[0040] In summary, AIGC, UGC, UGCtAI, and AItUGC exhibit significant differences at both the interaction intent and semantic feature levels. Therefore, deep machine learning can be used to differentiate AIGC, UGC, UGCtAI, and AItUGC texts by specifically applying these differences at the interaction intent and semantic feature levels. Furthermore, because this difference is at the intent level, it is unaffected by specific models or pragmatic statistical features, exhibiting strong generalization and accuracy.

[0041] Based on the above principles, this invention proposes an AI-generated content detection method and apparatus based on enhanced interactive intent.

[0042] Figure 1 A flowchart illustrating an exemplary embodiment of the AI-generated content detection method based on enhanced interactive intent provided by the present invention is shown. See also: Figure 1 The method includes: In step 101, the first text is obtained and vectorized to obtain the first word feature sequence of the first text.

[0043] The first text is the text that needs to undergo AI-generated content detection.

[0044] In step 102, based on the interactive intent dictionary, the first word feature sequence is enhanced with word-level interactive intent to obtain the first intent feature vector.

[0045] In step 103, the set of category intent words corresponding to each intent category in the interaction intent dictionary is obtained.

[0046] In step 104, based on the category intent word set of each category, the first word feature sequence is enhanced with category-level interactive intent to obtain the second intent feature vector.

[0047] In step 105, the first intention feature vector and the second intention feature vector are fused to obtain the third intention feature vector.

[0048] In step 106, the predicted category of the first text is predicted based on the third intent feature vector.

[0049] The prediction categories include AI-generated text (AIGC), human-generated text (UGC), text originating from humans but rewritten by AI (UGCtAI), and text originating from AI but rewritten by humans (AItUGC).

[0050] In this embodiment, the first lexical feature sequence is enhanced with lexical-level interaction intent through an interaction intent dictionary, introducing interaction intent information at the lexical level to obtain a first intent feature vector. Then, the first lexical feature sequence is enhanced with category-level interaction intent based on the category intent word set for each category, introducing interaction intent information at the global category level to obtain a second intent feature vector. Finally, the first and second intent feature vectors are fused to obtain a third intent feature vector. AIGC detection is then performed on the first text based on this third intent feature vector, which is equivalent to performing AIGC detection after fusing interaction intent information and semantic information. Since there are significant differences in interaction intent between AI-generated text, human-generated text, text rewritten by AI but originating from humans, and text rewritten by humans but originating from AI, fusing interaction intent information and semantic information before performing AIGC detection can effectively improve the accuracy of AIGC detection. Furthermore, since interaction intent information is deep-level prior information independent of text statistical features, AIGC detection based on this interaction intent information has good generalization and robustness.

[0051] In this embodiment, steps 101 to 106 are implemented using the IAD model. Figure 2 This is a structural diagram of the IPA model, such as... Figure 2 As shown, the IAD model includes a text representation learning unit 21, a word-level interactive intent enhancement unit 22, a category-level interactive intent enhancement unit 23, and an AIGC detection unit 24.

[0052] Among them, the text representation learning unit 21 is connected to the word-level interaction intent enhancement unit 22 and the category-level interaction intent enhancement unit 23 respectively, and the outputs of the word-level interaction intent enhancement unit 22 and the category-level interaction intent enhancement unit 23 are both connected to the AIGC detection unit 24.

[0053] The following section explains each unit in the IAD model.

[0054] In the IAD model, the text representation learning unit 21 is used to perform step 101; the word-level interaction intent enhancement unit 22 is used to perform step 102; the category-level interaction intent enhancement unit 23 is used to perform steps 103 to 104; and the AIGC detection unit 24 is used to perform steps 105 to 106.

[0055] The text representation learning unit 21 aims to learn a general text representation from the original text input based on a specific pre-trained language model. Here, the original text input is the first text, which is the text that needs to be detected by AI-generated content. Before the IAD model training is complete, the first text can be text from the training set or validation set; after the IAD model training is complete, the first text can be text from an online platform that needs to be detected by AI-generated content.

[0056] Optionally, the text representation learning unit 21 is implemented based on the RoBERTa model, such as the RoBERTa-Large model. The RoBERTa model can vectorize the input text sequence. The RoBERTa model has shown superior performance in a large number of natural language understanding tasks, and its robust masking language modeling strategy can capture deeper contextual dependencies.

[0057] In some embodiments, other pre-trained text encoders, such as BERT, DeBERT, DistilBERT, and ERNIE, can also be used to implement step 101. These models are also based on the Transformer architecture and can generate context-sensitive lexical representations. Furthermore, in scenarios with limited computational resources, lightweight models such as ALBERT and MobileBERT can be used as alternatives to reduce computational overhead.

[0058] When implementing step 101 in the text representation learning unit 21, the first text is first preprocessed, including removing hyperlinks, abnormal symbols, and noisy characters, as well as normalization operations, so that the first text... Include A valid token is represented as .in, Let j be the j-th word in the first text, where j is a positive integer and its value ranges from 1 to M. The maximum length of the first text is given. For example, when using the RoBERTa model for vectorized encoding, since the maximum processing length of the RoBERTa model is 512, therefore... The value is 512. When using other text encoders to implement vectorized encoding, Other values ​​are also possible.

[0059] If the length of the first text is greater than Then the text exceeding the limit in the first text will be... The first part is truncated; if the length of the first text is insufficient... Then, the padding mechanism of the pre-trained word segmenter is used to fill the first text to a length of M, and attention mask is used to distinguish the effective words from the filling positions.

[0060] Then, the preprocessed first text is input into the RoBERTa model to obtain the first lexical feature sequence of the first text output by the RoBERTa model. This first lexical feature sequence contains the lexical features of each word in the first text, represented as follows: , .in, The first word feature sequence, The j-th word in the first word feature sequence The corresponding lexical features, where M is the length of the first text and h is the hidden layer dimension of the lexical features.

[0061] In addition to outputting the first word feature sequence of the first text, the RoBERTa model also outputs the self-attention weight of each word in the first text. The self-attention weights of all words in the first text constitute the self-attention weight vector of the first text. , for Self-attention weights.

[0062] In addition, the RoBERTa model also outputs the text-level features of the first text. In this embodiment, these text-level features are the lexical features at the [CLS] position, represented as follows: In some embodiments, the first word feature sequence may also be used. Text-level features are obtained by performing average pooling or max pooling.

[0063] In reality, text generation on online platforms involves not only the expression of interactive intent, but also deeper cognitive processing activities such as semantic organization, logical planning, and contextual coordination. Therefore, text generation is essentially the result of the synergistic effect of expressing interactive intent and deep semantic processing.

[0064] Specifically, in real-world online interactions, users typically first formulate an interaction goal based on a specific interactive context, and then further complete language organization and semantic expression. That is, the global interaction intent provides directional constraints on the local semantic organization and content realization. Therefore, simply identifying differences in interaction intent across texts from different sources is insufficient to fully realize human-computer hybrid text detection in complex scenarios. It is also necessary to further characterize the guiding relationship between interaction intent and the semantic organization and text realization process, so that interaction intent knowledge can be more effectively integrated into AIGC detection tasks.

[0065] Dual-System Theory provides an important theoretical perspective for understanding this process. According to Dual-System Theory, human information processing is typically influenced by both a fast system and a slow system. The fast system is primarily responsible for rapid, intuitive, and context-driven information processing, quickly forming immediate interaction goals and behavioral intentions in interactive scenarios. The slow system, on the other hand, is mainly responsible for relatively deliberate semantic organization, logical reasoning, and content planning. In real online interactions, users typically first form interaction intentions based on specific contexts and then further complete linguistic expression and semantic organization. This process essentially reflects a dynamic synergy between the expression of interaction intentions and deep semantic processing.

[0066] Therefore, in AIGC detection tasks, modeling solely based on statistical patterns of interaction intent or surface semantic patterns may be insufficient to fully characterize the deep social interaction mechanisms involved in human text generation. Conversely, incorporating interaction intent knowledge as a theoretical prior into text representation learning, and further modeling the synergistic relationship between global interaction intent and local semantic organization, can help improve the model's ability to express and interpret the differences between AIGC and UGC in complex online platform scenarios.

[0067] Based on the above ideas, this invention constructs a word-level interactive intent enhancement unit 22 and a category-level interactive intent enhancement unit 23. The word-level interactive intent enhancement unit 22 can integrate interactive intent knowledge as a theoretical prior into text representation learning, while the category-level interactive intent enhancement unit 23 can further model the collaborative relationship between global interactive intent and local semantic organization.

[0068] The expression of interactive intent is significantly context-dependent, meaning that the same word may carry different communicative functions in different contexts. Therefore, before explicitly extracting the intent representation, it is necessary to first enhance the first lexical feature sequence using an interactive intent dictionary. In this embodiment, the lexical-level interactive intent enhancement unit 22 aims to use interactive intent knowledge as task-related prior guiding information to enhance the interactive intent of each lexical feature in the first lexical feature sequence at the lexical level, thereby integrating pragmatic information (interactive intent) into the text representation learning process and improving the IAD model's ability to perceive interactive intent. The lexical-level interactive intent enhancement unit 22 is used to execute step 102.

[0069] In step 102, the interaction intent refers to the communicative purpose embodied in the text within a specific context, and is a key variable connecting semantics and behavior. Optionally, step 102 includes steps 1021 to 1024 below.

[0070] Step 1021: Calculate the interaction intent score of each word feature in the first word feature sequence based on the interaction intent dictionary, and obtain the interaction intent weight vector of the first word feature sequence.

[0071] Before proceeding to step 1021, an interaction intent dictionary needs to be obtained. Based on speech act theory, this invention systematically categorizes interaction intents in online platform texts into five types: assertion, commitment, instruction, declaration, and expression. Based on this, an interaction intent dictionary for AI-generated content detection is constructed. Guided by speech act theory, this dictionary aims to capture the linguistic representation patterns of texts under different interactive functions at the lexical level, providing the model with structured pragmatic prior knowledge.

[0072] Among the five intent categories, assertion-type interactive intent words are narrative and descriptive words used to convey factual information, describe events, or express opinions, such as “believe,” “in fact,” and “find.”

[0073] Instruction-type interactive intent words are words used to express commands, requests, or guided actions, such as "must," "please," and "don't."

[0074] Commitment-type interaction intent words are words that indicate future action intentions or responsibilities, such as "I will," "I will," and "I promise."

[0075] Declaration-type interactive intent words are words used to express actions such as identity confirmation, institutional declaration, or formal decision, such as "announce," "determine," and "appoint."

[0076] Interactive intent words of the expressive type are words used to express emotions, attitudes, and emotional responses, such as "anger," "shock," and "understanding."

[0077] Optionally, the following three steps can be used to obtain the interaction intent dictionary.

[0078] The first step is to obtain the seed word set for each intent category.

[0079] In one possible implementation, representative corpus samples can be collected from a corpus, and annotators can systematically read and analyze the sample data to extract representative keywords or phrases from each intent category, thereby constructing a seed word set.

[0080] In another possible implementation, seed words for each intent category are generated using a large language model based on the definitions of different speech act types in speech act theory, resulting in a seed word set. In this case, to reduce the risk of semantic drift, the generated results must strictly conform to the theoretical definition of the corresponding speech act category during the seed word generation stage.

[0081] For example, the initial seed words were generated by GPT-5 and cross-validated using Qwen 3 and DeepSeek-V3.

[0082] The second step is to expand the seed word set for each intent category to obtain an expanded candidate word set.

[0083] In this embodiment, the seed word set of interactive intent words is expanded based on cosine similarity. For example, for the seed word set of interactive intent words of any intent category, the semantic vectors of the interactive intent words in the seed word set are obtained. Then, the Top-K words with the closest cosine similarity to the semantic vectors of the interactive intent words in the seed word set are selected from a large-scale corpus as expansion candidate words, thereby improving the semantic coverage of the interactive intent dictionary.

[0084] By using the above method, we can obtain extended candidate words for interactive intent words of each intent category, thus obtaining an extended candidate word set.

[0085] The third step is to filter the expanded candidate word set to obtain the interactive intent dictionary.

[0086] In this embodiment, the extended candidate words are screened in multiple rounds: (1) words with semantic ambiguity or irrelevant to the target intent are eliminated; (2) words that are only relevant in a specific context (low-relevance words) are removed; (3) the domain experts conduct manual review to further eliminate redundant and low-quality words, so as to ensure the conciseness and accuracy of the dictionary.

[0087] In addition, the words in the interaction intent dictionary can be classified, organized, and structurally represented according to five intent categories and stored in a standardized format for easy retrieval of the interaction intent dictionary later.

[0088] By performing multiple rounds of filtering on the expanded candidate word set, the accuracy and reliability of the interactive intent dictionary can be improved. Table 1 below shows an example of the interactive intent dictionary provided in this embodiment.

[0089] Table 1: Examples of interactive intent dictionaries.

[0090]

[0091] After obtaining the interaction intent dictionary, the RoBERTa model in step 101 can be used to vectorize the interaction intent words in the dictionary. For example, the interaction intent words can be... Encoded as semantic vectors .

[0092] After obtaining the interaction intent dictionary, step 1021 can be executed.

[0093] Optionally, step 1021 includes: for any one word feature in the first word feature sequence, calculating the similarity between the word feature and the semantic vector of each interactive intent word in the interactive intent dictionary, and taking the largest similarity as the interactive intent score of the word feature.

[0094] For example, for the j-th lexical feature in the first lexical feature sequence The interaction intent score is calculated using the following formula (1). .

[0095] (1) In formula (1), For the j-th word feature Interaction intent score, For the interactive intent dictionary Interactive intent words in yes semantic vectors, Indicates calculation and The similarity is calculated using cosine similarity.

[0096] In this embodiment, the similarity score is calculated as the interaction intent score. In some embodiments, other similarity metrics such as Euclidean distance, Manhattan distance, or dot product similarity can also be used as the interaction intent score. In addition, the normalization of the intent score can use max-min normalization or Z-score normalization instead of the original normalization method on effective lexical units.

[0097] By applying the same method to each lexical feature in the first lexical feature sequence... The interaction intent score is calculated in the same way, and the interaction intent weight vector is obtained as follows: .

[0098] To avoid the influence of padding or special words on the statistical results, the intent score is normalized only on the valid words; that is, padding words are not included in the intent score calculation.

[0099] Step 1022: A learnable gating mechanism is used to fuse the interaction intent weight vector and the self-attention weight vector to obtain a fused weight vector.

[0100] In this embodiment, a learnable gating mechanism is first used to calculate the first gating weight of each word (the first gating weight is a gating weight at the word level), and then the interaction intent weight vector and the self-attention weight vector are fused using the first gating weight to obtain the fused weight vector.

[0101] Optionally, the first gating weight is calculated using formula (2), and the fusion weight vector is calculated using formula (3).

[0102] (2) In formula (2), for The gating weights are used to balance the contributions of intent scores and attention scores. For non-linear activation functions, such as sigmoid or tanh, , The parameters that need to be learned when calculating the first gating weights, where This is the weight matrix. For bias.

[0103] (3) In formula (3), for The fusion weight, for Self-attention weights.

[0104] By calculating the fusion weight of each word feature using formula (3), the fusion weight vector can be obtained as follows: .

[0105] Step 1023: The first word feature sequence is enhanced by using a fusion weight vector to obtain the second word sequence.

[0106] Optionally, step 1023 can be represented by the following formula (4).

[0107] (4) In formula (4), To adopt fusion weights right The enhanced lexical features, that is, the features in the second lexical sequence that are related to... The corresponding enhanced features.

[0108] By enhancing each word feature in the first word feature sequence with the corresponding fusion weight (that is, by enhancing it in the way of formula (4)), the second word sequence can be obtained as follows: .

[0109] Step 1023 achieves co-modeling of semantics and intent at the word level, which can dynamically adjust the influence weight of intent information on each word.

[0110] Step 1024: The text-level feature vector is fused with the weighted pooled second word sequence to obtain the first intent feature vector.

[0111] Since the text-level feature vectors and the second word sequence have different dimensions, they cannot be directly fused. By weighted pooling of the second word sequence, the dimensions of the two can be made the same, thus enabling their fusion.

[0112] Optionally, step 1024 can be represented by the following formula (5).

[0113] (5) In formula (5), The first intention feature vector, The learnable second gating weights are used to fuse the original text representation (text-level feature vectors) with the intent-guided text representation (second lexical sequence). , , The parameters that need to be learned when calculating the second gating weights, where This is the weight matrix. For bias. Second gating weight Dimensions and , The same dimensions . This indicates that a weighted pooling operation is being performed. For the second word sequence, This is the text-level feature vector of the first text.

[0114] Step 102 enables the IAD model to obtain interactive intent information at the lexical level, achieving lexical-level interactive intent enhancement. That is, in text representation learning, interactive intent is integrated as pragmatic prior knowledge into semantic encoding, allowing the model to simultaneously capture "what was said" and "why it was said," thereby obtaining a more discriminative representation. Furthermore, step 1024 also enables the IAD model to obtain interactive intent information at the text level, thus achieving text-level interactive intent enhancement.

[0115] After completing the lexical-level intent guidance enhancement, the IAD model has obtained the first intent feature vector, which integrates semantic information and interactive intent cues. However, lexical-level enhancement mainly captures local intent cues and is insufficient to characterize the overall communicative functional structure of the text. Therefore, it is necessary to further abstract and aggregate intent information at the category level based on lexical-level enhancement to construct a macro-intent representation that can reflect the overall communicative function of the text.

[0116] In this embodiment, the category-level interactive intent enhancement unit 23 can further model the collaborative relationship between global interactive intent and local semantic organization. Based on the intent classification system of speech act theory, the category-level interactive intent enhancement unit 23 generates a global interactive intent vector by extracting, weighting, and fusing intent cues of various categories in the text, providing high-level semantic intent features for downstream detection tasks. Furthermore, the category-level interactive intent enhancement unit 23 is used to execute steps 103 to 104.

[0117] In step 103, the interaction intent dictionary has already been obtained, and the interaction intent words for each intent category have been classified and stored. Therefore, the set of category intent words corresponding to each intent category can be obtained directly.

[0118] Optionally, step 104 includes steps 1041 to 1043 as follows.

[0119] Step 1041: For any word feature in the first word feature sequence, calculate the category intent weight of the word feature in the c-th intent category based on the category intent word set of the c-th intent category.

[0120] Optionally, the j-th lexical feature in the first lexical feature sequence can be calculated using the following formula (6). Category intent weight for the c-th intent category: (6) In formula (6), Let be the category intent weight of the j-th word in the first word feature sequence for the c-th intent category. Let j be the lexical feature of the j-th lexical in the first lexical feature sequence. The set of category intent words for the c-th intent category Interactive intent words in yes semantic vectors, Indicates calculation and The similarity.

[0121] It should be noted that when calculating formula (6), invalid words filled by padding need to be masked, that is, invalid words filled by padding do not participate in the calculation of formula (6).

[0122] Step 1042: Based on the category intent weight of each word feature in the c-th intent category, perform a weighted summation on the first word feature sequence to obtain the global intent feature of the c-th intent category.

[0123] By calculating each lexical feature using formula (6), the category intent weight of each lexical feature in the c-th intent category can be obtained. Based on this, step 1042 is implemented using formula (7).

[0124] (7) In formula (7), Let be the global intent feature of the c-th intent category, and let represent the global intent distribution of the first text in the c-th intent category.

[0125] For each intent category, the global intent features can be calculated using steps 1041 to 1042 above, thereby obtaining the global intent features for each intent category.

[0126] Step 1043: After concatenating the global intent features of each intent category, process them through a linear layer mapping to obtain the second intent feature vector.

[0127] The process in step 1043 can be represented as follows: ,in This is the second intention feature vector. This represents a linear layer mapping operation. This represents a vector concatenation operation, where c ranges from 1 to C, where C is the total number of intent categories. In this embodiment, there are 5 intent categories, so C is 5. In some embodiments, there may be more or fewer intent categories, so the value of C is not limited to 5 and can also be other positive integers.

[0128] In fact, a post may contain multiple intents (expressed by multiple sentences). Therefore, step 1043 is to fuse the five intent categories to obtain the second intent feature vector, rather than selecting only a certain global intent feature as the second intent feature vector.

[0129] This second intent feature vector integrates semantic intent features across intent categories, providing high-level intent guidance for downstream discrimination tasks.

[0130] AIGC detection unit 24 is used to predict the predicted category of the first text based on the input first intent feature vector and second intent feature vector, and AIGC detection unit 24 is used to perform steps 105 to 106.

[0131] Since the contribution of local features and global features differs in different sentence structures, a learnable gating mechanism is used in step 105 to fuse the first intention feature vector and the second intention feature vector, so as to achieve adaptive feature fusion for different samples.

[0132] Therefore, before executing step 105, it is necessary to calculate the third gating weight using formula (8). .

[0133] (8) In formula (8), The third gating weight is used to fuse the first intent feature vector and the second intent feature vector. , The parameters that need to be learned when calculating the third gating weights, where This is the weight matrix. For bias.

[0134] The third gating weight Dimensions and , The dimensions are the same, thus enabling adaptive weight allocation for each sample at the feature dimension level. This indicates a pooling operation.

[0135] After obtaining the third gating weight, step 105 is implemented using formula (9).

[0136] (9) In formula (9), For the third intention feature vector, This represents element-wise multiplication. This indicates a pooling operation.

[0137] Step 105 enables the IAD model to dynamically adjust the balance between semantic evidence and interactive intent structure based on the specific features of the input text.

[0138] In this embodiment, a learnable gating vector is used to perform a weighted fusion of the first intent feature vector and the second intent feature vector. Alternatively, a multi-head attention mechanism, a cross-attention mechanism, or a gating loop unit can be used to fuse the first and second intent feature vectors. Furthermore, a simple vector concatenation followed by a wiring layer can also be used as an alternative, although this may slightly affect performance, it still achieves the core objective of this invention.

[0139] Here, steps 101 to 105 have already deeply fused the interactive intent into the first word feature sequence, resulting in the third intent feature vector. Therefore, in step 106, only a single deep learning model capable of performing classification prediction tasks is needed for prediction. In this embodiment, a multilayer perceptron is used to implement step 106. The input of this multilayer perceptron is the third intent feature vector after average pooling, and the output is the predicted category.

[0140] As mentioned earlier, the prediction categories include four types: AIGC, UGC, UGCtAI, and AItUGC. Therefore, the multilayer perceptron opportunistically outputs the probability of the first text in these four prediction categories. Then, the prediction category with the highest probability can be used as the final prediction category of the first text, thereby realizing the detection of AI-generated content of the first text.

[0141] The IAD model uses the cross-entropy loss function for optimization, and its learning objective is to minimize the multi-class classification loss across the entire dataset. When training the IAD model, in addition to optimizing the parameters of the multilayer perceptron, it is also necessary to optimize the parameters of each learnable gating mechanism, including... , , , , , wait.

[0142] Optionally, the loss function is expressed by formula (10).

[0143] (10) In formula (10), Let be the loss function of the IAD model. For predicting the set of categories, , For predicting the set of categories One of the prediction categories, Let be the true label of the i-th sample in the training set, where i is a positive integer ranging from 1 to N, and N is the total number of samples in the training set. As an indicator function, it is defined when the expression within the parentheses is true. The expression equals 1 when the expression within the parentheses is false. Equal to 0, Let be the probability of the i-th sample in the k-th predicted class output by the multilayer perceptron.

[0144] In some embodiments, other deep learning prediction models may also be used to implement step 106, such as convolutional neural networks, recurrent neural networks and other deep learning models and their hybrid structures.

[0145] This invention introduces prior knowledge of intent at the lexical level through a lexical-level interactive intent enhancement unit 22, weighting each lexical unit in the text for intent perception, enabling the IAD model to capture semantic cues highly relevant to the interactive intent. Simultaneously, a category-level interactive intent enhancement unit 23 abstracts and aggregates intent information at the category level, constructing a macro-intention representation reflecting the overall communicative function of the text. The synergistic effect of these two units allows the IAD model to simultaneously focus on "what is said" and "why it is said," thereby achieving higher discrimination accuracy in AI-generated content classification and detection tasks.

[0146] Table 2 compares the accuracy, precision, recall, and F1 score of the IAD model with other models in related technologies on the AI-generated content classification and detection task. It can be seen that the IAD model proposed in this invention outperforms existing baseline models in terms of accuracy, precision, recall, and F1 score. Specifically, the accuracy of IAD reaches 0.8448, an improvement of 2.10 percentage points compared to the best baseline model COCO (0.8238); the F1 score reaches 0.8443, an improvement of 2.42 percentage points compared to the best baseline model. These comparative experimental results verify the effectiveness of the AI-generated content detection method based on enhanced interactive intent proposed in this invention.

[0147] By fusing semantic and interactive intent information, the IAD model achieves a multi-dimensional characterization of text generation mechanisms, demonstrating significant performance advantages in AI-generated content classification and detection tasks. This fully demonstrates that introducing interactive intent from speech act theory as a pragmatic prior into the text representation learning process can effectively improve the model's ability to discriminate complex generation patterns.

[0148] Table 2: Comparison of accuracy, precision, recall, and F1 score of the IAD model and other models in related technologies on AI-generated content classification and detection tasks.

[0149]

[0150] To further explore the impact of interactive intent on AIGC detection performance, this study removed key modules from the IAD model one by one through ablation experiments to evaluate the contribution of each module to the overall performance. The experimental settings are as follows.

[0151] IAD_1: Removes the word-level interactive intent enhancement unit 22 and the category-level interactive intent enhancement unit 23. The AIGC detection unit 24 uses only the basic text representation as input for classification prediction. It is used to evaluate the impact of the text representation and interactive intent representation of intent enhancement on the overall performance of the model.

[0152] IAD_2: Remove the word-level interactive intent enhancement unit 22. The AIGC detection unit 24 uses the basic text representation and interactive intent representation as input for classification prediction to verify the impact of the word-level interactive intent enhancement unit 22 on the overall performance of the model.

[0153] IAD_3: Remove the category-level interactive intent enhancement unit 23. The AIGC detection unit 24 only uses the text representation of intent enhancement as input for classification prediction. This is used to assess the impact of the interactive intent extracted by the category-level interactive intent enhancement unit 23 on the overall performance of the model.

[0154] Table 3 presents the results of the ablation experiment. As can be seen from Table 3, both the word-level interactive intent enhancement unit 22 and the category-level interactive intent enhancement unit 23 significantly contribute to the model performance, and there is a synergistic effect between the two, jointly improving the model's ability to recognize AIGC content.

[0155] Table 3: Ablation Experiment Results

[0156] This invention introduces a theoretical prior of interactive intent as an inductive bias for the model. This theoretical prior is independent of specific data distributions and model parameters, providing a stable basis for the model's judgment, thus exhibiting good robustness and generalization in various scenarios. To address complex situations such as adversarial interference that may exist in online communities, this invention verifies the model's robustness and generalization through adversarial attack sample testing.

[0157] Table 4 shows the results of the adversarial attack sample test. The experimental results show that even after the input text has been perturbed by word substitution, rewriting, and other perturbations, the IAD model based on the prior theory of interaction intent can still maintain a high discrimination accuracy. This is because interaction intent, as a deep pragmatic feature, relies on a theoretical prior that does not change with data fluctuations or surface perturbations, thus giving the model good anti-interference ability and cross-scenario generalization ability.

[0158] Table 4: Test results of adversarial attack samples.

[0159]

[0160] To address the lack of interpretability in the detection process of existing deep learning models, this invention constructs an interactive intent dictionary based on speech act theory and calculates intent scores at the word level, thereby making the model's detection process explicitly interpretable.

[0161] For example, when determining the source of text, the model calculates the similarity score between each lexical unit and each category of intent words through the lexical-level interactive intent enhancement unit 22. This allows for the explicit location and focus on vocabulary and expressions related to specific intent categories (including assertions, instructions, commitments, expressions, and declarations). Since these intent categories have clear linguistic definitions and semantic orientations, the basis for the model's focus can be intuitively understood and traced. Therefore, the detection results of this invention have good interpretability.

[0162] On the one hand, interpretable detection results help enhance users' trust in the model's output and improve the acceptance of the invention in practical application scenarios. On the other hand, by analyzing the distribution of intent words that the model focuses on, researchers can intuitively judge whether the model's behavior meets expectations, providing clear guidance for subsequent model debugging, error analysis, and performance optimization.

[0163] The following are device embodiments of this application. For details not described in detail in the device embodiments, please refer to the above method embodiments.

[0164] Figure 3 This diagram illustrates the structure of an AI-generated content detection device based on enhanced interactive intent, according to an exemplary embodiment of the present invention. See also... Figure 3 The AI-generated content detection device 300 based on interactive intent enhancement includes: a text acquisition module 301, a first intent enhancement module 302, an intent dictionary acquisition module 303, a second intent enhancement module 304, a fusion module 305, and a prediction module 306.

[0165] The text acquisition module 301 is used to acquire the first text and perform vectorization encoding on the first text to obtain the first word feature sequence of the first text. The first text is the text that needs to be detected by AI-generated content.

[0166] The first intent enhancement module 302 is used to perform word-level interactive intent enhancement on the first word feature sequence based on the interactive intent dictionary to obtain the first intent feature vector.

[0167] The intent dictionary acquisition module 303 is used to acquire the set of category intent words corresponding to each intent category in the interaction intent dictionary.

[0168] The second intent enhancement module 304 is used to perform category-level interactive intent enhancement on the first word feature sequence based on the category intent word set of each category, so as to obtain the second intent feature vector.

[0169] The fusion module 305 is used to fuse the first intent feature vector and the second intent feature vector to obtain the third intent feature vector.

[0170] The prediction module 306 is used to predict the predicted category of the first text based on the third intent feature vector.

[0171] The prediction categories include AI-generated text, human-generated text, text originally from humans but rewritten by AI, and text originally from AI but rewritten by humans.

[0172] Optionally, after vectorizing the first text, a text-level feature vector and a self-attention weight vector of the first text are obtained. The first intent enhancement module 302 is also used to calculate the interaction intent score of each word feature in the first word feature sequence according to the interaction intent dictionary, and obtain the interaction intent weight vector of the first word feature sequence; the interaction intent weight vector and the self-attention weight vector are fused using a learnable gating mechanism to obtain a fused weight vector; the fused weight vector is used to enhance the first word feature sequence to obtain a second word sequence; the text-level feature vector is fused with the weighted pooled second word sequence to obtain the first intent feature vector.

[0173] Optionally, the first intent enhancement module 302 is further configured to calculate the similarity between any one of the word features in the first word feature sequence and the semantic vector of each interactive intent word in the interactive intent dictionary, and take the largest similarity as the interactive intent score of that word.

[0174] Optionally, the second intent enhancement module 304 is further configured to, for any word feature in the first word feature sequence, calculate the category intent weight of the word feature in the c-th intent category based on the category intent word set of the c-th intent category; perform a weighted summation on the first word feature sequence based on the category intent weight of each word feature in the c-th intent category to obtain the global intent feature of the c-th intent category; and concatenate the global intent features of each intent category and process them through a linear layer mapping to obtain the second intent feature vector.

[0175] Optionally, in the second intent enhancement module 304, the category intent weight of the j-th lexical feature in the first lexical feature sequence in the c-th intent category is calculated using the following formula:

[0176] in, Let be the category intent weight of the j-th word in the first word feature sequence for the c-th intent category. Let j be the lexical feature of the j-th lexical in the first lexical feature sequence. The set of category intent words for the c-th intent category Interactive intent words in yes semantic vectors, Indicates calculation and Similarity; The global intent feature for the c-th intent category is calculated using the following formula:

[0177] in, Let j be the global intent feature for the c-th intent category, where j ranges from 1 to... , The total number of lexical units in the first lexical unit feature sequence.

[0178] Optionally, the interaction intent dictionary includes five intent categories: assertion, commitment, instruction, declaration, and expression. The intent dictionary acquisition module 303 is also used to acquire the interaction intent dictionary in the following ways: acquire the seed word set for each intent category; expand the seed word set for each intent category to obtain an expanded candidate word set; and filter the expanded candidate word set to obtain the interaction intent dictionary.

[0179] It should be noted that the AI-generated content detection device based on enhanced interactive intent provided in the above embodiments is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the AI-generated content detection device based on enhanced interactive intent provided in the above embodiments and the AI-generated content detection method based on enhanced interactive intent belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0180] The module division in this embodiment of the invention is illustrative and represents only one logical functional division. In actual implementation, other division methods are possible. Furthermore, the functional modules in each embodiment of the invention can be integrated into a single processor, exist as separate physical entities, or consist of two or more modules integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0181] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a terminal device (which may be a personal computer, mobile phone, or communication device, etc.) or processor to execute all or part of the steps of the method of the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0182] Figure 4 This is a schematic diagram of the structure of a computer device provided in an exemplary embodiment of the present invention. For example... Figure 4 As shown, the computer device 400 includes a processor 401 and a memory 402.

[0183] Processor 401 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 401 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 401 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 401 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 401 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0184] The memory 402 may include one or more computer-readable storage media, which may be non-transitory. The memory 402 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 402 are used to store at least one instruction, which is executed by the processor 401 to implement the AI-generated content detection method based on enhanced interactive intent provided in this embodiment of the invention.

[0185] Those skilled in the art will understand that Figure 4 The structure shown does not constitute a limitation on computer device 400, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0186] This invention also provides a non-transitory computer-readable storage medium, which, when the instructions in the storage medium are executed by the processor of a computer device, enables the computer device to execute the AI-generated content detection method based on enhanced interactive intent provided in this invention.

[0187] This invention also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the AI-generated content detection method based on enhanced interactive intent provided in this invention.

[0188] The above description is merely an optional embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting AI-generated content based on enhanced interactive intent, characterized in that, The AI-generated content detection method based on enhanced interactive intent includes: The first text is obtained and vectorized to obtain the first word feature sequence of the first text. The first text is the text that needs to be detected by AI-generated content. Based on the interactive intent dictionary, the first word feature sequence is enhanced with word-level interactive intent to obtain the first intent feature vector. Obtain the set of category intent words corresponding to each intent category in the interaction intent dictionary; Based on the set of category intent words for each category, the first word feature sequence is enhanced with category-level interactive intent to obtain a second intent feature vector; The first intent feature vector and the second intent feature vector are fused to obtain the third intent feature vector; Based on the third intent feature vector, predict the predicted category of the first text; The prediction categories include AI-generated text, human-generated text, text originally written by humans but rewritten by AI, and text originally written by AI but rewritten by humans. After vectorizing the first text, a text-level feature vector and a self-attention weight vector are obtained. The step of performing word-level intent enhancement on the first word feature sequence based on the interactive intent dictionary to obtain a first intent feature vector includes: The interaction intent score of each word feature in the first word feature sequence is calculated based on the interaction intent dictionary to obtain the interaction intent weight vector of the first word feature sequence. A learnable gating mechanism is used to fuse the interaction intent weight vector and the self-attention weight vector to obtain a fused weight vector; The first word feature sequence is enhanced using the fusion weight vector to obtain the second word sequence; The text-level feature vector is fused with the weighted pooled second word sequence to obtain the first intent feature vector; The step of performing category-level interactive intent enhancement on the first lexical feature sequence based on the category intent word set for each category to obtain a second intent feature vector includes: For any word feature in the first word feature sequence, the category intent weight of the word feature in the c-th intent category is calculated based on the category intent word set of the c-th intent category. Based on the category intent weight of each lexical feature in the c-th intent category, the first lexical feature sequence is weighted and summed to obtain the global intent feature of the c-th intent category; The global intent features of each intent category are concatenated and then processed through a linear layer mapping to obtain the second intent feature vector.

2. The AI-generated content detection method based on enhanced interactive intent as described in claim 1, characterized in that, The step of calculating the interaction intent score for each lexical feature in the first lexical feature sequence based on the interaction intent dictionary, to obtain the interaction intent weight vector of the first lexical feature sequence, includes: For any lexical feature in the first lexical feature sequence, calculate the similarity between the lexical feature and the semantic vector of each interactive intent word in the interactive intent dictionary, and take the maximum similarity as the interactive intent score of the lexical feature.

3. The AI-generated content detection method based on enhanced interactive intent as described in claim 1, characterized in that, The category intent weight of the j-th word feature in the first word feature sequence for the c-th intent category is calculated using the following formula: in, Let be the category intent weight of the j-th word in the first word feature sequence for the c-th intent category. Let j be the lexical feature of the j-th lexical in the first lexical feature sequence. The set of category intent words for the c-th intent category Interactive intent words in yes semantic vectors, Indicates calculation and Similarity; The global intent feature of the c-th intent category is calculated using the following formula: in, The global intent feature of the c-th intent category is given, where j ranges from 1 to... , The total number of lexical units in the first lexical feature sequence.

4. The AI-generated content detection method based on enhanced interactive intent as described in claim 1 or 2, characterized in that, The interaction intent dictionary includes five intent categories: assertion, commitment, instruction, declaration, and expression. The interaction intent dictionary is obtained using the following method: Obtain the seed word set for each intent category; The seed word set for each intent category is expanded to obtain an expanded candidate word set; The expanded candidate word set is filtered to obtain the interactive intent dictionary.

5. A device for detecting AI-generated content based on enhanced interactive intent, characterized in that, The AI-generated content detection device based on enhanced interaction intent includes: The text acquisition module is used to acquire the first text and perform vectorization encoding on the first text to obtain the first word feature sequence of the first text. The first text is the text that needs to be detected by AI-generated content. The first intent enhancement module is used to perform word-level interactive intent enhancement on the first word feature sequence based on the interactive intent dictionary to obtain the first intent feature vector; The intent dictionary acquisition module is used to acquire the set of category intent words corresponding to each intent category in the interaction intent dictionary; The second intent enhancement module is used to perform category-level interactive intent enhancement on the first lexical feature sequence according to the category intent word set of each category, so as to obtain the second intent feature vector; The fusion module is used to fuse the first intent feature vector and the second intent feature vector to obtain a third intent feature vector; The prediction module is used to predict the predicted category of the first text based on the third intent feature vector; The prediction categories include AI-generated text, human-generated text, text originally written by humans but rewritten by AI, and text originally written by AI but rewritten by humans. After vectorizing the first text, a text-level feature vector and a self-attention weight vector of the first text are obtained. The first intent enhancement module is further used for: The interaction intent score of each word feature in the first word feature sequence is calculated based on the interaction intent dictionary to obtain the interaction intent weight vector of the first word feature sequence. A learnable gating mechanism is used to fuse the interaction intent weight vector and the self-attention weight vector to obtain a fused weight vector; The first word feature sequence is enhanced using the fusion weight vector to obtain the second word sequence; The text-level feature vector is fused with the weighted pooled second word sequence to obtain the first intent feature vector; The second intent enhancement module is also used for: For any word feature in the first word feature sequence, the category intent weight of the word feature in the c-th intent category is calculated based on the category intent word set of the c-th intent category. Based on the category intent weight of each lexical feature in the c-th intent category, the first lexical feature sequence is weighted and summed to obtain the global intent feature of the c-th intent category; The global intent features of each intent category are concatenated and then processed through a linear layer mapping to obtain the second intent feature vector.

6. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores at least one computer program, which is loaded and executed by the processor to implement the AI-generated content detection method based on interactive intent enhancement as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to implement the AI-generated content detection method based on enhanced interactive intent as described in any one of claims 1 to 4.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the AI-generated content detection method based on enhanced interactive intent as described in any one of claims 1 to 4.

Images