A fine-grained clickbait strength detection method based on hierarchical hint tuning

Abstract

The application discloses a fine-grained clickbait strength detection method based on hierarchical prompt tuning, which comprises the following steps: 1) obtaining the title text to be detected to obtain the prediction score corresponding to each clickbait strength level; 2) calculating the prediction probability corresponding to each clickbait strength level according to the prediction score, and performing main task training based on a cross-entropy loss function; 3) constructing a hierarchical score meeting the hierarchical ordered relationship based on the prediction probability, and generating a hierarchical supervision signal according to a real label to impose a hierarchical consistency constraint on the model; 4) constructing a shared semantic representation, and constructing an endpoint perception discrimination branch and an intermediate level disambiguation branch on the shared semantic representation; 5) jointly training the model, and outputting the final clickbait strength level according to the prediction probability in the model inference stage. The application can effectively model the ordered relationship between the levels while keeping the discrete level boundaries, and improve the recognition accuracy and stability of clickbait titles with different strengths.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing, text content quality analysis, and information content review and recommendation, and in particular to a fine-grained clickbait strength detection method based on hierarchical prompt optimization. Background Technology

[0002] With the rapid development of news platforms, search engines, social media platforms, and content recommendation systems, clickbait headlines have become a significant issue affecting information quality and user experience. Clickbait typically attracts users with exaggerated, vague, misleading, or suspenseful headlines. While it may boost traffic and exposure in the short term, it reduces the credibility of information, weakens user trust in the platform, and interferes with the platform's ranking, filtering, and review mechanisms.

[0003] Existing clickbait detection methods mainly fall into two categories: binary classification and continuous regression. Binary classification methods typically only determine whether a title is clickbait, grouping titles with varying degrees of misleading intent into the same category. This makes it difficult to distinguish between mildly misleading and highly misleading titles, thus failing to meet the differentiated handling needs of platforms in various governance scenarios such as demotion, warnings, and blocking. While continuous regression methods can output more granular intensity scores, continuous scores lack clear semantic boundaries. For example, two samples with similar scores cannot be directly mapped to specific platform actions, resulting in a poor correspondence between model output and actual strategies.

[0004] In recent years, prompt tuning methods based on pre-trained language models have achieved good results in text classification tasks. These methods transform the downstream task into a cloze-like prediction task by constructing prompt templates and label word mappers, achieving good performance in scenarios with few samples. However, existing prompt tuning methods typically treat category labels as independent planar classification targets, failing to explicitly utilize the naturally existing ordered relationships between clickbait strength labels. This can easily lead to confusion between adjacent strength levels and insufficient recognition ability on extreme strength samples. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a fine-grained clickbait intensity detection method based on hierarchical cue tuning. Clickbait detection is modeled as an ordered discrete intensity prediction task. Based on the pre-trained language model cue tuning backbone, hierarchical consistency constraints, endpoint awareness discrimination mechanisms, and intermediate level disambiguation mechanisms are introduced. This effectively models the sequential relationship between levels while maintaining the discrete level boundaries, thereby improving the accuracy, stability, and operability of identifying clickbait titles of different intensities.

[0006] The objective of this invention is achieved as follows: a fine-grained click decoy strength detection method based on hierarchical cue tuning, comprising the following steps:

[0007] 1) Obtain the title text to be detected, and construct a prompt input sequence using a manual prompt template and a tag word mapper. Input the prompt input sequence into a pre-trained language model for semantic encoding to obtain the prediction score corresponding to each click decoy strength level.

[0008] 2) Calculate the prediction probability corresponding to each click decoy strength level based on the prediction score, and perform main task training based on the cross-entropy loss function to achieve discrete prediction of the click decoy strength level of the title text;

[0009] 3) Construct hierarchical scores that satisfy the hierarchical order relationship based on the predicted probability, and generate hierarchical supervision signals based on the real labels to apply hierarchical consistency constraints to the model to ensure that the prediction results satisfy the order relationship between the intensity levels of the click decoys;

[0010] 4) Obtain a shared semantic representation based on the prompt input sequence, and construct an endpoint-aware discrimination branch and an intermediate-level disambiguation branch on the shared semantic representation to enhance the ability to distinguish between extreme-level samples and adjacent intermediate-level samples;

[0011] 5) The model is jointly trained using the main task loss, hierarchical consistency loss, endpoint discrimination loss, and intermediate level disambiguation loss, and the final click decoy strength level is output based on the predicted probability during the model inference stage.

[0012] Furthermore, step 1) specifically includes:

[0013] Step 1.1) Fill the title text into the manual prompt template and combine it with the tag word mapper to form a prompt input sequence;

[0014] Step 1.2) After inputting the prompt input sequence into the pre-trained language model, a four-dimensional logit vector is obtained:

[0015] (1)

[0016] in, This represents the unnormalized prediction score for click decoy strength level i.

[0017] Furthermore, step 2) specifically includes:

[0018] Step 2.1) Calculate the predicted probability corresponding to each intensity level based on the Softmax function:

[0019] (2)

[0020] Where, pi This indicates the predicted probability that the title text is classified as intensity level i; i and l j Let i and j represent the unnormalized predicted scores corresponding to intensity level i and intensity level j, respectively; i and j are the intensity level indices, and i and j ∈ {0, 1, 2, 3}.

[0021] Step 2.2) The main task loss adopts the cross-entropy loss function, the expression of which is:

[0022] (3)

[0023] in, The function is an indicator function, taking the value 1 if the condition within the parentheses is true, and 0 otherwise; y is the actual click decoy strength level label for the title text; p i This indicates the predicted probability that the title text is classified as intensity level i.

[0024] Furthermore, step 3) specifically includes:

[0025] Step 3.1) Construct three hierarchical scores that satisfy a monotonic relationship based on the four-level intensity prediction probabilities (p0, p1, p2, p3):

[0026] (4)

[0027] Wherein, s1 indicates that the intensity level of the title text is higher than 0, s2 indicates that the intensity level of the title text is at least 2, and s3 indicates that the intensity level of the title text is 3.

[0028] Step 3.2) Construct the corresponding hierarchical binary supervision signal based on the actual click decoy strength level label y:

[0029] (5)

[0030] Where z1 represents the binary supervision signal indicating whether the true intensity level of the title text is higher than level 0, z2 represents the binary supervision signal indicating whether the true intensity level of the title text reaches level 2, and z3 represents the binary supervision signal indicating whether the true intensity level of the title text reaches level 3; the value is 1 when the corresponding condition is met, and 0 otherwise.

[0031] Step 3.3) Apply a binary cross-entropy constraint to the hierarchical scores to obtain the hierarchical consistency loss:

[0032] (6)

[0033] Where BCE(⋅) represents the binary cross-entropy loss function.

[0034] Furthermore, step 4) specifically includes:

[0035] Step 4.1) Construct a shared semantic representation: When the input sequence contains M mask positions, the predicted scores for each mask position are aggregated using mean pooling to obtain the aggregated representation:

[0036] (7)

[0037] Among them, l (m) This represents the prediction score corresponding to the m-th mask position;

[0038] The aggregated representation is mapped to a shared representation through a projection layer:

[0039] (8)

[0040] Among them, W proj is the projection layer parameter matrix, and h is the shared representation used by the endpoint discrimination branch and the intermediate level disambiguation branch;

[0041] Step 4.2) Constructing the endpoint-aware discrimination branch specifically includes: constructing binary endpoint labels based on the actual click decoy strength level label y.

[0042] (9)

[0043] Where y0 indicates whether the title text belongs to level 0, and y3 indicates whether the title text belongs to level 3;

[0044] Input the shared representation h into the endpoint discrimination header. and And define the endpoint discrimination loss as:

[0045] (10)

[0046] in, Used to determine whether the title text belongs to the lowest intensity level. Used to determine whether the title text belongs to the highest intensity level;

[0047] Step 4.3) Constructing the intermediate-level disambiguation branch specifically includes: constructing an intermediate-level discrimination task for samples whose true labels belong to {1,2}, and defining the intermediate-level binary labels as:

[0048] (11)

[0049] Input the shared representation h into the intermediate level discrimination head. And define the intermediate level disambiguation loss as:

[0050] (12)

[0051] Among them, intermediate level discrimination head Used to differentiate between mildly and moderately suggestive headlines, in order to reduce semantic confusion between adjacent intermediate levels.

[0052] Furthermore, step 5) specifically includes:

[0053] Step 5-1) Jointly train the model. The joint training objective function is:

[0054] (13)

[0055] Where, λ hier , λ ep and λ mid These are the weight coefficients corresponding to the hierarchical consistency loss, endpoint discrimination loss, and intermediate level disambiguation loss, respectively, used to balance the contribution of each loss term in model training;

[0056] Step 5-2) In the model inference stage, the final click decoy strength level is determined based on the predicted probability output by the main task. Its expression is:

[0057] (14)

[0058] in, This represents the final predicted click decoy strength level.

[0059] Compared with the prior art, the beneficial effects of the present invention are: (1) The present invention reconstructs the click decoy detection from the traditional binary classification or continuous score regression task into an ordered discrete intensity prediction task, which retains the clear semantic boundaries between levels and is more in line with the actual decision-making methods of the platform in hierarchical review, filtering and sorting;

[0060] (2) By constructing hierarchical scores and applying hierarchical consistency constraints, the present invention enables the model to explicitly satisfy the hierarchical order logic while learning each discrete level, effectively reducing the contradictory prediction results between different intensity levels.

[0061] (3) By introducing an endpoint-aware discrimination mechanism, this invention enhances the model’s ability to identify the boundary between the lowest and highest intensity levels, which helps to reduce the misjudgment of non-inducing titles as clickbait and the failure to detect high-intensity clickbait.

[0062] (4) By introducing an intermediate level disambiguation mechanism, the present invention provides auxiliary learning for the most easily confused adjacent intermediate levels, thereby improving the accuracy and stability of fine-grained click decoy strength detection.

[0063] (5) This invention realizes fine-grained click decoy strength modeling based on pre-trained language models and prompt optimization framework, taking into account parameter efficiency, generalization ability and practical deployment feasibility, and can be applied to scenarios such as network content review, information recommendation, search ranking and media platform content quality assessment. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0065] Figure 1 This is a general framework diagram of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0067] like Figure 1 The method for fine-grained click decoy strength detection based on hierarchical cue tuning, as shown, includes the following steps:

[0068] 1) Obtain the title text to be detected, and construct a prompt input sequence using a manual prompt template and a tag word mapper. Input the prompt input sequence into a pre-trained language model for semantic encoding to obtain the prediction score corresponding to each click decoy strength level.

[0069] Step 1.1) Fill the title text into the artificial prompt template and combine it with a tag word mapper to form a prompt input sequence. Addressing the problem that traditional clickbait detection methods struggle to reflect differences in the degree of inducement, this embodiment defines the clickbait detection task as a four-level ordered discrete classification task. Given title text, the goal is to predict its clickbait strength level label y∈{0,1,2,3}, where: level 0 represents a non-inducing title; level 1 represents a slightly induceing title; level 2 represents a moderately induceing title; and level 3 represents a highly induceing title.

[0070] This four-level label not only has discrete boundaries but also satisfies the ordered relation 0<1<2<3. This definition method allows the model output to directly correspond to the platform's governance actions under different tolerance conditions, such as retention, downgrading, prompting, or blocking.

[0071] Step 1.2) In this embodiment, the title text is first filled into a manually designed prompt template, and a prompt input sequence is constructed using a tag word mapper. The prompt input sequence contains one or more [MASK] mask tags and is input into a pre-trained language model for encoding. The pre-trained language model outputs a four-dimensional logit vector:

[0072] (1)

[0073] in, This represents the unnormalized prediction score for click decoy strength level i.

[0074] 2) Calculate the prediction probability corresponding to each click decoy strength level based on the prediction score, and perform main task training based on the cross-entropy loss function to achieve discrete prediction of the click decoy strength level of the title text;

[0075] Step 2.1) After obtaining the four-dimensional logit vector, calculate the predicted probability corresponding to each intensity level based on the Softmax function:

[0076] (2)

[0077] Where, p i This indicates the predicted probability that the title text is classified as intensity level i; i and l j Let i and j represent the unnormalized predicted scores corresponding to intensity level i and intensity level j, respectively; i and j are the intensity level indices, and i and j ∈ {0, 1, 2, 3}.

[0078] Step 2.2) To achieve the main task training of the click decoy strength level of the title text, the main task loss adopts the cross-entropy loss function, the expression of which is:

[0079] (3)

[0080] in, The function is an indicator function, taking the value 1 if the condition within the parentheses is true, and 0 otherwise; y is the actual click decoy strength level label for the title text; p i This indicates the predicted probability that the title text is classified as intensity level i.

[0081] Through this step, the model can complete the basic prediction of the intensity of the four-level discrete click decoy.

[0082] 3) Construct hierarchical scores that satisfy the hierarchical order relationship based on the predicted probability, and generate hierarchical supervision signals based on the real labels to apply hierarchical consistency constraints to the model to ensure that the prediction results satisfy the order relationship between the intensity levels of the click decoys;

[0083] Step 3.1) Because there is a natural ordered relationship between the intensity labels of the four levels of click decoys, using only ordinary four-class targets is prone to producing prediction results that violate hierarchical logic. Therefore, based on the four-level intensity prediction probabilities (p0, p1, p2, p3), three hierarchical scores that satisfy a monotonic relationship are constructed:

[0084] (4)

[0085] Wherein, s1 indicates that the intensity level of the title text is higher than 0, s2 indicates that the intensity level of the title text is at least 2, and s3 indicates that the intensity level of the title text is 3.

[0086] Step 3.2) Construct the corresponding hierarchical binary supervision signal based on the actual click decoy strength level label y:

[0087] (5)

[0088] Where z1 represents the binary supervision signal indicating whether the true intensity level of the title text is higher than level 0, z2 represents the binary supervision signal indicating whether the true intensity level of the title text reaches level 2, and z3 represents the binary supervision signal indicating whether the true intensity level of the title text reaches level 3; the value is 1 when the corresponding condition is met, and 0 otherwise.

[0089] Step 3.3) Apply a binary cross-entropy constraint to the hierarchical scores to obtain the hierarchical consistency loss:

[0090] (6)

[0091] Where BCE(⋅) represents the binary cross-entropy loss function.

[0092] Through this step, the model not only learns to distinguish specific levels, but also needs to satisfy the hierarchical logic of "at least reaching a certain level", thereby reducing contradictory predictions between different intensities.

[0093] 4) Obtain a shared semantic representation based on the prompt input sequence, and construct an endpoint-aware discrimination branch and an intermediate-level disambiguation branch on the shared semantic representation to enhance the ability to distinguish between extreme-level samples and adjacent intermediate-level samples;

[0094] In real-world platform scenarios, the lowest and highest intensity levels typically have the most significant operational implications: the lowest intensity level corresponds to clearly non-click bait content, while the highest intensity level corresponds to strongly enticing titles most likely to be penalized or removed. Meanwhile, mildly and moderately enticing titles often share similar linguistic styles, thus necessitating a greater ability to differentiate between extreme intensity samples and adjacent intermediate intensity samples.

[0095] Step 4.1) Construct a shared semantic representation: When the input sequence contains M mask positions, the predicted scores for each mask position are aggregated using mean pooling to obtain the aggregated representation:

[0096] (7)

[0097] Among them, l (m) This represents the prediction score corresponding to the m-th mask position;

[0098] The aggregated representation is mapped to a shared representation through a projection layer:

[0099] (8)

[0100] Among them, W proj is the projection layer parameter matrix, and h is the shared representation used by the endpoint discrimination branch and the intermediate level disambiguation branch;

[0101] Step 4.2) Constructing the endpoint-aware discrimination branch specifically includes: constructing binary endpoint labels based on the actual click decoy strength level label y.

[0102] (9)

[0103] Where y0 indicates whether the title text belongs to level 0, and y3 indicates whether the title text belongs to level 3;

[0104] Input the shared representation h into the endpoint discrimination header. and And define the endpoint discrimination loss as:

[0105] (10)

[0106] in, Used to determine whether the title text belongs to the lowest intensity level. Used to determine whether the title text belongs to the highest intensity level;

[0107] Through the aforementioned endpoint perception and discrimination mechanism, the model can more accurately distinguish between non-inducing titles and high-intensity clickbait titles, thereby improving boundary recognition capabilities.

[0108] Step 4.3) Constructing the intermediate-level disambiguation branch specifically includes: constructing an intermediate-level discrimination task for samples whose true labels belong to {1,2}, and defining the intermediate-level binary labels as:

[0109] (11)

[0110] Input the shared representation h into the intermediate level discrimination head. And define the intermediate level disambiguation loss as:

[0111] (12)

[0112] Among them, intermediate level discrimination head Used to differentiate between mildly and moderately suggestive headlines, in order to reduce semantic confusion between adjacent intermediate levels.

[0113] This step allows the model to receive additional supervision for the most easily confused intermediate adjacent levels, thereby improving its fine-grained discrimination ability.

[0114] 5) The model is jointly trained using the main task loss, hierarchical consistency loss, endpoint discrimination loss, and intermediate level disambiguation loss, and the final click decoy strength level is output based on the predicted probability during the model inference stage.

[0115] Step 5-1) Jointly train the model. The joint training objective function is:

[0116] (13)

[0117] Where, λ hier , λ ep and λ mid These are the weight coefficients corresponding to the hierarchical consistency loss, endpoint discrimination loss, and intermediate level disambiguation loss, respectively, and their contribution relationships to balancing the main task prediction, level order preservation, endpoint discrimination enhancement, and intermediate level disambiguation.

[0118] Through joint training, the model can learn both the global hierarchical structure and local boundary features simultaneously, thereby improving the overall prediction robustness.

[0119] Step 5-2) In the model inference stage, the final click decoy strength level is determined based on the four-level strength prediction probability, and its expression is:

[0120] (14)

[0121] in, This represents the final predicted click decoy strength level.

[0122] Click decoy strength levels include: Level 0 non-decorative title, Level 1 slightly decoy title, Level 2 moderately decoy title, and Level 3 highly decoy title.

[0123] Among them, level 0 headings adopt neutral, objective, and factual expressions, and do not contain obvious exaggeration, emotional arousal or information gap patterns;

[0124] Level 1 headings contain slightly appealing expressions, but do not use highly emotional vocabulary or dramatic assertions;

[0125] Level 2 headings contain attention-grabbing expressions and use exaggeration, suspense, or missing information to entice users to click;

[0126] Level 3 headings contain strong emotional incitement, absolute statements, panic threats, or imperative expressions, and usually have both strong emotional and information gap characteristics.

[0127] This design keeps the inference process simple and efficient, making it easy to deploy quickly in real-world platform scenarios. Meanwhile, the auxiliary objectives introduced during the training phase can significantly improve the reliability and stability of the inference results.

[0128] To verify the effectiveness of the proposed fine-grained clickbait strength detection method based on hierarchical cue tuning, experiments were conducted on multiple Chinese clickbait datasets. This embodiment selects four clickbait title datasets from different online platforms: Tencent, Zongxiang, Wangyi, and WeChat. These datasets all originate from news titles or information published on real online platforms, and the clickbait strength of each title is graded through manual annotation.

[0129] During the data annotation process, several annotators with extensive experience in using online media were invited to participate in the annotation work, and the titles were graded according to a pre-defined standard for the intensity of clickbait. The intensity of clickbait was divided into four levels: level 0, level 1, level 2, and level 3, with higher level values ​​indicating a stronger degree of clickbait inducement in the title.

[0130] To ensure dataset quality, a portion of the samples were randomly selected for manual review, and an annotation consistency index was calculated. Experimental statistical results show that the labeled data exhibits high consistency and reliability. Dataset statistics are shown in Table 1.

[0131] Table 1. Dataset Statistics

[0132]

[0133] In the experiment, each dataset was divided into 60% training set, 20% validation set, and 20% test set, and a stratified sampling strategy was adopted to maintain a consistent distribution of labels for each category.

[0134] To comprehensively evaluate model performance, this embodiment uses Accuracy, Precision, Recall, and F1 as the main evaluation metrics. Furthermore, since click decoy strengths have a natural hierarchical structure, Mean Absolute Error (MAE) and Mean Squared Error (MSE) are also used to measure the degree of deviation between the predicted results and the true levels.

[0135] The Accuracy and F1 metrics are defined as follows:

[0136] (15)

[0137] (16)

[0138] in, (17)

[0139] (18)

[0140] Where TP represents the number of titles correctly predicted as clickbait titles by the algorithm; FP represents the number of titles predicted as clickbait titles but actually being non-clickbait titles; FN represents the number of titles predicted as non-clickbait titles but actually being clickbait titles; TN represents the number of titles correctly predicted as non-clickbait titles; and N represents the total number of test samples.

[0141] To verify the performance of the method of the present invention, several representative methods were selected as comparative baselines in the experiment, including:

[0142] (1) Hierarchical text classification method: HiEPS method;

[0143] (2) Order relation modeling method: MLL method;

[0144] (3) Click decoy detection method based on cue learning: PT-CD method;

[0145] (4) Large language model methods: GPT-4o model and DeepSeek-v3 model.

[0146] All the above methods were tested under the same dataset partitioning and experimental environment to ensure the fairness of the experimental results.

[0147] The experimental results of each method on the four datasets are shown in Table 2:

[0148]

[0149] As shown in Table 2, the fine-grained click decoy strength detection method based on hierarchical cue tuning proposed in this invention significantly outperforms other comparative methods in terms of accuracy and F1 scores on four datasets.

[0150] The reason is that, based on the traditional cue learning framework, the method of this invention introduces hierarchical consistency constraints, endpoint discrimination mechanisms, and intermediate level disambiguation mechanisms, enabling the model to better learn the ordered relationship between click decoy strength levels, thereby improving the ability to distinguish between different strength levels.

[0151] Experimental results show that the method of the present invention has higher detection accuracy and better stability in click decoy strength detection tasks, and can effectively support content filtering and sorting applications in actual platforms.

[0152] This invention proposes a fine-grained clickbait strength detection method based on hierarchical cue optimization. By constructing a cue learning framework and combining it with a pre-trained language model, semantic representation learning is performed on the input title text, enabling the prediction of clickbait strength levels. Building upon traditional cue learning models, this invention introduces hierarchical consistency constraints to characterize the ordered relationship between clickbait strength levels. Simultaneously, an endpoint-aware discrimination mechanism enhances the model's ability to identify extreme-level samples, and an intermediate-level disambiguation mechanism improves the distinguishability between adjacent levels. Through these methods, this invention significantly improves the accuracy of fine-grained clickbait strength detection while maintaining prediction efficiency. Experimental results show that the proposed method outperforms existing methods on multiple datasets, demonstrating its application value in network content quality analysis and information filtering tasks.

[0153] The above description of the embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principles of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A fine-grained click decoy strength detection method based on hierarchical cue tuning, characterized in that, Includes the following steps: 1) Obtain the title text to be detected, and construct a prompt input sequence using a manual prompt template and a tag word mapper. Input the prompt input sequence into a pre-trained language model for semantic encoding to obtain the prediction score corresponding to each click decoy strength level. 2) Calculate the prediction probability corresponding to each click decoy strength level based on the prediction score, and perform main task training based on the cross-entropy loss function to achieve discrete prediction of the click decoy strength level of the title text; 3) Construct hierarchical scores that satisfy the hierarchical order relationship based on the predicted probability, and generate hierarchical supervision signals based on the real labels to apply hierarchical consistency constraints to the model to ensure that the prediction results satisfy the order relationship between the intensity levels of the click decoys; 4) Obtain a shared semantic representation based on the prompt input sequence, and construct an endpoint-aware discrimination branch and an intermediate-level disambiguation branch on the shared semantic representation to enhance the ability to distinguish between extreme-level samples and adjacent intermediate-level samples; 5) The model is jointly trained using the main task loss, hierarchical consistency loss, endpoint discrimination loss, and intermediate level disambiguation loss, and the final click decoy strength level is output based on the predicted probability during the model inference stage.

2. The fine-grained click decoy strength detection method based on hierarchical cue tuning according to claim 1, characterized in that, Step 1) specifically includes: Step 1.1) Fill the title text into the manual prompt template and combine it with the tag word mapper to form a prompt input sequence; Step 1.2) After inputting the prompt input sequence into the pre-trained language model, a four-dimensional logit vector is obtained: （1）； in, This represents the unnormalized prediction score for click decoy strength level i.

3. The fine-grained click decoy strength detection method based on hierarchical cue tuning according to claim 2, characterized in that, Step 2) specifically includes: Step 2.1) Calculate the predicted probability corresponding to each intensity level based on the Softmax function: （2）； Where, p i This indicates the predicted probability that the title text is classified as intensity level i; i and l j Let i and j represent the unnormalized predicted scores corresponding to intensity level i and intensity level j, respectively; i and j are the intensity level indices, and i and j ∈ {0, 1, 2, 3}. Step 2.2) The main task loss adopts the cross-entropy loss function, the expression of which is: （3）； in, The function is an indicator function, taking the value 1 if the condition within the parentheses is true, and 0 otherwise; y is the actual click decoy strength level label for the title text; p i This indicates the predicted probability that the title text is classified as intensity level i.

4. The fine-grained click decoy strength detection method based on hierarchical cue tuning according to claim 1, characterized in that, Step 3) specifically includes: Step 3.1) Construct three hierarchical scores that satisfy a monotonic relationship based on the four-level intensity prediction probabilities (p0, p1, p2, p3): （4）； Wherein, s1 indicates that the intensity level of the title text is higher than 0, s2 indicates that the intensity level of the title text is at least 2, and s3 indicates that the intensity level of the title text is 3. Step 3.2) Construct the corresponding hierarchical binary supervision signal based on the actual click decoy strength level label y: （5）； Where z1 represents the binary supervision signal indicating whether the true intensity level of the title text is higher than level 0, z2 represents the binary supervision signal indicating whether the true intensity level of the title text reaches level 2, and z3 represents the binary supervision signal indicating whether the true intensity level of the title text reaches level 3; the value is 1 when the corresponding condition is met, and 0 otherwise. Step 3.3) Apply a binary cross-entropy constraint to the hierarchical scores to obtain the hierarchical consistency loss: （6）； Where BCE(⋅) represents the binary cross-entropy loss function.

5. The method for fine-grained click decoy strength detection based on hierarchical cue tuning according to claim 1, characterized in that, Step 4) specifically includes: Step 4.1) Construct a shared semantic representation: When the input sequence contains M mask positions, the predicted scores for each mask position are aggregated using mean pooling to obtain the aggregated representation: （7）； Among them, l (m) This represents the prediction score corresponding to the m-th mask position; The aggregated representation is mapped to a shared representation through a projection layer: （8）； Among them, W proj is the projection layer parameter matrix, and h is the shared representation used by the endpoint discrimination branch and the intermediate level disambiguation branch; Step 4.2) Constructing the endpoint-aware discrimination branch specifically includes: constructing binary endpoint labels based on the actual click decoy strength level label y. （9）； Where y0 indicates whether the title text belongs to level 0, and y3 indicates whether the title text belongs to level 3; Input the shared representation h into the endpoint discrimination header. and And define the endpoint discrimination loss as: （10）； in, Used to determine whether the title text belongs to the lowest intensity level. Used to determine whether the title text belongs to the highest intensity level; Step 4.3) Constructing the intermediate-level disambiguation branch specifically includes: constructing an intermediate-level discrimination task for samples whose true labels belong to {1,2}, and defining the intermediate-level binary labels as: （11）； Input the shared representation h into the intermediate level discrimination head. And define the intermediate level disambiguation loss as: （12）； Among them, intermediate level discrimination head Used to differentiate between mildly and moderately suggestive headlines, in order to reduce semantic confusion between adjacent intermediate levels.

6. The method for fine-grained click decoy strength detection based on hierarchical cue tuning according to claim 1, characterized in that, Step 5) specifically includes: Step 5-1) Jointly train the model. The joint training objective function is: （13）； Where, λ hier , λ ep and λ mid These are the weight coefficients corresponding to the hierarchical consistency loss, endpoint discrimination loss, and intermediate level disambiguation loss, respectively, used to balance the contribution of each loss term in model training; Step 5-2) In the model inference stage, the final click decoy strength level is determined based on the predicted probability output by the main task. Its expression is: （14）； in, This represents the final predicted click decoy strength level.

Images