Headline party content detection method based on self-enhancement prompt adjustment

Abstract

The invention belongs to the field of natural language processing and social media analysis, and relates to a title party content detection method based on self-enhancement prompt adjustment. The method comprises the following steps of: firstly, based on a prompt learning normal form, reconstructing a detection task into a complete blank-filling type mask prediction task by constructing an initial label word library and a classification prompt template, and finely tuning a pre-training language model by utilizing a small number of labeled samples; then, the model is used for classifying massive non-label texts, and candidate label words with high confidence are mined; performing semantic filtering on the candidate words based on consistency, diversity and purity indexes, and dynamically injecting high-quality new words into a label word library; and finally, carrying out iterative training by utilizing the enhanced label word bank. Through knowledge self-iteration and semantic self-adaption of the model, the problem of insufficient semantic coverage caused by static fixation of a tag word bank in a traditional prompt learning method is solved, and the title party content detection accuracy under the condition of few samples is improved.

Description

Technical Field

[0001] This invention relates to the fields of natural language processing and social media analysis, specifically to a method for detecting clickbait content based on self-enhancing prompts. Background Technology

[0002] With the rapid development of the internet and mobile information technology, news portals, social media, and short video platforms have seen a surge in textual information. To attract user attention and increase click-through rates, some content creators employ exaggeration, manipulation, and suspense to create "clickbait" content. These titles often contradict the main text or are intentionally misleading, leading to a negative user experience and even hindering the healthy development of online public opinion. Therefore, accurately and intelligently detecting and identifying clickbait content has become a key technological requirement in content moderation and internet governance.

[0003] Currently, existing methods for detecting clickbait content mainly fall into the following categories: One is methods based on manual feature engineering. These methods rely primarily on manual rules or experience, extracting features such as vocabulary, grammar, and sentiment from the text, and then using traditional machine learning models for classification. Although simple to implement and highly interpretable, this method is highly dependent on feature design and has limited generalization ability.

[0004] Second, there are detection methods based on deep neural networks. Pre-trained language models such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and BERT can automatically learn the latent semantic features of text, significantly improving detection performance. However, these methods typically require large amounts of high-quality labeled data, resulting in high training costs, and the model's decision-making process is difficult to interpret. Furthermore, because these models are not specifically designed to handle the diverse features of clickbait headlines, their recognition capabilities are prone to decline in complex real-world scenarios.

[0005] Thirdly, there are emerging methods such as Prompt-tuning. This method cleverly transforms the downstream classification task into a fill-in-the-blank task that aligns with the pre-training objectives of a large language model. It achieves excellent performance in scenarios requiring only a small number of labeled samples, significantly reducing reliance on labeled data. In this paradigm, the Verbalizer component is responsible for mapping the model's output to specific task labels. However, whether manually defined or generated through a fixed strategy, the label lexicon, once constructed, is typically fixed throughout the model's lifecycle. This static lexicon cannot capture the rich and dynamically changing semantic concepts of a specific domain, directly limiting the model's performance ceiling and adaptability. Summary of the Invention

[0006] To address the aforementioned problems, this invention provides a method for detecting clickbait content based on self-enhancing prompt adjustments, specifically including the following steps:

[0007] S1. Construct an initial detection space based on prompt adjustment: Construct an initial label lexicon (Verbalizer) and classification prompt templates. Based on the prompt learning paradigm, the clickbait detection task is transformed from a classification task into a cloze prediction task. Based on a small number of pre-labeled samples, the pre-trained language model is fine-tuned using the initial label lexicon and classification prompt templates to obtain the initial detection model.

[0008] S2. Self-enhancing candidate word mining based on generative prompts: The initial detection model is used to classify and predict unlabeled text samples, and high-confidence samples are selected based on the uncertainty of the prediction distribution; for the high-confidence samples, the model is stimulated to generate candidate labels describing their classification features using a preset generative prompt template, and a candidate word set is constructed.

[0009] S3. Multidimensional semantic index filtering and tag word library expansion: Calculate the consistency and confidence score, diversity score and purity score of each candidate tag word in the candidate word set, and select high-quality new tag words based on multidimensional indicators to form a new tag word set;

[0010] S4. Model Iteration and Optimization: Merge the newly added tag word set into the initial tag word library to build an enhanced tag word library; use the enhanced tag word library to expand the mapping function from tags to words, and combine the classification prompt template with a small number of labeled samples to iteratively fine-tune the model to obtain the final clickbait content detection model.

[0011] Furthermore, the step S1, which constructs an initial detection space based on cue adjustments, specifically includes:

[0012] S11. Constructing the initial tag vocabulary: For each target detection category y, manually define a set of initial, basic semantic description core seed words, y∈{clickbait content, non-clickbait content}, forming the core seed word set V. core Leveraging the generative capabilities of pre-trained language models, natural language prompt templates are constructed. These filled-in templates are then input into pre-trained language models such as the GPT series. Based on the probability distribution learned from large-scale corpora, a series of prompts related to V are generated. core Semantically related adjectives form an extended word set V. gen Using existing structured language knowledge bases such as CNKI and HowNet, synonyms, hypernyms, and hyponyms related to category names and core seed words are retrieved to form an extended vocabulary set V for the knowledge base. kb The above word sets are merged to form the initial tag lexicon V. seedEstablish a mapping relationship between categories and tags.

[0013] Specifically, construct the initial tag lexicon V seed The formula is:

[0014] V seed (y)=V core (y)∪V gen (y)∪V kb (y)={v1,v2,...,v n}

[0015] Where y represents the target classification category; v represents a single label word in the label word set; V seed (y) represents the initial tag vocabulary set corresponding to category y.

[0016] S12. Perform initial model training based on prompts: Transform the traditional classification task paradigm into a fill-in-the-blank task more suitable for the pre-trained language model. Based on a small number of labeled samples, fine-tune the parameters of the pre-trained language model by minimizing the cross-entropy loss function, and finally obtain an initial model M with basic detection capabilities. init .

[0017] Specifically, in step S12, for a given text sample x, it is filled into a classification hint template T containing the mask marker [MASK]. cls In (x), the probability p(y|x) that the model predicts for sample x to belong to category y is estimated as: the model uses the label vocabulary V of category y. seed The sum of the probabilities of all words in (y) filling the mask position [MASK] is calculated using the following formula:

[0018]

[0019] Where p(y|x) represents the posterior probability estimate of a given input text sample x belonging to a specific category y; x represents the original input text sequence to be detected; y represents the target classification category; v represents a single label word in the label word set; V seed (y) represents the set of tags corresponding to category y; T cls (x) represents the input sequence containing the masked [MASK] marker, formed by embedding the original text sample x into a predefined classification prompt template; p([MASK]=v|T) cls (x) represents the pre-trained language model receiving the templated input sequence T. cls After (x), predict the probability that the mask position [MASK] is the word v.

[0020] Furthermore, the S2 self-enhancing candidate word mining step based on generative prompts specifically includes:

[0021] S21. Filtering high-confidence samples based on the uncertainty of the predicted distribution: Input the unlabeled text samples into the initial text content detection model M. init To improve the reliability of classification, uncertainty assessment is introduced during the process to ensure the quality and stability of the self-learning process.

[0022] Furthermore, the operation of selecting high-confidence samples in step S21 is as follows: A random Dropout mechanism is introduced during model inference; for each unlabeled text sample x... i Perform m independent forward propagations to obtain m prediction result distributions; calculate the information entropy H(x) of the prediction result distributions. i ), and select samples with information entropy lower than a preset threshold as high-confidence samples.

[0023] Specifically, the formula for calculating information entropy in step S21 is as follows:

[0024]

[0025] Wherein, H(x) i ) represents the i-th unlabeled text sample x i The normalized information entropy value of the distribution of the prediction results; x i represents the i-th text sample to be analyzed in the unlabeled dataset; C represents the total number of categories in the classification task, where C=2 in clickbait content detection. In representing sample x i The average predicted probability of belonging to class j in m independent forward propagations.

[0026] S22. Using prompt templates to stimulate the model to generate candidate tags: Based on the classification results of high-confidence samples, using preset prompt templates, drive the initial text content detection model M. init Generate candidate tags describing the corresponding category, forming a candidate word set.

[0027] Furthermore, the operation of generating candidate label words in step S21 is as follows: Define a generation prompt template T to guide the model to explain the reason for classification. gen The template contains descriptive positions for filling in the blanks; based on the classification results of high-confidence samples, the corresponding category's prompt template is selected, driving the model to generate K descriptive words with the highest probability at the [MASK] position, which are used as descriptive feature words for that sample and stored in the candidate word set P. cand .

[0028] Furthermore, step S3 specifically includes: (1) consistency screening, which ensures that the selected words are closely related to the model's high-confidence judgment; (2) diversity screening, which avoids highly redundant semantically added label words; and (3) purity screening, which ensures that the selected words have a clear category orientation. Through the above screening, a small group of high-quality new label words V is finally obtained. new .

[0029] Specifically, the formula for calculating the consistency score in step S3 is as follows:

[0030] Score conf (v)=Freq(v)·Avg conf (v)

[0031] Among them, Score conf (v) represents the consistency score of candidate tag word v; v represents the score from the candidate word set P. cand The candidate tag word to be evaluated is any one of the candidate words P; Freq(v) represents the candidate word v in the entire candidate word set P. cand Frequency of occurrence in; Avg conf (v) represents the average confidence level of all samples described by word v in the classification. Score is retained. conf (v) Words that exceed a preset threshold.

[0032] Furthermore, the calculation process for the diversity score in step S3 is as follows:

[0033] Using the pre-trained Chinese language model bert-base-chinese as a feature extractor, high-dimensional word vector representations of each candidate word are extracted. The K-Means algorithm is used to dynamically set the number of clusters K based on the total number of candidate words, and K cluster centers are initialized. The Euclidean distance between each candidate word vector and each cluster center is calculated, and the candidate word is assigned to the cluster corresponding to the nearest cluster center. The average value of each cluster is recalculated to update the cluster centers, and this process is repeated iteratively until the cluster centers converge, resulting in K semantic clusters. Within each cluster, the candidate word v is calculated. i With the newly selected set of tags V new The word v in j The maximum semantic similarity R between r (v i ,v j ), preferentially select the cluster center with R closest to it. r Words with lower values ​​are used as new tag words to ensure the non-duplication and semantic diversity of the tag word library.

[0034] Specifically, the formula for calculating the diversity score in step S3 is as follows:

[0035]

[0036] Among them, R r (v i ,v j ) indicates the current candidate word v i With the selected set V new Maximum cosine similarity of all words in the text; v i This indicates that candidate words are currently being evaluated for inclusion in the tag library; v j This indicates that it has been selected into the new tag set V. new Any word in the vector; φ(v) represents the word vector representation of word v; φ(v) i )·φ(v j ) denotes the dot product operation of vectors; ‖φ(v)‖ denotes the magnitude of vector φ(v).

[0037] Specifically, the formula for calculating the purity score in step S3 is as follows:

[0038]

[0039] Among them, Score purity_pos (v) represents the purity score of word v relative to the positive category of "clickbait content"; Count pos (v) indicates the number of times the word v appears in the set categorized as "clickbait content"; Count neg (v) indicates the number of times word v appears in the "Non-clickbait content" category. Only words with a purity score close to 1 or 0 are retained to ensure that newly added tags have a clear category orientation.

[0040] Furthermore, the screening criteria include at least one of consistency criteria, diversity criteria, and purity criteria.

[0041] Furthermore, step S4 specifically includes: selecting the set of newly added tags V... new With the initial tag lexicon V seed Perform a union operation to obtain the enhanced tag vocabulary V. aug =V seed ∪V new ; Utilizing the enhanced tag lexicon V aug The label mapping space for the cloze test task is updated. Using the previously described classification hint template and a small number of labeled samples, the model is iteratively fine-tuned by minimizing the cross-entropy loss function to obtain the final clickbait content detection model M. final .

[0042] The beneficial effects of this invention are:

[0043] 1. High Adaptability and Evolutionary Capability: This invention enables the model to dynamically adapt to the rapid changes in online discourse by learning from unlabeled data. This overcomes the poor timeliness of traditional static models and significantly improves the robustness of the detection system in real and ever-changing environments.

[0044] 2. Extremely low cost and outstanding efficiency: This invention requires only a very small number of initial labeled samples to start and can effectively utilize massive amounts of unlabeled data for self-enhancement. This solution significantly reduces reliance on expensive manual annotation, and substantially lowers the application threshold and maintenance costs of content detection technology.

[0045] 3. Superior detection performance: By constructing a richer, data-driven enhanced tag lexicon, this invention enables the model to have a deeper understanding of the text, effectively breaking through the performance bottleneck of traditional prompt adjustment methods and improving the accuracy of clickbait content detection. Attached Figure Description

[0046] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0047] Figure 1 This is a flowchart of a clickbait content detection method based on self-enhancing prompt adjustment according to an embodiment of the present invention.

[0048] Figure 2 This is an overall architecture diagram of a clickbait content detection method based on self-enhancing prompt adjustment according to an embodiment of the present invention. Detailed Implementation

[0049] 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.

[0050] This invention proposes a method for detecting clickbait content based on self-enhancing prompts, which specifically includes the following steps:

[0051] S1. Construct an initial detection space based on prompt adjustment: Construct an initial label lexicon (Verbalizer) and classification prompt templates. Based on the prompt learning paradigm, the clickbait detection task is transformed from a classification task into a cloze prediction task. Based on a small number of pre-labeled samples, the pre-trained language model is fine-tuned using the initial label lexicon and classification prompt templates to obtain the initial detection model.

[0052] S2. Self-enhancing candidate word mining based on generative prompts: The initial detection model is used to classify and predict unlabeled text samples, and high-confidence samples are selected based on the uncertainty of the prediction distribution; for the high-confidence samples, the model is stimulated to generate candidate labels describing their classification features using a preset generative prompt template, and a candidate word set is constructed.

[0053] S3. Multidimensional semantic index filtering and tag word library expansion: Calculate the consistency and confidence score, diversity score and purity score of each candidate tag word in the candidate word set, and select high-quality new tag words based on multidimensional indicators to form a new tag word set;

[0054] S4. Model Iteration and Optimization: Merge the newly added tag word set into the initial tag word library to build an enhanced tag word library; use the enhanced tag word library to expand the mapping function from tags to words, and combine the classification prompt template with a small number of labeled samples to iteratively fine-tune the model to obtain the final clickbait content detection model.

[0055] like Figures 1-2 In this embodiment, a method for detecting clickbait content based on self-enhancing prompts includes the following steps:

[0056] S1. Constructing an initial detection space based on cue-based adjustment: This step aims to quickly construct an initial model M with basic detection capabilities using a small number of labeled samples based on a cue-based learning paradigm. init .

[0057] S11. Constructing an initial tag library: In this embodiment, for "clickbait content" (category y) pos ) and "non-clickbait content" (category y) neg ), respectively construct the initial tag vocabulary V seed (y pos ) and V seed (y neg The construction process specifically includes the following three sub-steps:

[0058] (1) Core seed word set V core Combining domain expert knowledge, a set of core seed keywords is manually selected for both clickbait and non-clickbait content; for example, y pos The category is defined by core seed words such as "shocked", "surprisingly", and "astonished", which are y neg The core seed words for the category definition are "report", "point out", "announce", etc.

[0059] (2) Generate an extended vocabulary set V genLeveraging the generative capabilities of pre-trained language models, a natural language prompt template is constructed. The filled-in template is then input into a pre-trained language model, the GPT series, which, based on its probability distribution learned from large-scale corpora, generates a series of prompts related to V. core Semantic-related adjectives. Specifically, given the input Prompt "Please list adjectives that are semantically similar to 'shocking' and are commonly used in exaggerated headlines", the GPT model outputs words such as "outrageous" and "scary".

[0060] (3) Knowledge Base Extended Vocabulary V kb Retrieve synonyms, hypernyms, and hyponyms related to category names or core seed words from the language knowledge base HowNet.

[0061] The above word sets are merged to form the initial tag word library V. seed Establish the mapping relationship between categories and tags, using the following formula:

[0062] V seed (y)=V core (y)∪V gen (y)∪V kb (y)={v1,v2,...,v n}

[0063] Where y represents the target classification category; v represents a single label word in the label word set; V seed (y) represents the initial tag vocabulary set corresponding to category y.

[0064] S12, Task Transformation and Initial Model Training: Based on the prompt learning paradigm, the clickbait detection task is transformed from a classification task into a cloze test prediction task. Define the classification prompt template T. cls (x) fills the input text x into a template containing the mask marker [MASK].

[0065] In this embodiment, the template is designed as follows:

[0066] T cls (x) = "Title: 'x', looks like [MASK]."

[0067] For example, if the input text x is "A man who eats this food will live forever", then the sequence of input to the model will be "Title: 'A man who eats this food will live forever', looks like [MASK]".

[0068] Then, based on a small number of pre-labeled samples, the pre-trained language model is fine-tuned. In this embodiment, the total number of samples is 50, with 25 positive and 25 negative samples.

[0069] Given a text sample x, populate it into a classification hint template T that includes the mask marker [MASK]. cls In (x), the probability p(y|x) that the model predicts for sample x to belong to category y is estimated as: the model uses the label vocabulary V of category y. seed The sum of the probabilities of all words in (y) filling the mask position [MASK] is calculated using the following formula:

[0070]

[0071] Where p(y|x) represents the posterior probability estimate of a given input text sample x belonging to a specific category y; x represents the original input text sequence to be detected; y represents the target classification category; v represents a single label word in the label word set; V seed (y) represents the set of tags corresponding to category y; T cls (x) represents the input sequence containing the masked [MASK] marker, formed by embedding the original text sample x into a predefined classification prompt template; p([MASK]=v|T) cls (x) represents the pre-trained language model receiving the templated input sequence T. cls After (x), predict the probability that the mask position [MASK] is the word v.

[0072] Based on this probability definition and the small number of labeled samples prepared in S11, the pre-trained language model is fine-tuned using the standard cross-entropy loss function. After training, an initial text content detection model M with preliminary clickbait detection capabilities is obtained. init .

[0073] S2. Self-enhancing candidate word mining based on generative prompts: This step aims to leverage the existing capabilities of M... init The system performs classification and prediction on a large amount of unlabeled text, filters high-confidence samples, and then uses generated prompt templates to drive the model to generate candidate labels.

[0074] S21. Filtering high-confidence samples based on the uncertainty of the predicted distribution: Obtain a large-scale unlabeled text dataset D from internet news portals, social media, and other channels. unlabeled Such as WeChat official accounts, NetEase News, Baidu Baijiahao, and Internet content farms.

[0075] A randomized Dropout mechanism is introduced during model inference to process each unlabeled text sample x. i Perform m independent forward propagations (m = 10 in this embodiment) to obtain m prediction result distributions; calculate the information entropy H(x) of the prediction result distributions. i), and select samples with information entropy below a preset threshold as high-confidence samples, calculated using the following formula:

[0076]

[0077] Wherein, H(x) i ) represents the i-th unlabeled text sample x i The normalized information entropy value of the distribution of the prediction results; x i represents the i-th text sample to be analyzed in the unlabeled dataset; C represents the total number of categories in the classification task, where C=2 in clickbait content detection. In representing sample x i The average predicted probability of belonging to class j in m independent forward propagations.

[0078] Set a preset threshold (H in this embodiment) threshold =0.1), select H(x) i ) <H threshold The samples are considered high-confidence samples. These samples are considered by the model to have a high degree of certainty regarding their classification results.

[0079] S22. Utilize prompt templates to stimulate the model to generate candidate labels: For each high-confidence sample selected in step S21, based on its classification result (category y)... pos or y neg Select the corresponding prompt template. This template is used to guide the model in explaining the reasons for classification.

[0080] In this embodiment, for the clickbait title category, template T gen The template for "Title 'x' is like clickbait because its wording is very [MASK]" is: "The reason title 'x' is like clickbait is because its wording is very [MASK]". For non-clickbait categories, template T... gen The explanation given was: "The headline 'x' is objective news because its wording is very [MASK]."

[0081] Using prompt templates to drive the initial model M init At the [MASK] position, generate K descriptive words with the highest probability (K=5 in this embodiment) as descriptive feature words of the sample and store them in the candidate word set P. cand .

[0082] S3. Multidimensional semantic index filtering and tag lexicon expansion: This step aims to filter the candidate word set P. cand For each candidate word w, multiple criteria are applied for comprehensive scoring and screening to ensure the quality of the final selected new tag words, ultimately constructing a new tag word library V. new .

[0083] S31. Calculate the consistency score: This criterion ensures that the selected words are closely related to the model's high-confidence judgments.

[0084] The consistency score of candidate word v is calculated using the following formula. conf (v):

[0085] Score conf (v)=Freq(v)·Avg conf (v)

[0086] Among them, Score conf (v) represents the consistency score of candidate tag word v; v represents the score from the candidate word set P. cand The candidate tag word to be evaluated is any one of the candidate words P; Freq(v) represents the candidate word v in the entire candidate word set P. cand Frequency of occurrence in; Avg conf (v) represents the average confidence level of all samples described by word v in the classification. Score is retained. conf (v) Words that exceed a preset threshold.

[0087] S32. Calculate the diversity score: This standard avoids highly redundant semantically added tags. It specifically includes the following three steps:

[0088] (1) Word vector representation: Using the pre-trained Chinese language model bert-base-chinese as the feature extractor, for each Chinese candidate tag word v that has passed the consistency screening in the candidate word set... i The [CLS] and [SEP] tags are added to encapsulate it into a model input format. The processed sequence is then input into a pre-trained language model. All token vectors output from the last hidden state of the model are extracted, and their average value (Mean Pooling) is calculated to obtain the candidate word v. i High-dimensional dense word vector representation φ(v i ).

[0089] (2) K-Means clustering algorithm: The K-Means clustering algorithm is used to cluster these word vectors. In order to adapt to candidate word sets of different sizes, the number of clusters K is dynamically set based on the total number of candidate words N. In this embodiment, Alternatively, a fixed value can be set based on actual computing resources; from the candidate word vector set {φ(v1),φ(v2),...,φ(v... N K vectors are randomly selected from the subset {μ1, μ2, ..., μ} as the initial cluster centers. K}; Calculate the vector φ(v) for each candidate word. i ) and each cluster center μ j The Euclidean distance between them is calculated using the following formula:

[0090] dist(φ(v i ),μ j )=‖φ(v i )-μ j ||2

[0091] Where ||·||2 represents the L2 norm of the vector. Each candidate word is assigned to the cluster C corresponding to its nearest cluster center. j For each cluster C j Calculate the mean of all word vectors within this cluster, and use it as the new cluster center μ′. j The formula is:

[0092]

[0093] Repeat the above steps until the change in cluster centers is less than the preset convergence threshold ∈ or the maximum number of iterations is reached, finally outputting K semantic clusters. In this embodiment, the convergence threshold ∈ = 1 × 10 -4 The maximum number of iterations is 300.

[0094] (3) Calculate diversity score: In each cluster, calculate candidate word v i With the newly selected set of tags V new The word v in j The maximum semantic similarity R between r (v i ,v j ), preferentially select the cluster center with R closest to it. r Words with lower values ​​are used as new tag words to ensure the non-duplication and semantic diversity of the tag word library.

[0095] Specifically, the formula for calculating the diversity score in step S3 is as follows:

[0096]

[0097] Among them, R r (v i ,v j ) indicates the current candidate word v i With the selected set V new Maximum cosine similarity of all words in the text; v i This indicates that candidate words are currently being evaluated for inclusion in the tag library; v j This indicates that it has been selected into the new tag set V. new Any word in the vector; φ(v) represents the word vector representation of word v; φ(v) i )·φ(v j ) denotes the dot product operation of vectors; ‖φ(v)‖ denotes the magnitude of vector φ(v).

[0098] Prioritize selecting the cluster center with R closest to it. r Using words with lower values ​​as new tags ensures that the new tags can cover multiple different aspects of clickbait rhetoric.

[0099] S33. Calculate the purity score: Ensure that the selected words have a clear category orientation. Calculate the purity score (Score) of the candidate word w for its respective category using the following formula. purity_pos (v):

[0100]

[0101] Among them, Score purity_pos (v) represents the purity score of word v relative to the positive category of "clickbait content"; Count pos (v) indicates the number of times the word v appears in the set categorized as "clickbait content"; Count neg (v) indicates the number of times word v appears in the "non-clickbait content" category. Only words with a purity score close to 1 or 0 are retained to ensure that newly added tags have a clear category orientation, while ambiguous neutral words are removed.

[0102] Through the above screening, the final set of high-quality new tags V is obtained. new .

[0103] S4. Model Iteration and Optimization: This step aims to merge the newly added tag words with the initial tag word library to form an enhanced tag word library, and use this enhanced word library to retrain and continue to fine-tune the model, thereby obtaining a final model with superior performance.

[0104] S41. Construct an enhanced tag lexicon V aug : The set of newly added tags V selected in step S2 new Compared with the initial tag vocabulary V of step S1 seed The merging process will create a richer, more comprehensive, and data-driven enhanced tag library V. aug The merging method is the union of sets:

[0105] V aug (y)=V seed (y)∪V new (y)

[0106] S42. Train and deploy the final model M final Using the above enhanced tag vocabulary V aug And use the classification prompt template T from step S12. cls (x) and a small number of labeled samples from step S11, for model M initAfter further fine-tuning, the final detection model M with significantly improved performance was obtained. final Because of the richness of the tag lexicon, the model is able to capture more diverse clickbait features, thus obtaining a high-performance text content detection model.

[0107] In embodiments of the present invention, steps S2 and S3 can constitute an iterative loop. M can be... final As the new M init The self-reinforcement process is repeated on the next round of unlabeled data, thus achieving continuous and automated model evolution. Ultimately, the optimal M-model will be determined. final Deployed to an online system for real-time detection of new web titles.

[0108] This invention aims to address the problems of high cost and insufficient adaptability of existing clickbait content detection technologies when facing dynamically changing online environments. Through the aforementioned implementation methods, this invention proposes an innovative, closed-loop, self-reinforcing framework. This framework efficiently starts an initial model with a small amount of labeled data. Then, leveraging the model's inherent consistency, it performs high-confidence pseudo-classification from massive amounts of unlabeled data, further driving the model to generate and extract high-quality new tags that reflect the current data distribution characteristics. Finally, by integrating this data-driven new knowledge into the model, it achieves simultaneous improvement in model performance and adaptability without relying on continuous manual annotation.

[0109] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting clickbait content based on self-enhancing prompts, characterized in that: The method includes the following steps: S1. Constructing an initial detection space based on prompts: Building an initial tag lexicon V seed Using classification prompt templates and a prompt learning paradigm, the clickbait detection task is transformed from a classification task into a cloze prediction task. Based on a small number of pre-labeled samples, the pre-trained language model is fine-tuned using an initial label lexicon and classification prompt templates to obtain the initial detection model M. init ; S2. Self-enhancing candidate word mining based on generative prompts: The initial detection model is used to classify and predict unlabeled text samples, and high-confidence samples are selected based on the uncertainty of the prediction distribution; for the high-confidence samples, the model is stimulated to generate candidate labels describing their classification features using a preset generative prompt template, and a candidate word set is constructed. S3. Multidimensional Semantic Index Filtering and Tag Lexicon Expansion: Calculate the consistency and confidence scores, diversity scores, and purity scores of each candidate tag in the candidate word set. Based on multidimensional indicators, select high-quality new tag words to form a new tag word set V. new ; S4. Model Iteration and Optimization: Add a new tag word set V new Merged into the initial tag lexicon V seed In the middle, construct an enhanced tag lexicon V aug ; Utilizing the enhanced tag lexicon V aug The model is iteratively fine-tuned by extending the tag-to-vocabulary mapping function (Verbalizer) and combining the aforementioned classification hint template with a small number of labeled samples to obtain the final clickbait content detection model M. final .

2. The method for detecting clickbait content based on self-enhancing prompt adjustment according to claim 1, characterized in that: The construction of the initial tag vocabulary in step S1 specifically includes: manually selecting a set of core seed words for both clickbait and non-clickbait content, forming a core seed word set V. core Utilizing the generative capabilities of pre-trained language models, natural language prompt templates are constructed. The filled-in templates are then input into the pre-trained language model (such as the GPT series). Based on the probability distribution learned from large-scale corpora, a series of prompts related to V are generated. core Semantically related adjectives form an extended word set V. gen Using existing structured language knowledge bases (such as HowNet), we retrieve synonyms, hypernyms, and hyponyms related to category names and core seed words, forming an extended vocabulary set V for the knowledge base. kb The above word sets are merged to form the initial tag lexicon V. seed Establish the mapping relationship between categories and tags, using the following formula: V seed (y)=V core (y)∪V gen (y)∪V kb (y)={v1,v2,...,v n } Where y represents the target classification category; v represents a single label word in the label word set; V seed (y) represents the initial tag vocabulary set corresponding to category y.

3. The method for detecting clickbait content based on self-enhancing prompt adjustment according to claim 1, characterized in that: The step S1, which transforms the task into a cloze prediction task, specifically includes: for a given text sample x, filling it into a classification hint template T containing a mask marker [MASK]. cls In (x), the probability p(y|x) that the model predicts for sample x to belong to category y is estimated as follows: the model uses the tag vocabulary V of category y. seed The sum of the probabilities of all words in (y) filling the mask position [MASK] is calculated using the following formula: Where p(y|x) represents the posterior probability estimate of a given input text sample x belonging to a specific category y; x represents the original input text sequence to be detected; y represents the target classification category; v represents a single label word in the label word set; V seed (y) represents the set of tags corresponding to category y; T cls (x) represents the input sequence containing the masked [MASK] marker, formed by embedding the original text sample x into a predefined classification prompt template; p([MASK]=v|T) cls (x) represents the pre-trained language model receiving the templated input sequence T. cls After (x), predict the probability that the mask position [MASK] is the word v.

4. The method for detecting clickbait content based on self-enhancing prompt adjustment according to claim 1, characterized in that: Step S2, which involves filtering high-confidence samples based on the uncertainty of the predicted distribution, specifically includes: introducing a random Dropout mechanism during model inference, and for each unlabeled text sample x... i Perform m independent forward propagations to obtain m prediction result distributions; calculate the information entropy H(x) of the prediction result distributions. i ), and select samples with information entropy below a preset threshold as high-confidence samples, calculated using the following formula: Wherein, H(x) i ) represents the i-th unlabeled text sample x i The normalized information entropy value of the distribution of the prediction results; x i represents the i-th text sample to be analyzed in the unlabeled dataset; C represents the total number of categories in the classification task, where C=2 in clickbait content detection. In representing sample x i The average predicted probability of belonging to class j in m independent forward propagations.

5. The method for detecting clickbait content based on self-enhancing prompt adjustment according to claim 1, characterized in that: Step S2, which uses a prompt template to stimulate the model to generate candidate labels, specifically includes: defining a prompt template T for guiding the model to explain the reasons for classification. gen The template contains descriptive positions for filling in the blanks; based on the classification results of high-confidence samples, the corresponding category's prompt template is selected, driving the model to generate K descriptive words with the highest probability at the [MASK] position, which are used as descriptive feature words for that sample and stored in the candidate word set P. cand .

6. The method for detecting clickbait content based on self-enhancing prompt adjustment according to claim 1, characterized in that: The calculation of the consistency score in step S3 specifically includes: calculating the consistency score (Score) of candidate word v using the following formula. conf (v): Score conf (v)=Freq(v)·Avg conf (v) Among them, Score conf (v) represents the consistency score of candidate tag word v; v represents the score from the candidate word set P. cand Any candidate tag word to be evaluated is selected from the list; F req (v) represents the candidate word v in the entire candidate word set P. cand Frequency of occurrence in; Avg conf (v) represents the average confidence level of all samples described by word v in the classification; Keep Score conf (v) Words that exceed a preset threshold.

7. The method for detecting clickbait content based on self-enhancing prompt adjustment according to claim 1, characterized in that: The calculation of diversity score in step S3 specifically includes: using the pre-trained Chinese language model bert-base-chinese as a feature extractor to extract high-dimensional word vector representations for each candidate word; using the K-Means algorithm, dynamically setting the number of clusters K based on the total number of candidate words, and initializing K cluster centers; calculating the Euclidean distance between each candidate word vector and each cluster center, and assigning the candidate word to the cluster corresponding to the nearest cluster center; recalculating the average value of each cluster to update the cluster centers, repeating the iteration until the cluster centers converge, resulting in K semantic clusters; and in each cluster, calculating the candidate word v i With the newly selected set of tags V new The word v in j The maximum semantic similarity R between r (v i ,v j The calculation formula is: Among them, R r (v i ,v j ) indicates the current candidate word v i With the selected set V new Maximum cosine similarity of all words in the text; v i This indicates that candidate words are currently being evaluated for inclusion in the tag library; v j This indicates that it has been selected into the new tag set V. new Any word in the vector; φ(v) represents the word vector representation of word v; φ(v) i )·φ(v j ) denotes the dot product operation of vectors; ||φ(v)|| denotes the magnitude of vector φ(v); Within each cluster, the cluster closest to the cluster center and R is selected first. r Words with lower values ​​are used as new tag words to ensure the non-duplication and semantic diversity of the tag word library.

8. The method for detecting clickbait content based on self-enhancing prompt adjustment according to claim 1, characterized in that: The calculation of the purity score in step S3 specifically includes: calculating the purity score (Score) of candidate word v for its category using the following formula. purity_pos (v): Among them, Score purity_pos (v) represents the purity score of word v relative to the positive category of "clickbait content"; Count pos (v) indicates the number of times the word v appears in the set categorized as "clickbait content"; Count neg (v) indicates the number of times the word v appears in the set categorized as "non-clickbait content"; Only words with purity scores close to 1 or 0 are retained to ensure that newly added tags have a clear category orientation.

9. The method for detecting clickbait content based on self-enhancing prompt adjustment according to claim 1, characterized in that: Step S4 specifically includes: selecting the newly added tag set V... new With the initial tag lexicon V seed Perform a union operation to obtain the enhanced tag vocabulary V. aug ; Utilizing the enhanced tag lexicon V aug The label mapping space for the cloze test task is updated. Using the previously described classification hint template and a small number of labeled samples, the model is iteratively fine-tuned by minimizing the cross-entropy loss function to obtain the final clickbait content detection model M. final .

Images