A clickbait detection method using only title for prompt learning
By expanding the tag word space using the RoBERTa-base model and five optimization strategies, this method addresses the issues of high sample requirements and limited detection performance in existing clickbait detection methods, achieving efficient clickbait detection with a prediction accuracy of 95%.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- YANGZHOU UNIV
- Filing Date
- 2023-08-23
- Publication Date
- 2026-05-29
Smart Images
Figure CN117033639B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of natural language processing research, and in particular to a clickbait detection method that uses only titles for cue learning. Background Technology
[0002] With the rapid development of the Internet, for most online services such as e-commerce, more clicks and web traffic mean more profits and business revenue. Therefore, many content producers and publishers use gimmicks and sensational news headlines to lure and deceive users into clicking, which is known as clickbait.
[0003] Currently, scholars both domestically and internationally have conducted extensive research on clickbait detection, including detection algorithms based on feature engineering, detection algorithms based on neural networks in deep learning, and detection algorithms based on pre-trained language models.
[0004] The three types of clickbait detection methods mentioned above have achieved some success, but they still fall short of current detection tasks. Feature engineering-based detection algorithms require expert knowledge for feature selection, and some manually generated features are limited in representing semantic information, thus affecting the final detection results. While deep learning-based neural network detection algorithms can effectively detect clicks, they require a large amount of labeled data for training, and in practical applications with limited training samples, they cannot accurately detect clickbait. Pre-trained language models have become powerful language understanding and generation tools, and their rich knowledge can be well applied to downstream tasks, but this usually requires auxiliary information to fine-tune the model. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a clickbait detection method that uses only titles for cue learning. This method can train the model with a small number of labeled samples and expand the label word space for cue learning using five optimization strategies, making the pre-trained model more suitable for the clickbait detection task and improving classification accuracy.
[0006] The objective of this invention is achieved as follows: a clickbait detection method that uses only titles for cue learning, comprising the following steps:
[0007] Step 1: Select a suitable pre-trained model and use it as the backbone to construct the tag words and templates for cue learning;
[0008] Step 2: Expand the category tag words using five optimization strategies to obtain the optimized tag word set;
[0009] Step 3: Using the expanded tags, transform the classification task into a probability calculation problem for category tags;
[0010] Step 4: Map the predicted probability of each tag to the corresponding category to obtain the final prediction score of the tag as the classification result.
[0011] Furthermore, step 1 specifically includes:
[0012] Step 1.1: Select the RoBERTabase model as the pre-trained language model. This model uses a dynamic masking strategy. Each time a sequence is input into the model, a new masking pattern is generated. As a large amount of data is continuously input, the model will gradually adapt to different masking strategies and learn different language representations.
[0013] Step 1.2: Use auxiliary tags [head] and [tail] to indicate the start and end of the template. Combine the template with the original input title to form the input of the model. The template is "[head][placeholder:x].In summary,it was[mask].[tail]", where "[placeholder:x]" is the position to fill the news title of the input.
[0014] Furthermore, step 2 specifically includes:
[0015] Step 2.1: In the expansion of tag words, the special category V associated with the specific tag. y V1 = {clickbait} can be expanded to V1 = {clickbait, hyperlink, misleading, ...};
[0016] Step 2.2: In the expansion of the tag terms, select the tag class names of both click decoys and non-click decoys as the initial words for concept search in Probase. Select N concepts from the probabilistic retrieval ranking results of Probase, denoted as N(v); then, calculate the distance dist(V) between the tag class name y in the embedding space and the expanded tag term set. y For all inputs x, y), i {N1,…,N i ,…,N n The list is a similarity ranking result obtained through concept retrieval, and the top M words are selected as the expanded tag word set.
[0017] Step 2.3: After obtaining the above tag word set, the tag words are further filtered using four optimization strategies, including using BERT to predict the probability distribution of expanded words filling the mask position, calculating the semantic similarity between the fastText vector of each class tag name and the fastText vector of the expanded tag word, based on word frequency provided by Wikipedia, and considering the contextual information of the mask position; for each strategy, N is selected. a =15 words. Finally, the words selected by the four strategies are combined to form the final extended word set.
[0018] Furthermore, step 3 specifically includes:
[0019] Step 3.1: Given a set of input data x = {x1, ..., x2} n They will all be classified into category y∈Y, and the set of category labels is denoted as V. y ={v1,…,v n}, where V is the set of the entire category, V y It is a subset of V; in the pre-trained language model M, the probability that each word v is filled into [MASK] is represented as P([MASK]=v∈V). y |x p Then, clicking on the decoy detection task can be transformed into the probability calculation problem of category tag words, as shown in equation (1).
[0020] P(y∈Y|x)=p([MASK]=v∈V y |x p (1);
[0021] Step 3.2: During the detection of click decoys, if the probability of belonging to y1 = "click decoy" is greater than the probability of y2 = "non-click decoy", then the text will be classified as "click decoy".
[0022] Furthermore, step 4 specifically includes:
[0023] Step 4.1: After calculating the probabilities of all category labels, map the predicted probabilities of each label to the corresponding category, assign a learnable weight parameter to each label, and then normalize it, as shown in equation (2).
[0024]
[0025] Where, α v w is the predicted probability of the class label v. v These are learnable weight parameters associated with label v, where u represents the set of all labels V under category y. y One of the tags, exp(wv ) indicates w v The exponential function.
[0026] Assuming that each word in the expanded vocabulary contributes equally to the predicted label, a simple average is performed, and the mean of the predicted scores is used as the predicted score for that label. Finally, the category with the highest predicted score is selected as the final prediction result, as shown in Equation (3).
[0027]
[0028] In operation, this invention selects a suitable pre-trained language model to construct label words and templates for cue learning. Five optimization strategies are used to optimize the label words in cue learning. Using the expanded label words, the classification task is transformed into a probability calculation problem of category label words. The input text and the preset cue template are combined to construct a cue text with a mask, which is used as the input to the model. Click decoy detection is performed using the optimized label words. Finally, the predicted probability of each label word is mapped to the corresponding category to obtain the final prediction score of the label as the classification result.
[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: First, the present invention makes full use of cues to predict labels, thereby improving the effectiveness of representation learning and feature extraction.
[0030] Secondly, the click decoy detection method of the present invention expands the tag word space of cue learning through various optimization strategies, which can effectively improve the detection effect of click decoys.
[0031] Second, the click decoy detection method of the present invention is the first to propose applying a cue learning model to click decoy detection, and its performance on three benchmark datasets surpasses existing baseline methods, with prediction results reaching 95%. Attached Figure Description
[0032] Figure 1 This is a general framework diagram of the click decoy detection method of the present invention. Detailed Implementation
[0033] like Figure 1 The click decoy detection method shown includes the following steps:
[0034] Step 1: Select a suitable pre-trained model and use it as the backbone to construct the tag words and templates for cue learning;
[0035] Step 2: Expand the category tag words using five optimization strategies to obtain the optimized tag word set;
[0036] Step 3: Using the expanded tags, transform the classification task into a probability calculation problem for category tags;
[0037] Step 4: Map the predicted probability of each label to its corresponding category to obtain the final predicted score for that label as the classification result. The specific implementation method is as follows:
[0038] The specific implementation of this method is carried out according to the following steps:
[0039] Step 1: Select the RoBERTabase model and use this trained model as the backbone to build the tag words and templates in the cue learning process;
[0040] Step 1.1: Select the RoBERTabase model as the pre-trained language model. This model uses a dynamic masking strategy. Each time a sequence is input into the model, a new masking pattern is generated. As a large amount of data is continuously input, the model will gradually adapt to different masking strategies and learn different language representations.
[0041] Step 1.2: Use auxiliary tags [head] and [tail] to indicate the start and end of the template. Combine the template with the original input title to form the model input. For example, when taking a news headline x: “For 13 years family hear storming sound in walls nightly as repairman in disbelief of finding” as input, the template is “[head][placeholder:x].In summary,it was[mask].[tail]”, where “[placeholder:x]” is the position to fill the news headline in the input.
[0042] Step 2: Expand the category tag words using five optimization strategies to obtain the optimized tag word set;
[0043] Step 2.1: In the expansion of tag words, the special category V associated with the specific tag. y V1 = {clickbait} can be expanded to V1 = {clickbait, hyperlink, misleading, ...};
[0044] Step 2.2: In the expansion of the tag terms, select the tag class names of both click decoys and non-click decoys as the initial words for concept search in Probase. Select N concepts from the probabilistic retrieval ranking results of Probase, denoted as N(v); then, calculate the distance dist(V) between the tag class name y in the embedding space and the expanded tag term set.y For all inputs x, y), i {N1,…,N i ,…,N n The list is a similarity ranking result obtained through concept retrieval, and the top M words are selected as the expanded tag word set.
[0045] Step 2.3: After obtaining the above tag word set, the tag words are further filtered using four optimization strategies, including using BERT to predict the probability distribution of expanded words filling the mask position, calculating the semantic similarity between the fastText vector of each class tag name and the fastText vector of the expanded tag word, based on word frequency provided by Wikipedia, and considering the contextual information of the mask position; for each strategy, N is selected. a =15 words. Finally, the words selected by the four strategies are combined to form the final extended word set.
[0046] Step 3: Using the expanded tags, transform the classification task into a probability calculation problem for category tags;
[0047] Step 3.1: Given a set of input data x = {x1, ..., x2} n They will all be classified into category y∈Y, and the set of category labels is denoted as V. y ={v1,…,v n}, where V is the set of the entire category, V y It is a subset of V; in the pre-trained language model M, the probability that each word v is filled into [MASK] is represented as P([MASK]=v∈V). y |x p Then, clicking on the decoy detection task can be transformed into the probability calculation problem of category tag words, as shown in equation (1).
[0048] P(y∈Y|x)=p([MASK]=v∈V y |x p (1);
[0049] Step 3.2: During the detection of click decoys, if the probability of belonging to y1 = "click decoy" is greater than the probability of y2 = "non-click decoy", then the text will be classified as "click decoy".
[0050] Step 4: Map the predicted probability of each tag to the corresponding category to obtain the final prediction score of the tag as the classification result;
[0051] Step 4.1: After calculating the probabilities of all category labels, map the predicted probabilities of each label to the corresponding category, assign a learnable weight parameter to each label, and then normalize it, as shown in equation (2).
[0052]
[0053] Where, α v w is the predicted probability of the class label v. v These are learnable weight parameters associated with label v, where u represents the set of all labels V under category y. y One of the tags, exp(w v ) indicates w v The exponential function.
[0054] Assuming that each word in the expanded vocabulary contributes equally to the predicted label, a simple average is performed on them, and the mean of the predicted scores is used as the predicted score for that label. Finally, the category with the highest predicted score is selected as the final prediction result, as shown in equation (3).
[0055]
[0056] This invention can be further illustrated by the following experiments:
[0057] To test the effectiveness of this invention, prediction results were implemented on three benchmark datasets: SC-Clickbait, DL-Clickbait, and W23-Clickbait. SC-Clickbait contains 32,000 news headlines, including 16,000 clickbaits and 16,000 non-clickbaits, collected from multiple news websites such as BuzzFeed, Upworthy, and ViralNova. DL-Clickbait contains 2,388 pieces of content crawled from Reddit, Facebook, and Twitter, including 814 clickbait samples. W23-Clickbait comes from a clickbait challenge and includes two versions; we used the larger version. To ensure a realistic measurement of detection performance, this paper selected accuracy and F1 score as representative evaluation metrics, defined as follows:
[0058]
[0059]
[0060]
[0061]
[0062] Where tp is the number of click decoys correctly predicted by the algorithm, fp is the number of non-click decoys predicted by the algorithm, fn is the number of non-click decoys predicted by the algorithm but which are actually click decoys, tn is the number of non-click decoys correctly predicted by the algorithm, N is the total number of predicted articles, and F1 score is the harmonic mean of precision and recall.
[0063] To demonstrate the performance of the test results, we selected several baseline methods for comparison, including traditional neural network models such as CNN, LSTM, and BiLSTM, as well as some pre-trained language model-based methods such as BERT, RoBERTa, UNIFIEDM2, PT, and KPT. The prediction results are shown in Table 1. Table 1 shows that the prediction result of this invention on the click decoy dataset reaches 95%, which is superior to other methods.
[0064] Table 1 shows the experimental results on the three datasets.
[0065]
[0066] The decoy click detection method of this invention selects a suitable pre-trained language model to construct label words and templates for cue learning. It optimizes the label words in cue learning through five optimization strategies. Using the expanded label words, the classification task is transformed into a probability calculation problem for category label words. The input text and the preset cue template are combined to construct a masked cue text, which serves as the model input. Click decoy detection is performed using the optimized label words. Finally, the predicted probability of each label word is mapped to the corresponding category, and the final prediction score of that label is obtained as the classification result. This method can be applied to text classification.
[0067] This invention is not limited to the above embodiments. Based on the technical solutions disclosed in this invention, those skilled in the art can make some substitutions and modifications to some of the technical features without creative effort, and all such substitutions and modifications are within the protection scope of this invention.
Claims
1. A click decoy detection method that uses only titles for cue learning, characterized in that, Includes the following steps: Step 1: Select a suitable pre-trained model and use it as the backbone to construct the tag words and templates for cue learning; Step 2: Expand the category tag words using five optimization strategies to obtain the optimized tag word set. The specific steps are as follows: Step 2.1: In the expansion of tag words, the special category V associated with the specific tag. y V1={clickbait} can be expanded to V1={clickbait,hyperlink,misleading,…}; Step 2.2: In the expansion of the tag terms, select the tag class names of both click decoys and non-click decoys as the initial words for concept search in Probase. Select N concepts from the probabilistic retrieval ranking results of Probase, denoted as N(v); then, calculate the distance dist(V) between the tag class name y in the embedding space and the expanded tag term set. y For all inputs x, y), i {N1,…,N i ,…,N n The list is a similarity ranking result obtained through concept retrieval, and the top M words are selected as the expanded tag word set. Step 2.3: After obtaining the above tag word set, the tag words are further filtered using four optimization strategies, including using BERT to predict the probability distribution of expanded words filling the mask position, calculating the semantic similarity between the fastText vector of each class tag name and the fastText vector of the expanded tag word, based on word frequency provided by Wikipedia, and considering the contextual information of the mask position; for each strategy, N is selected. a =15 words, and finally the union of the words filtered by the four strategies is taken to form the final extended word set; Step 3: Using the expanded tags, transform the classification task into a probability calculation problem for category tags; Step 4: Map the predicted probability of each tag to the corresponding category to obtain the final prediction score of the tag as the classification result.
2. The click decoy detection method based solely on title-based cue learning according to claim 1, characterized in that: Step 1 specifically includes: Step 1.1: Select the RoBERTabase model as the pre-trained language model. This model uses a dynamic masking strategy. Each time a sequence is input into the model, a new masking pattern is generated. As a large amount of data is continuously input, the model will gradually adapt to different masking strategies and learn different language representations. Step 1.2: Use auxiliary tags [head] and [tail] to indicate the start and end of the template. Combine the template with the original input title to form the input of the model. The template is "[head][placeholder: x] .In summary, it was [mask].[tail]", where "[placeholder: x]" is the position to fill the news title of the input.
3. The click decoy detection method based solely on title-based cue learning according to claim 1, characterized in that: Step 3 specifically includes: Step 3.1: Given a set of input data x={x1,…,x...} n They will all be classified into category y∈Y, and the set of category labels is denoted as V. y ={v1,…,v n }, where V is the set of the entire category, V y It is a subset of V; in the pre-trained language model M, the probability that each word v is filled into [MASK] is represented as P([MASK]=v∈V). y |x p Then, clicking on the decoy detection task can be transformed into the probability calculation problem of category tag words, as shown in equation (1). （1）， Step 3.2: During the detection of clickbait, if the probability of belonging to y1="clickbait" is greater than the probability of y2="non-clickbait", then the text will be classified as "clickbait".
4. The click decoy detection method based solely on title-based cue learning according to claim 1, characterized in that: Step 4 specifically includes: Step 4.1: After calculating the probabilities of all category labels, map the predicted probabilities of each label to the corresponding category, assign a learnable weight parameter to each label, and then normalize it, as shown in equation (2). （2）， Where, α v w is the predicted probability of the class label v. v These are learnable weight parameters associated with label v, where u represents the set of all labels V under category y. y One of the tags, exp(w v ) indicates w v The exponential function; Assuming that each word in the expanded vocabulary contributes equally to the predicted label, a simple average is performed, and the mean of the predicted scores is used as the predicted score for that label. Finally, the category with the highest predicted score is selected as the final prediction result, as shown in equation (3). （3）。