A Chinese privacy policy automatic labeling and abstract extraction method and device

Abstract

The application discloses a kind of Chinese privacy policy automatic marking and abstract extraction method and device, including the division of privacy policy text into clause sentence, after the marking of each clause sentence is carried out category label, important clause label and potential risk clause label according to privacy policy marking criterion, carry out data enhancement and word segmentation processing to build training sample;Supervised learning training is carried out to category identification model, key sentence identification model and risk identification model respectively using training sample and corresponding label;After the privacy policy text to be marked is processed into test sample, at least one model in the three models trained is used to predict the label of test sample, and each clause sentence in privacy policy text is marked according to label prediction result;According to the label marking result, the clause sentence group meeting the demand is extracted from the privacy policy text to form abstract text.The method and device are high in marking quality, strong in universality and easy to use.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of computer natural language processing and privacy policy, and particularly relates to a Chinese privacy policy automatic labeling and abstract extraction method and device. BACKGROUND

[0002] Privacy policy labeling is a basic way to obtain large-scale privacy policy corpus for natural language processing tasks. However, manual labeling is costly and inefficient. In contrast, crowdsourcing labeling reduces the cost of privacy policy labeling, but it is difficult to guarantee the quality of privacy policy labeling. Therefore, privacy policy automatic labeling based on text classification has gradually attracted people's attention. However, compared with foreign research progress, there are fewer solutions to Chinese privacy policy automatic labeling and the performance is insufficient.

[0003] Privacy policy abstract is a natural language processing technology for obtaining short or valuable key information from lengthy privacy policies. It has extremely important practical significance for reducing privacy policy reading time and identifying privacy clause risks. The main fields of privacy policy abstract involve two different subfields. One is to extract important clauses in privacy policies using classification technology. The other is to use text generation technology to translate lengthy and difficult-to-understand privacy policies in privacy policies into short and easy-to-understand privacy clauses.

[0004] Initially, "The Creation and Analysis of a Website Privacy Policy Corpus" provides an automatic labeling model for English. It uses logistic regression, SVM, and HMM to label privacy policies, respectively. However, the complex relationship between data instances can lead to problems such as ambiguity, contradiction, and unclear scope in the constructed database, affecting the consistency of manual labeling and the accuracy of the labeling model. "The Creation and Analysis of a Website Privacy Policy Corpus" studies the accuracy and scalability of crowdsourcing labeling. It designs a labeling scheme based on common content in privacy lawsuits. Studies show that whether the details of the labeling scheme are easy to understand and whether the boundaries are clear have a great impact on the results of crowdsourcing labeling. It also emphasizes that we should not have too high expectations for crowdsourcing results.

[0005] The patent "An automatic classification method based on Chinese privacy policy clauses" introduces a privacy policy classification method based on SVM. This method first extracts the features of privacy clauses, and then uses the SVM classifier to obtain the privacy clauses of different categories, and classifies and displays them. This method has only achieved satisfactory results in some categories. "A Tale of Two Regulatory Regimes: Creation and Analysis of a Bilingual Privacy Policy Corpus" applies Bert and m-Bert models to the automatic labeling task of English and German privacy policies, and the results are significantly better than previous labeling models. "A Tale of Two Regulatory Regimes: Creation and Analysis of a Bilingual Privacy Policy Corpus" proposes a rule-based privacy policy abstract technology for extracting different types of content in privacy policies and extracting them to the corresponding category. This method is based on an English corpus, and the rule base is difficult to build and cover all samples, which makes the model prone to overfitting. The patent "A privacy policy abstract method based on natural language processing" provides a generative text abstract method, but this model still uses an English corpus for training and verification. "Toward Domain-Guided Controllable Summarization of Privacy Policies" proposes a method for summarizing only the risk clauses of English privacy policies. It considers the summary as a binary classification problem, trains a CNN model to predict the risk probability of the privacy policy, and proposes two risk-centered and coverage-centered summary methods that are greatly affected by the threshold. SUMMARY

[0006] In view of the above, in view of the problems of large-scale privacy policy corpus labeling required for training high-quality machine learning models, high cost, and the problems of long privacy policies and ordinary users' difficulty in obtaining privacy clauses with security risks. The present application provides a Chinese privacy policy automatic labeling and abstract extraction method and device to realize fast, low-cost high-quality privacy policy labeling and privacy policy abstract extraction.

[0007] To achieve the above invention purpose, the embodiment provides a Chinese privacy policy automatic labeling and abstract extraction method, which comprises the following steps:

[0008] Text processing: divide the privacy policy text into clause sentences, label each clause sentence with category label, important clause label, and potential risk clause label according to the privacy policy labeling criteria, then perform data augmentation and word segmentation processing to construct training samples;

[0009] Model training: use the training samples and corresponding category labels, important clause labels, and potential risk clause labels to perform supervised learning training on the category recognition model, key sentence recognition model, and risk recognition model, respectively;

[0010] Labeling: after processing the privacy policy text to be labeled into test samples, use at least one of the trained category recognition model, key sentence recognition model, and risk recognition model to predict the labels of the test samples, and label each clause sentence in the privacy policy text according to the label prediction results;

[0011] Abstract extraction: extract the required clause sentences from the privacy policy text according to the labeling results to form an abstract text.

[0012] In an optional embodiment, during text processing, the privacy policy text is divided into clause sentence granularity using regular expressions, and the sentence break symbols include single character sentence break symbol, English ellipsis sentence break symbol, Chinese ellipsis sentence break symbol, and double quotation marks as the end point sentence break symbol when there is a sentence break symbol before the double quotation marks.

[0013] In an optional embodiment, during text processing, data augmentation is performed using TextBlob through back-translation technology, and word segmentation processing is performed using Tokenizer and Bert's Vocab word table.

[0014] In an optional embodiment, the network structure of the category recognition model, key sentence recognition model, and risk recognition model includes an input layer, an embedding layer, an encoding layer, a convolution layer, a pooling layer, a fully connected layer, and an output layer. Except for the output layer, the other layers of the three models are the same, wherein the input layer is used to obtain the Token sequence T n of the sample; the embedding layer is used to calculate the embedding matrix E n of T n ; the encoding layer is used to use Bert encoder and LSTM encoder to encode E nThe encoding is performed to obtain vectors B and L; a convolution layer is used to perform convolution operation on the spliced result BL of vectors B and L to obtain matrix C; a pooling layer is used to perform maximum pooling processing on C to obtain vector P; a full connection layer is used to perform full connection mapping on vector P, and an output layer is used to perform activation mapping according to the full connection mapping result to output a prediction value, wherein the output layer of the category recognition model has N dimensions, corresponding to N category labels, and the output layers of the key sentence recognition model and the risk recognition model each have one dimension, corresponding to the prediction values of the important clause label and the potential risk clause label.

[0015] In an optional embodiment, the full connection layer adopts a ReLU activation function for activation mapping, and the output layer adopts a Sigmoid activation function for activation mapping to output a prediction result.

[0016] In an optional embodiment, when the category recognition model, the key sentence recognition model and the risk recognition model are supervised learning trained, the cross-entropy between the prediction result of the model and the training sample is used as a loss function.

[0017] In an optional embodiment, when the abstract is extracted, the clause sentences meeting the needs are extracted from the privacy policy text according to the label annotation results to form an abstract text, including:

[0018] First, the clause sentences annotated with the important clause label are extracted from the privacy policy text to form a candidate clause sentence set; then, the clause sentences annotated with the category label of interest are selected from the candidate clause sentence set to form an abstract text.

[0019] In an optional embodiment, when the abstract is extracted, when the risk clause needs to be extracted, the clause sentences annotated with the potential risk clause label are extracted from the privacy policy text as part of the abstract text.

[0020] To achieve the above-mentioned purposes, the embodiments further provide a Chinese privacy policy automatic labeling and abstract extraction device, which comprises a text processing module, a model training module, a label annotation module and an abstract extraction module.

[0021] The text processing module is used to divide the privacy policy text into clause sentences, and after each clause sentence is annotated with a category label, an important clause label and a potential risk clause label according to a privacy policy annotation criterion, data enhancement and word segmentation processing are performed to construct training samples;

[0022] The model training module is used to perform supervised learning training on the category recognition model, the key sentence recognition model and the risk recognition model respectively by using the training samples and the corresponding category labels, important clause labels and potential risk clause labels;

[0023] The labeling module is used to process the privacy policy text to be labeled into test sample data, and then use at least one of the trained category recognition model, key sentence recognition model and risk recognition model to predict the labels of the test samples, and label each clause sentence in the privacy policy text according to the label prediction results.

[0024] The summary extraction module is used to extract relevant clause sentences from the privacy policy text based on the tagging results to form a summary text.

[0025] Compared with the prior art, the beneficial effects of the present invention include at least the following:

[0026] By leveraging label prediction results from category recognition, key sentence recognition, and risk recognition models, clause sentences can be quickly labeled to build a large-scale, high-quality privacy policy corpus. This corpus then yields privacy policy summaries tailored to different needs and can detect risky content within privacy policies. This approach effectively reduces manual costs associated with related content creation, improves task efficiency, and offers advantages such as high labeling quality, strong versatility, and ease of use. Attached Figure Description

[0027] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0028] Figure 1 This is a flowchart of the method for automatically annotating and extracting Chinese privacy policies provided in the embodiment;

[0029] Figure 2 This is a schematic diagram of the model structure provided in the embodiment;

[0030] Figure 3 This is a flowchart of the tag annotation and summary extraction process provided in the embodiment;

[0031] Figure 4 This is a schematic diagram of the summary extraction results provided in the embodiment;

[0032] Figure 5 This is a schematic diagram of the structure of the Chinese privacy policy automatic annotation and summary extraction device provided in the embodiment. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0034] like Figure 1 As shown in the example, the method for automatically annotating and extracting Chinese privacy policies includes the following steps:

[0035] Step 1, Text Processing

[0036] In this embodiment, the text processing process includes: obtaining the privacy policy text divided into clause sentences, labeling each clause sentence with category tags, important clause tags, and potential risk clause tags according to the privacy policy labeling guidelines, and then performing data augmentation and word segmentation to construct training samples.

[0037] A total of 125 privacy policy texts were collected from mainstream or best-selling apps, of which 85 were from application software and 40 were from game software.

[0038] The Privacy Policy Labelling Guidelines are established based on legal and regulatory requirements and the characteristics of privacy terms. The labeling standards in these guidelines fully encompass the complete content required by laws and regulations to be mapped to the policy, including the following 11 category labels (1)-11), the important clause label (12), and the potential risk clause label (13).

[0039] 1) "Data collected by the application subject": refers to the privacy terms for user identity information, device information, etc. collected in specific application scenarios or functions, as well as the privacy terms that do not authorize the impact of specific collected content on application functions;

[0040] 2) "Permissions obtained by the application subject": This refers to the privacy terms regarding application device permissions obtained for specific application scenarios or functions, as well as the privacy terms regarding the impact of not granting specific permissions on application functions.

[0041] 3) "Third-party collection, sharing with third parties, entrustment, public disclosure, and transfer": refers to privacy clauses describing whether third parties collect data through SDKs or other means, share or transfer data with third parties, entrust third parties to manage data, or disclose data to the public.

[0042] 4) "Use": refers to the use of data for building data analysis models, personalized content recommendation or service models, automated decision-making models, etc., in accordance with privacy terms.

[0043] 5) "Storage": refers to the privacy terms regarding the storage method, location, duration, time, and medium of personal information;

[0044] 6) "Security Measures": refers to privacy clauses related to the security measures, principles, technologies, etc., collected to protect the security of users' personal data;

[0045] 7) "Data and Access Management": This refers to privacy terms that clearly define how specific personal information can be modified and how to modify it, including disabling personalized ads, content recommendations, adding, deleting, modifying, and querying personal information, revoking permissions, and canceling accounts;

[0046] 8) "Specific groups": refers to special provisions for groups with specific identities or groups from specific regions;

[0047] 9) "Contact Information": Includes the email address and phone number for contacting customer service to modify or file complaints regarding the privacy policy;

[0048] 10) "Privacy Policy Authorization, Modification and Revision": Clearly define the data behavior that may result from authorizing the privacy policy and the consequences of not authorizing it, as well as the content of the privacy policy relative to previous revisions and changes;

[0049] 11) "Cessation of Operations": Relevant clauses regarding how collected data will be handled when the application ceases operations;

[0050] 12) "Important Clauses": These refer to clauses among the above 11 types that explicitly involve collection, sharing, and management activities, clauses that require a summary explanation, and other clauses that require a summary but are not included in the above tags.

[0051] 13) "Potential Risk Clauses": The privacy terms described in the privacy policy may contain clauses that harm users' personal information or are inconsistent with legal provisions.

[0052] In this embodiment, regular expressions are used to divide the privacy policy text into clause sentences. The segmentation markers include single-character segmentation markers (([。!?;;\?])([^”'])), English ellipsis segmentation markers (\.{6})([^”']), Chinese ellipsis segmentation markers (\…{2})([^”']), and segmentation markers ending with double quotes when there is a segmentation marker before double quotes ([。!?\?][”'])([^,。!?\?]).

[0053] In this embodiment, after obtaining the clause sentences, law graduate students with a background in privacy policy-related laws label the clause sentences corresponding to the privacy policy text according to the privacy policy labeling guidelines. Then, TextBlob is used to perform data augmentation through reverse translation technology. Tokenizer and Bert's Vocabulary are used to segment the tagged clause sentences to construct training samples. Each training sample is represented as a token sequence.

[0054] Step 2, Model Training

[0055] In this embodiment, the category recognition model, key sentence recognition model, and risk recognition model are trained in a supervised manner using training samples and corresponding category labels, important clause labels, and potential risk clause labels, respectively.

[0056] The category recognition model is used to categorize the clause sentences of the privacy policy text. It is a multi-label classifier, such as... Figure 2 As shown, its structure includes an input layer, an embedding layer, an encoder layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer.

[0057] The input layer is used to obtain the token sequence T of the clause sentences corresponding to the privacy policy text. n .

[0058] Embedded layers are used to compute T n Embedding matrix E n Preferably, T can be obtained using a BERT pre-trained model. n Embedding matrix E n =[e1,e2,…,e m ], E n Any row in the dictionary corresponds to a word vector e in the vocabulary. m =[e m1 ,e m2 ,…,e ml ] T That is, any privacy policy terms of E n It can be represented as follows:

[0059]

[0060] Where l is the length of the embedding vector, specifically l = 768.

[0061] The encoding layer is used to encode E using a Bert encoder and an LSTM encoder respectively. n Encode to obtain vector B = h b And vector L = h l .

[0062] The convolutional layer uses six identical 2×1 convolutional kernels with a stride of 1. The concatenation result of matrices B and L is BL = [h b ,h l Perform a convolution operation to obtain matrix C = l × 6.

[0063] The pooling layer is used to perform max pooling on matrix C to obtain a vector P of dimension l.

[0064] Fully connected layers are used to map vector P to fully connected layers. A four-layer fully connected layer structure is adopted, with lengths of 1024, 512, 128, and 64 respectively. ReLU is used as the activation function of the fully connected layer, and Dropout is used to prevent the model from overfitting.

[0065] The output layer performs activation mapping based on the fully connected mapping results to output predicted values. This output layer consists of an 11-dimensional vector, with each dimension's output value corresponding to a class label. The output layer uses a sigmoid activation function, and the output value can be considered as the probability that a sample belongs to the corresponding label.

[0066] When training a category recognition model, training samples are used as input, and the corresponding category labels of the training samples are used as supervision. Cross-entropy is used as the loss function to measure the error. Training is represented as:

[0067]

[0068] Where C represents the number of category labels, y (i) Represents the category labels as vectors; This is the predicted probability value output by the category recognition model, where i represents the i-th sample and m represents the number of samples.

[0069] The key sentence recognition model is used to annotate important clauses in privacy policy texts. Specifically, it's a binary classifier used to identify whether a clause in a privacy policy is a crucial privacy clause requiring a summary. Its structure is basically the same as the category recognition model, including an input layer, embedding layer, encoding layer, convolutional layer, pooling layer, fully connected layer, and output layer. The difference is that the output layer of the key sentence recognition model has a dimension of 1, corresponding to the important clause label, and represents the predicted probability value of the important clause label.

[0070] When training the key sentence recognition model, training samples are used as input, and the corresponding important clause labels are used as supervision. Cross-entropy is used as the loss function to measure the error. Training is represented as: Where y represents an important terms label, This represents the predicted probability value output by the key sentence recognition model.

[0071] The risk identification model is used to label potentially risky clauses in the terms and conditions of a privacy policy text. Essentially, it's a binary classifier used to identify whether any clauses in a privacy policy contain potentially risky terms. Its structure is basically the same as the category identification model, including an input layer, embedding layer, encoding layer, convolutional layer, pooling layer, fully connected layer, and output layer. The difference is that the output layer of the risk identification model has a dimension of 1, corresponding to the potential risk clause label, and represents the predicted probability value of the potential risk clause label.

[0072] When training the key sentence recognition model, the training samples are used as input, and the potential risk clause labels corresponding to the training samples are used as supervision. Cross-entropy is used as the loss function to measure the error.

[0073] The trained category recognition model, key sentence recognition model, and risk recognition model can be used to predict categories, important clauses, and potential risk clauses, respectively, so as to label the corresponding clause sentences according to the predicted probability values.

[0074] Step 3, Labeling

[0075] In this embodiment, during the labeling process, the privacy policy text to be labeled is processed into a test sample. At least one of the trained category recognition model, key sentence recognition model, and risk recognition model is used to predict the label of the test sample. Based on the label prediction results, each clause sentence in the privacy policy text is labeled to obtain an automatic labeling corpus of privacy policies.

[0076] like Figure 3 As shown, the privacy policy text to be annotated is segmented and word-segmented using the method in step 1 to obtain test samples represented by token sequences. After the test samples are filtered, the text to be annotated is selected, and the token sequences corresponding to the text to be annotated are input into the category recognition model, key sentence recognition model, and risk recognition model respectively for category recognition, important clause recognition, and potential risk clause recognition to obtain the tag prediction results. Then, the clause sentences are annotated according to the tag prediction results to obtain the annotated text.

[0077] It is also possible to integrate category recognition models, key sentence recognition models, and risk recognition models together, and set a function switch for each model. When the model is applied, the function switch is turned on, and otherwise the function switch is turned off, so as to achieve prediction of different functions.

[0078] Step 4, Abstract Extraction

[0079] In this embodiment, during summary extraction, clause sentences that meet the requirements are extracted from the privacy policy text based on the tagging results to form the summary text. Specifically, this includes: first, extracting clause sentences tagged with important clauses from the privacy policy text to form a candidate clause sentence set; then, selecting clause sentences tagged with categories of interest from the candidate clause sentence set to form the summary text. When it is necessary to extract risk clauses, clause sentences tagged with potential risk clauses are extracted from the privacy policy text as part of the summary text, and can be highlighted in bold red.

[0080] An implementation example of this invention is carried out on a machine configured with Ubuntu 22.04, equipped with an NVIDIA A100 Tensor Core GPU processor and 128GB of memory. Using all the parameter values ​​listed in the specific implementation, the privacy policy text is automatically annotated. The automatic annotation is stored in .json format, consisting of three parts: a privacy policy-specific ID, the text itself, and a list of labels. The summary text extracted based on the annotation results is as follows: Figure 4 As shown, the privacy policy summary is stored as a .docx file, which displays the important key information in the privacy policy by section.

[0081] To achieve the above-mentioned objectives, such as Figure 5 As shown in the embodiment, an automatic annotation and summary extraction device for Chinese privacy policies is also provided, including a text processing module, a model training module, a tag annotation module, and a summary extraction module. The text processing module divides the privacy policy text into clause sentences, and annotates each clause sentence with category tags, important clause tags, and potential risk clause tags according to privacy policy annotation guidelines. Then, it performs data augmentation and word segmentation to construct training samples. The model training module uses the training samples and corresponding category tags, important clause tags, and potential risk clause tags to perform supervised learning training on a category recognition model, a key sentence recognition model, and a risk recognition model, respectively. The tag annotation module processes the privacy policy text to be annotated into test sample data, uses at least one of the trained category recognition model, key sentence recognition model, and risk recognition model to predict the tags for the test samples, and annotates each clause sentence in the privacy policy text with tags based on the tag prediction results. The summary extraction module extracts clause sentences that meet the requirements from the privacy policy text based on the tag annotation results to form a summary text.

[0082] It should be noted that the automatic annotation and summary extraction device for Chinese privacy policies provided in the above embodiments should be illustrated using the above-described functional module divisions. The functions can be assigned to different functional modules as needed, i.e., the internal structure of the terminal or server can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the automatic annotation and summary extraction device for Chinese privacy policies provided in the above embodiments and the method embodiments for automatic annotation and summary extraction of Chinese privacy policies belong to the same concept. For details of its implementation, please refer to the method embodiments for automatic annotation and summary extraction of Chinese privacy policies, which will not be repeated here.

[0083] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for automatically annotating and extracting summaries of Chinese privacy policies, characterized in that, Includes the following steps: Text processing: The privacy policy text is divided into clause sentences. Each clause sentence is labeled with category tags, important clause tags, and potential risk clause tags according to the privacy policy labeling guidelines. Then, data augmentation and word segmentation are performed to build training samples. Model training: Supervised learning training is performed on the category recognition model, key sentence recognition model, and risk recognition model using training samples and corresponding category labels, important clause labels, and potential risk clause labels, respectively. Tagging: After processing the privacy policy text to be annotated into test samples, at least one of the trained category recognition model, key sentence recognition model and risk recognition model is used to predict the labels of the test samples, and each clause sentence in the privacy policy text is labeled according to the label prediction results. Summary extraction: Based on the tagging results, extract the relevant clause sentences from the privacy policy text to form a summary text. Specifically, when it is necessary to extract risk clauses, extract the clause sentences tagged with potential risk clauses from the privacy policy text as part of the summary text. The network structures of the category recognition model, key sentence recognition model, and risk recognition model all include an input layer, an embedding layer, an encoding layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Except for the output layer, the other layers of the three models are the same. The input layer is used to obtain the token sequence of the sample. T n The embedding layer is used for computation. T n Embedding matrix E n The encoding layer is used to encode data using a Bert encoder and an LSTM encoder, respectively. E n The process involves encoding vectors B and L; a convolutional layer is used to convolve the concatenated result BL of vectors B and L to obtain matrix C; a pooling layer is used to perform max pooling on matrix C to obtain vector P; a fully connected layer is used to map vector P to the fully connected layer; and the output layer is used to perform activation mapping based on the fully connected mapping result to output the predicted value. The output layer of the category recognition model has N dimensions, corresponding to N category labels, while the output layers of the key sentence recognition model and the risk recognition model have one dimension, corresponding to the predicted values ​​of the important clause label and the potential risk clause label, respectively.

2. The method for automatically annotating and extracting summaries of Chinese privacy policies according to claim 1, characterized in that, In text processing, regular expressions are used to divide the privacy policy text into clause sentences. The sentence breakers used in the division include single-character sentence breakers, English ellipsis sentence breakers, Chinese ellipsis sentence breakers, and sentence breakers ending with double quotes when there is a sentence breaker before double quotes.

3. The method for automatically annotating and extracting summaries of Chinese privacy policies according to claim 1, characterized in that, In text processing, TextBlob is used for data augmentation through reverse translation technology, and Tokenizer and Bert's Vocabulary are used for word segmentation.

4. The method for automatically annotating and extracting summaries of Chinese privacy policies according to claim 1, characterized in that, The fully connected layer uses the ReLU activation function for activation mapping, and the output layer uses the Sigmoid activation function for activation mapping to output the prediction result.

5. The method for automatically annotating and extracting summaries of Chinese privacy policies according to claim 1, characterized in that, When conducting supervised learning training on the category recognition model, key sentence recognition model, and risk recognition model, the cross-entropy between the model's prediction results and the training samples is used as the loss function.

6. The method for automatically annotating and extracting summaries of Chinese privacy policies according to claim 1, characterized in that, In the abstract extraction process, the abstract text is composed of relevant clauses extracted from the privacy policy text based on the tagging results. This includes: First, extract clause sentences marked with important clause tags from the privacy policy text to form a candidate clause sentence set; then, select clause sentences marked with category tags of interest from the candidate clause sentence set to form a summary text.

7. A device for automatically annotating and extracting Chinese privacy policies, characterized in that, It includes a text processing module, a model training module, a labeling module, and a summary extraction module. The text processing module is used to divide the privacy policy text into clause sentences, and after annotating each clause sentence with category tags, important clause tags, and potential risk clause tags according to the privacy policy annotation guidelines, it performs data augmentation and word segmentation to construct training samples. The model training module is used to perform supervised learning training on the category recognition model, key sentence recognition model and risk recognition model using training samples and corresponding category labels, important clause labels and potential risk clause labels, respectively. The labeling module is used to process the privacy policy text to be labeled into test sample data, and then use at least one of the trained category recognition model, key sentence recognition model and risk recognition model to predict the labels of the test samples, and label each clause sentence in the privacy policy text according to the label prediction results. The summary extraction module is used to extract the required clause sentences from the privacy policy text based on the tagging results to form a summary text. Specifically, when it is necessary to extract risk clauses, the clause sentences tagged with potential risk clauses are extracted from the privacy policy text as part of the summary text. The network structures of the category recognition model, key sentence recognition model, and risk recognition model all include an input layer, an embedding layer, an encoding layer, a convolutional layer, a pooling layer, a fully connected layer, and an output layer. Except for the output layer, the other layers of the three models are the same. The input layer is used to obtain the token sequence of the sample. T n The embedding layer is used for computation. T n Embedding matrix E n The encoding layer is used to encode data using a Bert encoder and an LSTM encoder, respectively. E n The process involves encoding vectors B and L; a convolutional layer is used to convolve the concatenated result BL of vectors B and L to obtain matrix C; a pooling layer is used to perform max pooling on matrix C to obtain vector P; a fully connected layer is used to map vector P to the fully connected layer; and the output layer is used to perform activation mapping based on the fully connected mapping result to output the predicted value. The output layer of the category recognition model has N dimensions, corresponding to N category labels, while the output layers of the key sentence recognition model and the risk recognition model have one dimension, corresponding to the predicted values ​​of the important clause label and the potential risk clause label, respectively.

Images