Intelligent management method for data security classification of language science and technology resources

Abstract

The invention provides a data security level intelligent management method for language science and technology resources, which comprises the following steps of: collecting related data of the language science and technology resources, and preprocessing the collected data to obtain an experimental data set; performing pre-training on the experimental data set by utilizing the BERT word vector to generate a language science and technology word list; segmenting the target document into a sentence set, and based on the language science and technology word list, performing training and fine tuning by utilizing a BERT model to generate a weighted sentence vector of the sentence set; and in combination with the BERT model and the TextRank algorithm, obtaining a secret level setting algorithm, and judging the security level of the target document by using the secret level setting algorithm. According to the method, the key information in the document can be accurately extracted, the document security level is automatically judged in combination with the language science and technology word list, intelligence and automation of language science and technology resource management are promoted, the efficiency and quality of security work are improved, and the security of the language science and technology resources is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of data confidentiality, and in particular to a method for intelligent management of data confidentiality levels of language science and technology resources. Background Art

[0002] With the rapid development of information technology and the advent of the big data era, language technology is increasingly being used in various fields. Language technology resources, including text corpora, speech data, linguistic research data, and computational linguistics materials, provide important support for the advancement of technologies such as natural language processing, machine translation, speech recognition, and intelligent customer service. However, the management and protection of language technology resources, especially the confidentiality management of confidential data, has become an important issue that needs to be addressed. In the context of increasingly stringent information security and confidentiality requirements, how to efficiently and accurately manage the confidentiality of language technology resources has become a research focus.

[0003] Traditional confidentiality determination methods rely on manual operations, usually with dedicated personnel responsible for receiving documents, formulating confidentiality determination rules, and conducting confidentiality determination and classification management. This method has problems such as low efficiency, strong subjectivity, and difficulty in ensuring quality. At present, the research on the intelligent and digitalization of confidentiality determination is in its infancy. In terms of confidentiality determination methods, traditional methods are mainly based on keyword matching, that is, comparing documents with confidential keywords in the confidentiality determination rules. The more confidential keywords and word pairs appear in the document, the greater the possibility that the document will be classified as a corresponding confidentiality level. However, this method ignores the problem of expressions that may appear in the document that are similar to keywords. Although the document does not contain confidential keywords, the meaning of the sentence may still be confidential, which leads to problems such as inaccurate confidentiality determination and low efficiency. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide an intelligent management method for the data confidentiality of language technology resources, which can accurately extract key information in documents and realize automatic determination of document confidentiality in combination with language technology vocabulary, thereby promoting the intelligent and automated management of language technology resources, improving the efficiency and quality of confidentiality work, and ensuring the security of language technology resources.

[0005] To achieve the above object, the present invention provides the following solution: a method for intelligent management of data confidentiality of language technology resources, comprising the following steps:

[0006] Collect data related to language technology resources, and perform data screening, Chinese word segmentation, and stop word removal on the collected data to obtain the experimental data set;

[0007] Use the BERT word vector method to perform pre-training on the experimental data set to generate a language technology vocabulary;

[0008] Divide the target document into a sentence set, and train and fine-tune the BERT model based on the language technology vocabulary to generate weighted sentence vectors for the sentence set;

[0009] The BERT model and the TextRank algorithm are combined to obtain a confidentiality determination algorithm, and the confidentiality determination algorithm is used to determine the confidentiality level of the target document through the similarity between the weighted sentence vector and the preset confidentiality determination rule.

[0010] Optionally, the BERT word vector method is used for pre-training on the experimental dataset to generate a language technology vocabulary, including:

[0011] Based on the experimental data set, a masked language model is used to predict masked words, and a next sentence prediction task is used to drive the masked language model to determine whether two sentences are adjacent in the original text, so as to perform pre-training of the BERT model;

[0012] Select a labeled dataset and fine-tune the parameters of the pre-trained BERT model to obtain the BERT model;

[0013] During the pre-training process of the BERT model, the WordPiece segmentation technology is used to segment the text and generate a vocabulary including common words and subword units;

[0014] Based on the vocabulary, each word in the experimental data set is encoded to generate a related word vector for capturing the semantic features of the word;

[0015] The vocabulary list and the related word vectors are combined to obtain a language technology vocabulary list.

[0016] Optionally, the target document is divided into a sentence set, and based on the language technology vocabulary, a BERT model is used for training and fine-tuning to generate a weighted sentence vector for the sentence set, including:

[0017] The target document is divided into a sentence set, and each word in the sentence set is converted into a word vector using the BERT model;

[0018] Defining weight factors according to the part of speech, frequency or semantic importance of the language technology vocabulary, and defining a weighting formula in combination with the weight factors;

[0019] Based on the language technology vocabulary in the target document and the word vector, a weighted sentence vector is generated using the weighting formula.

[0020] Optionally, the weighted formula is expressed as:

[0021]

[0022] Among them, S is the weighted sentence vector, N is the number of words in the sentence, and a i is the weight factor of the i-th word, v i is the vector representation of the th word generated by BERT, and f is a function defined according to the task requirements.

[0023] Optionally, combining the BERT model and the TextRank algorithm to obtain a confidentiality determination algorithm, using the confidentiality determination algorithm to determine the confidentiality level of the confidential language technology resources in the target document through the similarity between the weighted sentence vector and a preset confidentiality determination rule, including:

[0024] Segment the target document, remove stop words and punctuation marks, and encode the processed text using the BERT model to obtain context-related word vectors for each word;

[0025] A language technology resource graph model is constructed using the context-related word vectors; in the language technology resource graph model, each node is a sentence, the edge is the relationship between the nodes, and the weight of the edge is determined by the similarity of the context-related word vectors;

[0026] Using the TextRank algorithm to calculate the ranking score of each node in the language technology resource graph model to obtain the importance of the node;

[0027] The encryption rules are preset, sensitivity scores are given to each node, and the sensitivity scores are integrated into the ranking calculation of TextRank to obtain the ranking scores of weighted sentences;

[0028] The information in the target document is extracted according to the ranking score of the weighted sentences, and the confidentiality score of the target document is calculated according to the confidentiality rules to obtain the document confidentiality level, and the high-ranking words and sentences in the target document are marked as sensitive information that requires access authorization protection.

[0029] Optionally, the edge weight is calculated as:

[0030]

[0031] Among them, v i and v j are the vectors of nodes i and j respectively, and sim(.,.) represents the cosine similarity of the vectors.

[0032] Optionally, the node ranking score is calculated as:

[0033]

[0034] Among them, r i is the ranking score of node i, d is the damping coefficient, In(i) and Out(j) represent the in-edge and out-edge of node i respectively, and wji is the weight of edge (j,i).

[0035] Optionally, the ranking score of the weighted sentence is calculated as:

[0036] S i =max(S i,base ,sim(v i ,v sensitive ))

[0037] r i ′=r i ×S i

[0038] Among them, S i S is the sensitivity score of word i. i,base As the basic sensitivity score, sim(v i ,v sensitive ) is the similarity between node i and the secret rule, r i ′ is the weighted ranking score of the sentence, r i is the ranking score of the unweighted sentence.

[0039] The present invention discloses the following technical effects by providing a method for intelligent management of data confidentiality of language technology resources:

[0040] 1. By introducing BERT word vectors, the present invention can generate high-dimensional semantic vectors for each word, capture the unique meaning of words in different contexts, and the similarity calculation between these vectors provides more reasonable weights for the edges of the graph. This construction method not only considers the explicit co-occurrence of words, but also deeply explores the implicit semantic connections between them, so that the graph structure can better reflect the true semantic layout of the text. Compared with the traditional edge construction method that relies on co-occurrence information, the semantic similarity introduced by BERT makes the connection between words more in line with the actual context, especially when processing long texts or complex topics, it can significantly improve the relevance of keyword extraction and the accuracy of summary generation.

[0041] 2. The present invention combines BERT word vectors with TextRank to enable the TextRank algorithm to more flexibly deal with texts in different fields. The BERT pre-training model covers a wide range of language phenomena, and its word vectors can adapt to different contexts, thereby maintaining high expressiveness in various types of text processing. TextRank can generate keywords and summaries close to the text content in different application scenarios, reflecting stronger generalization ability and applicability.

[0042] 3. The present invention also effectively avoids the errors that may occur in traditional methods when facing synonyms, near-synonyms or polysemous words through the improvement of the TextRank algorithm, so that the generated text analysis results are more informative and practical. The TextRank algorithm that introduces BERT word vectors not only improves the accuracy of graph model construction, but also optimizes the overall performance of the algorithm. The keywords and abstracts generated by this deep fusion are not only more accurate, but also more logical and readable in content and expression, making the text processing results closer to human understanding of semantics.

[0043] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0045] Figure 1 A schematic diagram of a method flow chart provided by an embodiment of the present invention;

[0046] Figure 2 A schematic diagram of the logic flow of the encryption algorithm provided in the embodiment of the present invention;

[0047] Figure 3 A schematic diagram of experimental results of the accuracy of document confidentiality calculation provided by an embodiment of the present invention;

[0048] Figure 4 A schematic diagram of the comparison results between the embodiment of the present invention and the traditional TextRank experiment. DETAILED DESCRIPTION

[0049] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0050] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] like Figure 1-2 As shown, the present invention provides a method for intelligent management of data confidentiality of language technology resources, comprising the following steps:

[0052] 1. Collect data related to language technology resources, screen the collected data, perform Chinese word segmentation, and remove stop words, convert the original unstructured text data into a format that is easy for computers to process and analyze, so as to obtain the experimental data set.

[0053] Data closely related to the field of language technology are collected through a variety of means, including crawling data from the Internet, experimental data entered by experimenters themselves, and importing data from a dedicated language technology experimental platform. In the process of data collection, attention is paid to screening data sources that are highly relevant to language technology research to ensure that the collected data is highly relevant and practical.

[0054] Data screening: In the early stages of data processing of language technology resources, the main task is to screen the data. The pictures, videos, hyperlinks and various non-text symbols contained in the original data are both redundant and disruptive to text analysis. By excluding these irrelevant contents, not only can the data be purified, but also the efficiency and accuracy of subsequent processing can be improved, laying the foundation for in-depth analysis.

[0055] Chinese word segmentation: As the core link of data preprocessing, Chinese word segmentation has a direct impact on the subsequent tag generation and text analysis effects. Jieba word segmentation technology, combined with rule-based and statistical-based methods, effectively solves the problem of Chinese word segmentation. Jieba word segmentation improves scanning efficiency by constructing a triangular tree structure scanning word graph, and uses directed acyclic graph (DAG) to divide all possible word formation situations in detail. Finally, it determines the maximum probability path through dynamic programming to achieve accurate word segmentation. For new words outside the dictionary, Jieba word segmentation also uses the hidden Markov model (HMM) and Viterbi algorithm based on Chinese characters to further improve the accuracy and flexibility of word segmentation. In addition, Jieba word segmentation supports user-defined dictionaries, which greatly enhances the adaptability and accuracy of the word segmentation system in the field of language technology. By introducing field-specific proper nouns, it effectively avoids the situation where language technology professional terms are incorrectly segmented, thereby ensuring the accuracy and completeness of the word segmentation results.

[0056] Stop word removal: Stop word removal is an important step in text preprocessing. Frequently appearing conjunctions, prepositions, pronouns, etc. have limited effects in capturing the core meaning of the text and conducting in-depth analysis. By removing these stop words, data noise can be further reduced and key information in the text can be highlighted.

[0057] 2. Use the BERT word vector method to pre-train on the experimental data set to generate a language technology vocabulary. Including:

[0058] Based on the experimental dataset, a masked language model (MLM) is used to predict masked words, and a next sentence prediction task (NSP) is used to drive the masked language model to determine whether two sentences are adjacent in the original text, so as to pre-train the BERT model; the combination of the masked language model and the next sentence prediction task enables BERT to generate a language technology vocabulary with deep semantic information.

[0059] A specific task (labeled dataset) is selected, and parameters of the pre-trained BERT model are fine-tuned on the specific task to make it better adapt to the needs of specialized vocabularies such as language technology, and finally the BERT model is obtained.

[0060] During the pre-training process of BERT, the language technology vocabulary is finally obtained through word segmentation and encoding. BERT uses WordPiece word segmentation technology, which splits the vocabulary into smaller sub-word units so that it can handle unseen vocabulary and new words.

[0061] Vocabulary construction: During training, BERT uses the WordPiece algorithm to segment the text and generate a vocabulary containing common words and subword units. Each unit in the vocabulary has a unique ID, and these IDs correspond to the embedding vectors in the pre-trained model. This enables the model to map each word or subword unit in the text into a high-dimensional vector space.

[0062] Word vector generation: BERT generates word vectors for each word in the experimental dataset. These word vectors can capture the semantic features of the words and generate different vector representations when processing different contexts.

[0063] The vocabulary list and the related word vectors are combined to obtain a language technology vocabulary list. Language technology keywords include but are not limited to: Natural Language Processing (NLP); Machine Translation; Speech Recognition; Text-to-Speech (TTS); Natural Language Generation (NLG); Sentiment Analysis; Semantic Understanding; Named Entity Recognition (NER); Word Embeddings; Syntax Parsing; Topic Modeling; Dialogue Systems; Automatic Summarization; Machine Reading Comprehension; Language Models; Contextual Awareness; Multilingual Processing; Semantic Role Labeling; Knowledge Graphs; Information Retrieval. Retrieval; Information Extraction; Transfer Learning in NLP; Pattern Recognition; Computational Linguistics; Deep Learning in NLP; Reinforcement Learning in NLP; Natural Language Inference (NLI); Text Classification; Voice Interaction; Text Similarity, etc.

[0064] 3. Divide the target document (target language technology resource document) into a sentence set, and train and fine-tune the BERT model based on the language technology vocabulary to generate a weighted sentence vector for the sentence set. This includes:

[0065] Word vector extraction: First, each word in the sentence is converted into a word vector through the BERT model. Assume that the words in the sentence are w1, w2, ..., w n , the corresponding word vectors are v1,v2,...,v n , where v i is the vector representation of the i-th word generated by BERT.

[0066] Weighting formula: weight factors are defined according to the part of speech, frequency or semantic importance of language technology vocabulary, and a weighting formula is defined in combination with the weight factors; the expression of the weighting formula is:

[0067]

[0068] Among them, S is the weighted sentence vector, N is the number of words in the sentence, and a i is the weight factor of the i-th word, v i is the vector representation of the word generated by BERT, and f is a function defined according to the task requirements, such as TF-IDF (language technology term frequency-inverse document frequency) or part-of-speech weight.

[0069] Weighted calculation: Based on the language technology vocabulary and the word vector in the target document, a weighted sentence vector is generated using the weighted formula.

[0070] The language technology vocabulary and weighted formula generated by BERT can provide a more accurate representation for sentences. This representation can better capture the language technology semantic information of the sentence and provide higher accuracy in subsequent natural language processing tasks.

[0071] 4. Combining the BERT model and the TextRank algorithm, a confidentiality determination algorithm is obtained. The confidentiality determination algorithm is used to determine the confidentiality level of the target document through the similarity between the weighted sentence vector and the preset confidentiality determination rule.

[0072] The classification algorithm combines the word vectors generated by BERT with the TextRank algorithm for classification tasks in language technology resources. Based on the language technology vocabulary, BERT (Bidirectional Encoder Representations from Transformers) provides high-quality word vector representations, while the TextRank algorithm extracts information based on a graph model. Based on the language technology vocabulary, by combining BERT's word vectors with TextRank, sensitive information in the text can be more accurately identified.

[0073] like Figure 2 As shown in the figure, the logical steps of the encryption algorithm include:

[0074] 4.1 Text Preprocessing

[0075] The target document is segmented, stop words and punctuation marks are removed, and the processed text is encoded using the BERT model to obtain the context-related word vector of each word.

[0076] 4.2 Building a Graph Model

[0077] A language technology resource graph model is constructed using the context-related word vectors; in the language technology resource graph model, each node is a sentence, the edge is the relationship between nodes, and the weight of the edge is determined by the similarity of the context-related word vectors.

[0078] The edge weight is calculated as:

[0079]

[0080] Among them, v i and v j are the vectors of nodes i and j respectively, and sim(.,.) represents the cosine similarity of the vectors.

[0081] 4.3 Calculating Node Importance

[0082] The TextRank algorithm is used to calculate the ranking score of each node in the language technology resource graph model to obtain the importance of the node. The core idea of ​​TextRank is to determine the importance of a node by iteratively calculating its influence.

[0083] The calculation formula for node ranking score is:

[0084]

[0085] Among them, r i is the ranking score of node i, d is the damping coefficient, In(i) and Out(j) represent the in-edge and out-edge of node i respectively, and w ji is the weight of edge (j,i).

[0086] 4.4 Calculation of sensitivity scores

[0087] The preset confidentiality rules are used to evaluate the sensitivity of each node. The preset confidentiality rules contain various possible sensitive information, and the sensitivity score is set based on the task requirements. The sensitivity score calculation formula is:

[0088] S i =max(S i,base ,sim(v i ,v sensitive ))

[0089] Among them, Si S is the sensitivity score of word i. i,base As the basic sensitivity score, sim(v i ,v sensitive ) is the similarity between node i and the encryption rule.

[0090] 4.5 Weighted ranking score

[0091] The sensitivity score is then incorporated into the ranking calculation of TextRank to obtain the ranking score of the weighted sentence. The calculation formula for the ranking score of the weighted sentence is:

[0092] r i ′=r i ×S i

[0093] Among them, r i ′ is the weighted ranking score of the sentence, r i is the ranking score of the unweighted sentence, S i is the sensitivity score.

[0094] 4.6 Information Extraction and Classification

[0095] The information in the target document is extracted according to the ranking score of the weighted sentences, and the confidentiality score of the target document is calculated according to the confidentiality rules to obtain the document confidentiality level, and the high-ranking words and sentences in the target document are marked as sensitive information that requires access authorization protection.

[0096] The confidentiality determination algorithm based on the BERT word vector of the language technology vocabulary and TextRank provides an effective solution to improve the ability to identify sensitive information in text analysis by combining the word vector generated by BERT with the graph model and ranking mechanism of TextRank. The advantages of the algorithm include: High-quality word vectors: The word vectors generated by BERT can capture contextual information, improve the understanding of the semantics of language technology vocabulary, and make the relationship between vocabulary more accurate. Accurate graph construction: The graph model is constructed based on the similarity of word vectors, so that the edge weights between nodes can more accurately reflect the actual semantic relationship. Enhanced sensitivity identification: Through the introduction of confidentiality determination rules, the algorithm can effectively identify sensitive information in the text and improve the accuracy of confidentiality determination. Improved ranking mechanism: The TextRank algorithm combined with sensitivity scoring can more accurately evaluate the importance of nodes and improve the effect of information extraction.

[0097] 5. Experimental environment

[0098] 5.1 Hardware Environment

[0099] Computer: Intel Xeon CPU, 64GB RAM

[0100] GPU: NVIDIA Tesla V100

[0101] 5.2 Software Environment

[0102] Operating system: Ubuntu 20.04

[0103] Python version: 3.8

[0104] Deep Learning Framework: PyTorch 1.9

[0105] Main libraries: Transformers 4.6.1, Scikit-learn 0.24, NetworkX 2.5

[0106] 5.3 Dataset:

[0107] Training data: Use a large-scale language technology resource library.

[0108] Test data: A collection of language technology resources containing sensitive information used to evaluate the accuracy of classification algorithms.

[0109] 6. Experimental results analysis

[0110] In the experiment, the text was first preprocessed and the BERT model was used to generate a language technology vocabulary on the language technology resource library. Then, the text was classified using the BERT word vector and the improved TextRank algorithm. The following are the key results of the experiment:

[0111] In the iterative calculation of the improved TextRank algorithm, the damping coefficient is set to 0.85, and the experiment is carried out under different weighted sentence vector weights α and different confidentiality similarity sim conditions. The evaluation is carried out by calculating the accuracy of document confidentiality. The specific experimental results are as follows: Figure 3 As shown, it can be seen that, in general, under the same weight α, the accuracy increases with the increase of the density similarity sim value, reaching the maximum value at about sim = 0.63. Under the same sim, the overall accuracy increases with the increase of α. Because as α increases, the vocabulary weight based on the preset language technology vocabulary in the construction of the sentence vector becomes larger and larger. However, when α = 0.8, the accuracy does not continue to increase with the increase of α. Therefore, it is believed that when the pre-trained language technology vocabulary intervenes too much, the specific situation of the target document itself is ignored, resulting in the accuracy no longer increasing.

[0112] like Figure 4As shown in the figure, under the same data set, the traditional TextRank classification method is compared with the improved TextRank algorithm in terms of accuracy, precision and recall. It can be seen that the improved TextRank algorithm has improved accuracy, precision and recall compared with the traditional method.

[0113] Therefore, the present invention provides an intelligent management method for the data confidentiality of language technology resources, which can accurately extract key information in documents and realize automatic determination of document confidentiality in combination with the language technology vocabulary, thereby promoting the intelligent and automated management of language technology resources, improving the efficiency and quality of confidentiality work, and ensuring the security of language technology resources.

[0114] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0115] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for intelligent management of data confidentiality of language technology resources, characterized in that: The following steps are involved: Collect data related to language technology resources, and perform data screening, Chinese word segmentation, and stop word removal on the collected data to obtain the experimental data set; Use the BERT word vector method to perform pre-training on the experimental data set to generate a language technology vocabulary; Divide the target document into a sentence set, and train and fine-tune the BERT model based on the language technology vocabulary to generate weighted sentence vectors for the sentence set; The BERT model and the TextRank algorithm are combined to obtain a confidentiality determination algorithm, and the confidentiality determination algorithm is used to determine the confidentiality level of the target document through the similarity between the weighted sentence vector and the preset confidentiality determination rule.

2. The method for intelligent management of data confidentiality of language technology resources according to claim 1, characterized in that: The BERT word vector method is used to perform pre-training on the experimental dataset to generate a language technology vocabulary, including: Based on the experimental data set, a masked language model is used to predict masked words, and a next sentence prediction task is used to drive the masked language model to determine whether two sentences are adjacent in the original text, so as to perform pre-training of the BERT model; Select a labeled dataset and fine-tune the parameters of the pre-trained BERT model to obtain the BERT model; During the pre-training process of the BERT model, the WordPiece segmentation technology is used to segment the text and generate a vocabulary including common words and subword units; Based on the vocabulary, each word in the experimental data set is encoded to generate a related word vector for capturing the semantic features of the word; The vocabulary list and the related word vectors are combined to obtain a language technology vocabulary list.

3. The method for intelligent management of data confidentiality of language technology resources according to claim 2, characterized in that: The target document is divided into a sentence set, and based on the language technology vocabulary, the BERT model is used for training and fine-tuning to generate weighted sentence vectors for the sentence set, including: The target document is divided into a sentence set, and each word in the sentence set is converted into a word vector using the BERT model; Defining weight factors according to the part of speech, frequency or semantic importance of the language technology vocabulary, and defining a weighting formula in combination with the weight factors; Based on the language technology vocabulary in the target document and the word vector, a weighted sentence vector is generated using the weighting formula.

4. The method for intelligent management of data confidentiality of language technology resources according to claim 3, characterized in that: The expression of the weighted formula is: a i =f(w i ) Among them, S is the weighted sentence vector, N is the number of words in the sentence, and a i is the weight factor of the i-th word, v i is the vector representation of the th word generated by BERT, and f is a function defined according to the task requirements.

5. The method for intelligent management of data confidentiality of language technology resources according to claim 4, characterized in that: Combining the BERT model and the TextRank algorithm, a confidentiality determination algorithm is obtained. The confidentiality determination algorithm is used to determine the confidentiality level of the target document through the similarity between the weighted sentence vector and the preset confidentiality determination rule, including: Segment the target document, remove stop words and punctuation marks, and encode the processed text using the BERT model to obtain context-related word vectors for each word; A language technology resource graph model is constructed using the context-related word vectors; in the language technology resource graph model, each node is a sentence, the edge is the relationship between the nodes, and the weight of the edge is determined by the similarity of the context-related word vectors; Using the TextRank algorithm to calculate the ranking score of each node in the language technology resource graph model to obtain the importance of the node; The encryption rules are preset, sensitivity scores are given to each node, and the sensitivity scores are integrated into the ranking calculation of TextRank to obtain the ranking scores of weighted sentences; The information in the target document is extracted according to the ranking score of the weighted sentences, and the confidentiality score of the target document is calculated according to the confidentiality rules to obtain the document confidentiality level, and the high-ranking words and sentences in the target document are marked as sensitive information that requires access authorization protection.

6. The method for intelligent management of data confidentiality of language technology resources according to claim 5, characterized in that: The edge weight is calculated as: Among them, v i and v j are the vectors of nodes i and j respectively, and sim(.,.) represents the cosine similarity of the vectors.

7. The method for intelligent management of data confidentiality of language technology resources according to claim 6, characterized in that: The calculation formula for node ranking score is: Among them, r i is the ranking score of node i, d is the damping coefficient, In(i) and Out(j) represent the in-edge and out-edge of node i respectively, and w ji is the weight of edge (j,i).

8. The method for intelligent management of data confidentiality of language technology resources according to claim 7, characterized in that: The formula for calculating the ranking score of a weighted sentence is: S i =max(S i,base ,sim(v i ,v sensitive )) r i ′＝r i ×S i Among them, S i S is the sensitivity score of word i. i,base As the basic sensitivity score, sim(v i ,v sensitive ) is the similarity between node i and the secret rule, r i ′ is the weighted ranking score of the sentence, r i is the ranking score of the unweighted sentence.

Images