Method for preventing office information from being sent mistakenly based on AI technology

Through AI-based methods to prevent missend office information, natural language processing and deep learning identify information intentions and applicable objects, the problem of missed office information is solved, automated review and reduced risk of violations, and solutions to adapt to the needs of different enterprises are realized. .

CN118916909AInactive Publication Date: 2024-11-08NANJING TECH UNIV PUJIANG INST
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202410926053.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-11
Publication Date
2024-11-08
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The problem of missed sending office information is common in modern enterprises, resulting in leakage of sensitive information, leakage of commercial secrets, and increased legal risks, making it difficult for existing technologies to effectively identify and prevent them.

Method used

Adopt AI technology to prevent the missend of office information, through the personalized settings of user groups, AI inspection before information sending, and sensitive word filtering detection mechanism, combined with natural language processing and deep learning, identify information intentions and applicable objects, reduce human errors, and Determine whether the information contains sensitive words through word segmentation model, encoding and decoder.

Benefits of technology

It realizes automated information review, reduces employee burden, improves work efficiency, adapts to different user communication styles and enterprise needs, reduces the risk of violations, and can be integrated into existing IT infrastructure to expand and upgrade.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118916909A_ABST
    Figure CN118916909A_ABST
Patent Text Reader

Abstract

A method for preventing office information from being sent mistakenly based on an AI technology comprises the following steps that 1, a user selects options only receiving group information in group personality setting; 2, a user selects AI check options before information sending; 3, switching to an AI sensitive word filtering detection mechanism before information sending, identifying character information, and comparing the character information with a sensitive word library; and 4, if the sensitive word bank is met, popping up an early warning confirmation sending dialog box, and if the sensitive word bank is not met, directly turning to sending. According to the method, a perfect information sensitive word filtering and checking mechanism is established by adding new personalized settings and introducing components based on the AI technology and the sensitive word bank, so that the method has relatively high operability, the probability of mistakenly sending office information can be reduced to a great extent, and loss and harm caused by mistakenly sending the information are avoided.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of AI, and in particular to a method for preventing office information from being sent by mistake based on AI technology. Background Art

[0002] With the rapid development of artificial intelligence technology, office automation systems play an increasingly important role in modern enterprises. Communication tools such as email, instant messaging, and corporate social platforms have become the main ways for employees to communicate in their daily lives. However, the problem of mis-sending information has always been a common problem. According to a survey conducted by Tessian in 2019, on average, each employee mis-sends information at least once a month. This mistake may lead to serious consequences such as the leakage of sensitive information, the leakage of commercial secrets, the decline of customer trust, and the increase of legal risks. Therefore, the development of a method based on AI technology to prevent the mis-sending of office information has important practical significance and application value.

[0003] AI technology has great potential in preventing office information from being mis-sent. By applying technologies such as natural language processing, machine learning, and deep learning, it is possible to effectively identify and prevent the problem of mis-sending information. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a method for preventing office information from being sent by mistake based on AI technology. The technology can understand the complexity of language and contextual relationships, so as to more accurately judge the intention and applicable objects of the information, reduce human errors, and ensure that such information will not be accidentally sent to inappropriate recipients, thereby protecting the data security of the enterprise.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for preventing office information from being sent by mistake based on AI technology includes the following steps:

[0007] Step 1: In the group personalization settings, the user selects the option to only receive group messages;

[0008] Step 2: The user selects the AI ​​check option before sending the message;

[0009] Step 3: Enter the AI ​​sensitive word filtering detection mechanism before sending the information, identify the text information, and compare it with the sensitive word library;

[0010] Step 4: If the word matches the sensitive word library, a warning confirmation dialog box will pop up; if the word does not match the sensitive word library, the program will be sent directly.

[0011] Furthermore, the process of user group personality setting in step 1 is expressed as:

[0012] Step 1.1, the user chooses to only receive group messages. The user can only receive messages from the group and cannot send any messages. At this time, the AI ​​check option before sending messages is grayed out;

[0013] In step 1.2, if the user does not select to only receive group messages, the user can both receive and send group messages.

[0014] Furthermore, the process of the AI ​​sensitive word filtering detection mechanism in step 3 is expressed as follows:

[0015] Before AI inspection, user information needs to be segmented. A segmentation model is established through a statistical method model. The segmentation model is as follows:

[0016] Step 3.1.1, prepare the corpus, build a corpus of Chinese words that have been divided into words, and count the word frequency and the adjacency relationship between words;

[0017] Step 3.1.2, preprocess the corpus, including removing punctuation, unifying character encoding, converting to lowercase, etc.;

[0018] Step 3.1.3, traverse the preprocessed corpus and count the frequency of each word and the frequency of each pair of words;

[0019] Step 3.1.4, for each word, calculate the probability of its occurrence in the corpus:

[0020]

[0021] In the formula, w i represents the i-th word;

[0022] Step 3.1.5, for each pair of words (w i ,w j ), and calculate the probability of its occurrence in the corpus:

[0023]

[0024] Step 3.1.6, using the probabilities calculated in steps 3.1.4 and 3.1.5, build a Bigram language model, which is used to evaluate the probability of a word sequence;

[0025] Step 3.1.7, given an unsegmented sentence S = c1, c2, ..., c n , where c i is the i-th character in the sentence, setting the word segmentation algorithm target:

[0026]

[0027] Where W *represents the optimal word sequence, i.e., the word segmentation result; W represents the set of all possible word sequences; P(W) represents the probability of the word sequence W; represents the search for W that maximizes P(W); Π represents the multiplication symbol, which is used to calculate the product of multiple probabilities; P(w i |w i-1 ) means that when the previous word w is known i-1 Under the condition that the current word w i The conditional probability of occurrence; m represents the number of words in the word sequence W; i represents the iteration variable, which is used to traverse each word in the word sequence W;

[0028] Step 3.1.8, combined with the word segmentation algorithm goal of step 3.1.7, use the dynamic programming algorithm: Viterbi algorithm, calculate the optimal word segmentation path, and obtain the user information for word segmentation results.

[0029] Furthermore, the process of the AI ​​sensitive word filtering detection mechanism in step 3 is expressed as follows:

[0030] After the user information is segmented, each segmented word is encoded and decoded:

[0031] The encoder and decoder consist of a word embedding layer, a positional encoding, an attention layer, a position-aware feedforward layer, a residual connection layer, and a normalization layer;

[0032] Position encoding can indicate the relative position relationship between model words, and the expression is as follows:

[0033]

[0034] Where PE represents the position encoding of the word, pos represents the position of the word, 2i and 2i+1 represent the corresponding dimensions in the position encoding vector, and d corresponds to the total dimension of the position encoding;

[0035] The self-attention network is used to improve the model's ability to understand the complex relationships within the data. Its expression is as follows:

[0036]

[0037] In the formula, A represents the attention weight, sofmax() is the softmax function, Q is the output of the pooling layer, K is the key, V is the value, and d k is the rank of the key;

[0038] The position-aware feedforward layer accepts the output of the self-attention sublayer as input and is expressed as follows:

[0039] FFN(A)=Relu(AW3+b1)W4+b2

[0040] Where W3, W4, b1, b2 are the parameters of the position-aware feedforward layer;

[0041] The normalization layer can speed up the network training process and prevent the gradient from disappearing or exploding. LN normalization is used;

[0042] The residual connection layer allows the gradient to propagate directly to the deeper layers of the network, which helps to train deep networks. Its expression is as follows:

[0043] y=W2σ1(W1u)+W s u

[0044] Where y is the output of the residual connection layer, u is the input of the residual connection layer, W1 and W2 are weight matrices, σ1() is the activation function ReLU, and W s is the dimension-matched weight matrix;

[0045] The feedforward network can perform nonlinear transformation on features and enhance the expressiveness of the model;

[0046] O=σ2(∑W o y+b)

[0047] Where O is the output of the feedforward network layer, y is the input of the feedforward network layer, σ2() is the activation function Sigmoid, and W o is the feedforward network weight, b is the feedforward network layer bias;

[0048] Determine whether the user input information contains sensitive words based on the output classification results of encoding and decoding.

[0049] The method for preventing office information from being sent by mistake based on AI technology of the present invention has the following beneficial effects:

[0050] 1. The traditional information review of the present invention requires manual participation, while the technology of the present invention can automatically perform this process, reduce the burden on employees and improve work efficiency. ;

[0051] 2. The present invention continuously optimizes the model through machine learning to adapt to the communication styles of different users and the specific needs of enterprises, thereby improving the accuracy of recognition;

[0052] 3. For enterprises that need to comply with specific industry regulations and legal standards, the present invention can help ensure that all information sent complies with these requirements and reduce the risk of violations;

[0053] 4. The present invention can be easily integrated into the existing IT infrastructure and can be expanded and upgraded according to the development needs of the enterprise. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a flow chart of the present invention;

[0055] Figure 2 This is a network model diagram of the present invention. DETAILED DESCRIPTION

[0056] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments:

[0057] The present invention proposes a method for preventing office information from being sent by mistake based on AI technology. The invention flowchart is as follows: Figure 1 As shown in the figure, the network model is as follows Figure 2 As shown, the steps of the present invention are described in detail below.

[0058] Step 1: In the group personalization settings, the user selects the option to only receive group messages;

[0059] Step 1.1, the user chooses to only receive group messages. The user can only receive messages from the group and cannot send any messages. At this time, the AI ​​check option before sending messages is grayed out;

[0060] In step 1.2, if the user does not select to only receive group messages, the user can both receive and send group messages.

[0061] Step 2: The user selects the AI ​​check option before sending the message;

[0062] Step 2.1: If the user does not select AI check before sending messages, group message sending will remain in the default state, and messages can be sent without any restrictions within the group.

[0063] Step 2.2: User selects AI check before sending information.

[0064] Step 3: Enter the AI ​​sensitive word filtering detection mechanism before sending the information, identify the text information, and compare it with the sensitive word library;

[0065] Before AI inspection, user information needs to be segmented. A segmentation model is established through a statistical method model. The segmentation model is as follows:

[0066] Step 3.1.1, prepare the corpus, build a corpus of Chinese words that have been divided into words, and count the word frequency and the adjacency relationship between words;

[0067] Step 3.1.2, preprocess the corpus, including removing punctuation, unifying character encoding, converting to lowercase, etc.;

[0068] Step 3.1.3, traverse the preprocessed corpus and count the frequency of each word and the frequency of each pair of words;

[0069] Step 3.1.4, for each word, calculate the probability of its occurrence in the corpus:

[0070]

[0071] In the formula, w i represents the i-th word;

[0072] Step 3.1.5, for each pair of words (w i ,w j ), and calculate the probability of its occurrence in the corpus:

[0073]

[0074] Step 3.1.6, using the probabilities calculated in steps 3.1.4 and 3.1.5, build a Bigram language model, which is used to evaluate the probability of a word sequence;

[0075] Step 3.1.7, given an unsegmented sentence S = c1, c2, ..., c n , where c i is the i-th character in the sentence, setting the word segmentation algorithm target:

[0076]

[0077] Where W * represents the optimal word sequence, i.e., the word segmentation result; W represents the set of all possible word sequences; P(W) represents the probability of the word sequence W; represents the search for W that maximizes P(W); Π represents the multiplication symbol, which is used to calculate the product of multiple probabilities; P(w i |w i-1 ) means that when the previous word w is known i-1 Under the condition that the current word w i The conditional probability of occurrence; m represents the number of words in the word sequence W; i represents the iteration variable, which is used to traverse each word in the word sequence W;

[0078] Step 3.1.8, combined with the word segmentation algorithm goal of step 3.1.7, use the dynamic programming algorithm: Viterbi algorithm, calculate the optimal word segmentation path, and obtain the user information for word segmentation results.

[0079] After the user information is segmented, each segmented word is encoded and decoded:

[0080] The encoder and decoder consist of a word embedding layer, a position encoding layer, an attention layer, a position-aware feedforward layer, a residual connection layer, and a normalization layer. The network model is shown in the figure below. Figure 2 As shown;

[0081] Position encoding can indicate the relative position relationship between model words, and the expression is as follows:

[0082]

[0083] Where PE represents the position encoding of the word, pos represents the position of the word, 2i and 2i+1 represent the corresponding dimensions in the position encoding vector, and d corresponds to the total dimension of the position encoding;

[0084] The self-attention network is used to improve the model's ability to understand the complex relationships within the data. Its expression is as follows:

[0085]

[0086] In the formula, A represents the attention weight, sofmax() is the softmax function, Q is the output of the pooling layer, K is the key, V is the value, and d k is the rank of the key;

[0087] The position-aware feedforward layer accepts the output of the self-attention sublayer as input and is expressed as follows:

[0088] FFN(A)=Relu(AW3+b1)W4+b2

[0089] Where W3, W4, b1, b2 are the parameters of the position-aware feedforward layer;

[0090] The normalization layer can speed up the network training process and prevent the gradient from disappearing or exploding. LN normalization is used;

[0091] The residual connection layer allows the gradient to propagate directly to the deeper layers of the network, which helps to train deep networks. Its expression is as follows:

[0092] y=W2σ1(W1u)+W s u

[0093] Where y is the output of the residual connection layer, u is the input of the residual connection layer, W1 and W2 are weight matrices, σ1() is the activation function ReLU, and W s is the dimension-matched weight matrix;

[0094] The feedforward network can perform nonlinear transformation on features and enhance the expressiveness of the model;

[0095] O=σ2(∑W o y+b)

[0096] Where O is the output of the feedforward network layer, y is the input of the feedforward network layer, σ2() is the activation function Sigmoid, and W o is the feedforward network weight, b is the feedforward network layer bias;

[0097] Determine whether the user input information contains sensitive words based on the output classification results of encoding and decoding.

[0098] Step 4: If the word matches the sensitive word library, a warning confirmation dialog box will pop up; if the word does not match the sensitive word library, the program will be sent directly.

[0099] The above description is only a preferred embodiment of the present invention and does not constitute any other form of limitation to the present invention. Any modification or equivalent change made based on the technical essence of the present invention still falls within the scope of protection required by the present invention.

Claims

1. A method based on AI technology to prevent office information from being sent by mistake. It is characterized in that The following steps are involved: Step 1: In the group personalization settings, the user selects the option to only receive group messages; Step 2: The user selects the AI ​​check option before sending the message; Step 3: Enter the AI ​​sensitive word filtering detection mechanism before sending the information, identify the text information, and compare it with the sensitive word library; Step 4: If the word matches the sensitive word library, a warning confirmation dialog box will pop up; if the word does not match the sensitive word library, the program will be sent directly.

2. According to claim 1, a method for preventing office information from being sent by mistake based on AI technology, Features: The process of user group personality setting in step 1 is expressed as follows: Step 1.1, the user chooses to only receive group messages. The user can only receive messages from the group and cannot send any messages. At this time, the AI ​​check option before sending messages is grayed out; In step 1.2, if the user does not select to only receive group messages, the user can both receive and send group messages.

3. According to claim 1, a method for preventing office information from being sent by mistake based on AI technology, Features: The process of AI sensitive word filtering detection mechanism in step 3 is as follows: Before AI inspection, user information needs to be segmented. A segmentation model is established through a statistical method model. The segmentation model is as follows: Step 3.1.1, prepare the corpus, build a corpus of Chinese words that have been divided into words, and count the word frequency and the adjacency relationship between words; Step 3.1.2, preprocess the corpus, including removing punctuation, unifying character encoding, converting to lowercase, etc.; Step 3.1.3, traverse the preprocessed corpus and count the frequency of each word and the frequency of each pair of words; Step 3.1.4, for each word, calculate the probability of its occurrence in the corpus: In the formula, w i represents the i-th word; Step 3.1.5, for each pair of words (w i ,w j ), and calculate the probability of its occurrence in the corpus: Step 3.1.6, using the probabilities calculated in steps 3.1.4 and 3.1.5, build a Bigram language model, which is used to evaluate the probability of a word sequence; Step 3.1.7, given an unsegmented sentence S = c 1 ,c 2 ,...,c n , where c i is the i-th character in the sentence, setting the word segmentation algorithm target: Where W * represents the optimal word sequence, i.e., the word segmentation result; W represents the set of all possible word sequences; P(W) represents the probability of the word sequence W; represents the search for W that maximizes P(W); Π represents the multiplication symbol, which is used to calculate the product of multiple probabilities; P(w i |w i-1 ) means that when the previous word w is known i-1 Under the condition that the current word w i Conditional probability of occurrence; m represents the number of words in the word sequence W; i represents the iteration variable, which is used to traverse each word in the word sequence W; Step 3.1.8, combined with the word segmentation algorithm goal of step 3.1.7, use the dynamic programming algorithm: Viterbi algorithm, calculate the optimal word segmentation path, and obtain the user information for word segmentation results.

4. According to claim 1, a method for preventing office information from being sent by mistake based on AI technology, Features: The process of AI sensitive word filtering detection mechanism in step 3 is as follows: After the user information is segmented, each segmented word is encoded and decoded: The encoder and decoder consist of a word embedding layer, a positional encoding, an attention layer, a position-aware feedforward layer, a residual connection layer, and a normalization layer; Position encoding can indicate the relative position relationship between model words, and the expression is as follows: Where PE represents the position encoding of the word, pos represents the position of the word, 2i and 2i+1 represent the corresponding dimensions in the position encoding vector, and d corresponds to the total dimension of the position encoding; The self-attention network is used to improve the model's ability to understand the complex relationships within the data. Its expression is as follows: In the formula, A represents the attention weight, sofmax() is the softmax function, Q is the output of the pooling layer, K is the key, V is the value, and d k is the rank of the key; The position-aware feedforward layer accepts the output of the self-attention sublayer as input and is expressed as follows: FFN(A))Relu(AW). 3 +b 1 )W 4 +b 2 Where W 3 , W 4 、b 1 、b 2 are the parameters of the position-aware feed-forward layer; The normalization layer can speed up the network training process and prevent the gradient from disappearing or exploding. LN normalization is used; The residual connection layer allows the gradient to propagate directly to the deeper layers of the network, which helps to train deep networks. Its expression is as follows: y=W 2 σ 1 (W 1 u)+W s u Where y is the output of the residual connection layer, u is the input of the residual connection layer, and W 1 and W 2 is the weight matrix, σ 1 () is the activation function ReLU, W s is the dimension-matched weight matrix; The feedforward network can perform nonlinear transformation on features and enhance the expressiveness of the model; O=s 2 (∑W o y+b) Where O is the output of the feedforward network layer, y is the input of the feedforward network layer, σ 2 () is the activation function Sigmoid, W o is the feedforward network weight, b is the feedforward network layer bias; Determine whether the user input information contains sensitive words based on the output classification results of encoding and decoding.

Citation Information

Patent Citations

  • Information monitoring method and device, terminal and storage medium

    CN112396071A

  • Text processing method and device

    CN114138969A

  • Chinese word segmentation method, Chinese word segmentation device, electronic equipment and storage medium

    CN117493485A

  • Information identification method and device, equipment, storage medium and program product

    CN118194131A