Risk compliance type intelligent interaction method and system

Abstract

The invention relates to the technical field of intelligent interaction, solves the technical problems that an answer with a legal risk is easily given when artificial intelligence is matched with artificial customer service, and the answer is difficult to be satisfied and satisfactorily given to a customer when artificial intelligence of the artificial customer service is not added, and particularly relates to a risk compliance type intelligent interaction method and system. The method comprises the following steps: S1, constructing a trust domain retrieval library and an unpopular problem database based on historical data; s2, acquiring a plurality of pieces of question information input by a client, and preprocessing the question information to obtain a plurality of pieces of question data Ca, and S3, converting the question data C (a) into question representation f (a) through the convolutional neural network. Through detection of the bottom end, a proper template can be found in the unpopular question database, and the accuracy of the unpopular question database is improved. In the detection process, keywords are recognized and found out through a keyword search algorithm, and appropriate answers are found out in the unpopular question database.

Description

Technical Field

[0001] The present invention relates to the field of intelligent interaction technology, and in particular to a risk compliance intelligent interaction method and system. Background Art

[0002] Compliance generally refers to the consistency of commercial banks' business activities with laws, rules and standards. During the company's system operation, in order to prevent the company from making legal mistakes and causing losses, the company will choose some systems that identify and avoid risky behaviors during operation. In the process of human-computer interaction between the system and the customer, the language given by the system is given by artificial intelligence through calculations, and it is easy for unpopular questions to be difficult to answer. In order to solve this problem, many companies have added emergency manual options to the interactive system. However, this method not only requires manual participation, which increases labor costs, but also the manual customer service's mastery of legal knowledge cannot fully meet the standards, which makes the system prone to legal risks. Therefore, the use of artificial intelligence in combination with manual customer service is prone to giving answers with legal risks. Without the addition of manual customer service, artificial intelligence is difficult to give customers compliant and satisfactory answers. Summary of the invention

[0003] In view of the deficiencies in the prior art, the present invention provides a risk compliance intelligent interaction method and system, which solves the technical problem that artificial intelligence combined with manual customer service can easily give answers with legal risks, while without the addition of manual customer service, artificial intelligence is difficult to give customers compliant and satisfactory answers. The method achieves the purpose of answering unpopular questions given by customers by combining artificial intelligence with a trust domain retrieval library and a multi-dimensional review library, and using a keyword recognition algorithm, and ensuring that the wording of the answer is compliant before answering.

[0004] To solve the above technical problems, the present invention provides the following technical solutions: a risk compliance intelligent interaction method, the method comprising the following steps:

[0005] S1. Build a trust domain retrieval library and unpopular question database based on historical data;

[0006] S2. Obtaining several question information input by the client, and preprocessing the question information to obtain several question data C a ;

[0007] S3, convert the question data C(a) into question representation f(a) through a convolutional neural network, and obtain question representation d(b) similar to question representation f(a) from the trust domain retrieval library;

[0008] S4. Calculate the content similarity value sim based on f(a) and f(b) cos and the average similarity Determine whether there is a matching answer in the trust domain search library;

[0009] like If they are not similar, a request is sent to the bottom end;

[0010] like Similarly, the answer matching the question representation f(b) in the trust domain retrieval library is sent to the review end;

[0011] S5. The audit end receives the answer and calculates the key value P i Obtain keywords, compare the keywords with the words in the multi-dimensional audit library established based on network information, and determine whether they are compliant words;

[0012] If yes, a reply is sent to the client;

[0013] If not, return to S2 until the wording in the answer complies with the regulations;

[0014] S6. After receiving the request, the back-end sends the answer template formed based on the unpopular question database to the client according to the customer's question information;

[0015] S7. Regularly perform vulnerability scans on the scan targets and provide vulnerability detection reports.

[0016] Preferably, in step S1, the specific implementation steps are as follows:

[0017] S11. Obtain the number Z of all questions asked by customers in historical data and the number C of repeated questions, and calculate the question repetition rate P c , the calculation formula is as follows:

[0018]

[0019] Among them, P c represents the repetition rate of questions, C represents the number of repeated questions, and Z represents the number of all questions asked by customers;

[0020] S12, determining whether the problem is a common problem;

[0021] If P c ≥0.5, it is marked as a common problem;

[0022] If P c <0.5, it is marked as an unpopular question and a database of unpopular questions is constructed;

[0023] S13. Build a trust domain search library based on frequently asked questions, government websites, official media and encyclopedia knowledge. The trust domain search library includes questions and matching answers.

[0024] Preferably, in step S2, the specific implementation steps are as follows:

[0025] S21, using the NLPIR word segmenter to perform word segmentation and tagging on the obtained question information and obtain tagging information;

[0026] S22. Use the extractive attention mechanism method to compress the redundant information in the annotation information to obtain the question data C a .

[0027] Preferably, in step S3, the specific implementation steps are as follows:

[0028] S31. Setting the analysis core D a , the analysis kernel D a and question data C a Perform convolution operation to obtain convolution data B a ;

[0029] S32, determine the pooling kernel H c , the pooling kernel H c Convolution data B a Perform pooling operation to obtain pooling data G b ;

[0030] S33, using sigmoid function as activation function, the pooled data G b Substitute the activation function for full connection calculation to obtain the question representation f(a), the calculation formula is as follows:

[0031] f(a)=K(w b G b +d)

[0032] Among them, f(a) represents the question representation, G b represents the bth pooled data, w b Represents the weight coefficient in the activation function, K represents the activation function, and d represents the bias coefficient;

[0033] S34, obtaining similar question data from the trust domain retrieval library and converting it into a question representation f(b) using the method of steps S31 to S33.

[0034] Preferably, in step S4, the specific implementation steps are as follows:

[0035] S41. Calculate the content similarity value sim between the question representation f(a) and the question representation f(b) cos ;

[0036] S42. Obtain the historical content similarity value sim in the past six months his , calculate the average similarity value by the average method The calculation formula is as follows:

[0037]

[0038] in, Represents the average similarity value, sim his Indicates the historical content similarity value, n represents sim his The number of

[0039] Preferably, the content similarity value sim cos The calculation formula is as follows:

[0040]

[0041] Among them, sim cos represents the content similarity value, f(a) represents the question representation input by the customer, and f(b) represents the question representation obtained from the trust domain retrieval library.

[0042] Preferably, in step S5, the specific implementation steps are as follows:

[0043] S51. A multi-dimensional audit database is established based on Internet dictionaries, news sensitive words and illegal words. The audit end obtains the question data C corresponding to the dissimilar question representation f(a) a ;

[0044] S52, obtaining a number of pre-processed question data C a The frequency e of word i in i , calculate the word frequency value fre of word i i , the calculation formula is as follows:

[0045]

[0046] Among them, fre i Indicates the word frequency value, e i Indicates the frequency of occurrence of word i;

[0047] S53. Obtaining question data C a The frequency u of the sentence beginning of word i in i 、The frequency of the end of the sentence v i and the frequency in the sentence t i , calculate the question data C a The position frequency of word i in add i , the calculation formula is as follows:

[0048]

[0049] Among them, add i Indicates the position frequency of the word, u i represents the frequency of word i appearing at the beginning of the sentence, v i represents the frequency of word i appearing at the end of the sentence, t i Indicates the frequency of word i appearing in the sentence;

[0050] S54, obtaining the total number of words tw of the Internet dictionary and the frequency O of the word i in the Internet dictionary i Sum frequency e i And calculate the frequency value select i , the calculation formula is as follows:

[0051]

[0052] Among them, select i represents the frequency value of word i, tw represents the total number of words in the Internet dictionary;

[0053] S55, according to the word frequency value fre i 、Position frequencyadd i and frequency value select i Calculate the key value P of word i i , the calculation formula is as follows:

[0054] P i =A×fre i +B×add i +C×select i

[0055] Among them, P i represents the key value, A, B and C represent the weight coefficients respectively;

[0056] S56, setting keyword threshold P based on adaptive threshold method g , determine whether word i is a keyword;

[0057] If P i ≥P g , then it is a keyword and is compared with the words in the multi-dimensional audit library;

[0058] If P i <P g , it is not a keyword and is ready to detect the next word.

[0059] Preferably, in step S6, the specific implementation steps are as follows:

[0060] S61. Establishing an answer template based on unpopular questions in the unpopular question database;

[0061] S62: After receiving the request, the back-end finds similar questions and corresponding answer templates in the unpopular question database according to the question information;

[0062] S63. Send the answer template to the client.

[0063] Preferably, in step S7, the specific implementation steps are as follows:

[0064] S71, constructing a timing mechanism in the startup program of the vulnerability scanning program through a time wheel algorithm;

[0065] S72. Determine regular scanning targets in the vulnerability scanning program;

[0066] S73. Give a vulnerability detection report after the scan is completed.

[0067] The technical solution also provides a system for the above-mentioned risk compliance intelligent interaction method, which includes:

[0068] Database module, which builds a trust domain search library and a database of unpopular questions based on the questions and answers of customers in the past three years;

[0069] The preprocessing module obtains several question information input by the client and preprocesses the question information to obtain several question data C a ;

[0070] The intelligent processing module converts the question data C(a) into question representation f(a) through a convolutional neural network, and obtains similar question representation f(b) from the trust domain retrieval library;

[0071] Matching module, calculates the content similarity value sim based on f(a) and f(b) cos and the average similarity Determine whether there is a matching answer in the trust domain search library;

[0072] like If they are not similar, a request is sent to the bottom end;

[0073] like Similarly, the answer matching the question representation f(b) in the trust domain retrieval library is sent to the review end;

[0074] In the compliance review module, the audit end receives the answer and calculates the key value P i Obtain keywords, compare the keywords with the words in the multi-dimensional audit library established based on network information, and determine whether they are compliant words;

[0075] If yes, a reply is sent to the client;

[0076] If not, return to S2 until the wording in the answer complies with the regulations;

[0077] The backup module, after receiving the request, sends the answer template formed based on the database of unpopular questions to the client according to the customer's question information;

[0078] The vulnerability scanning module regularly performs vulnerability scans on the scan targets and provides vulnerability detection reports.

[0079] Through the above technical solution, the present invention provides a risk compliance intelligent interaction method and system, which has at least the following beneficial effects:

[0080] 1. The present invention processes and compares the processed question information through convolutional neural networks and similarity value comparison, extracts the features in the question information and calculates the similarity values ​​of the features, and can quickly compare the matching question results in the trust domain retrieval library based on the similarity values. If there is no result, a backup request can be sent to the backup end to detect and answer unpopular questions.

[0081] 2. The present invention can detect unmatched question information and search for suitable templates in the unpopular question database through bottom-end detection. During the detection process, the keyword search algorithm is used to quickly identify and find keywords, and find suitable answers in the unpopular question database based on the keyword information.

[0082] 3. Through the function of the unpopular question database, the present invention can quickly obtain the corresponding answer template in the database according to the keyword after obtaining the keyword information, and send the answer template to the client, thereby avoiding the situation where the client's question is too unpopular to be answered, thereby improving the customer experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0084] Figure 1 This is a flow chart of a risk compliance intelligent interaction method of the present invention;

[0085] Figure 2 This is a structural block diagram of a risk compliance intelligent interactive system of the present invention. DETAILED DESCRIPTION

[0086] In order to make the above-mentioned purposes, 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 implementation methods, so that the implementation process of how the present application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0087] Since artificial intelligence combined with human customer service can easily give answers with legal risks, and artificial intelligence without human customer service is difficult to answer customers in a compliant and satisfactory manner, please refer to Figure 1 and Figure 2This embodiment provides a risk compliance intelligent interaction method, which can use artificial intelligence, trust domain search library and multi-dimensional audit library to answer unpopular questions given by customers using keyword recognition algorithm, and ensure that the answer words are compliant before answering. The method includes the following steps:

[0088] S1. Build a trust domain retrieval library and a database of unpopular questions based on historical data to facilitate subsequent judgments. The specific implementation steps are as follows:

[0089] S11. Obtain the number Z of all questions asked by customers in historical data and the number C of repeated questions, and calculate the question repetition rate P c , the calculation formula is as follows:

[0090]

[0091] Among them, P c represents the repetition rate of questions, C represents the number of repeated questions, and Z represents the number of all questions raised by customers. The number of all questions raised by customers Z and the number of repeated questions C can be directly obtained. By automatically identifying and deleting repeated content, our system ensures that stationery information can convey key information, avoid redundancy and improve efficiency.

[0092] S12, determine whether the problem is a common problem, if P c ≥0.5, it is marked as a common problem. c <0.5, it is marked as an unpopular question and a database of unpopular questions is constructed;

[0093] S13. Build a trust domain search library based on frequently asked questions, government websites, official media and encyclopedia knowledge. The trust domain search library includes questions and matching answers.

[0094] It should be noted that the present invention can construct a trust domain retrieval library based on public government websites, such as the announcements regularly released on the Beijing Government Website and the Guangdong Provincial Government Website. Official media refers to official media, such as Xinhua News Agency and People's Daily, in which the public information content is periodically updated and highly accurate. Encyclopedia knowledge generally refers to Baidu Encyclopedia and "Encyclopedia Knowledge: China Volume". These contents have strong reference value and cover a wide variety of issues, ensuring that security issues such as politics, hot public opinions, etc. can be accurately answered according to official standards.

[0095] S2. Obtaining several question information input by the client, and preprocessing the question information to obtain several question data C a , in order to facilitate the processing of question information and the judgment of subsequent steps, the specific implementation steps are as follows:

[0096] S21. Use NLPIR word segmenter to segment and annotate the obtained question information and obtain annotated information. NLPIR word segmenter is a common word segmentation tool. In the present invention, only NLPIR word segmenter is needed to complete the processing of stationery information, and no further elaboration is needed.

[0097] S22. Use the extractive attention mechanism method to compress the redundant information in the annotation information to obtain the question data C a The attention mechanism method is a mechanism widely used in deep learning, which aims to help the model process information more effectively. The extraction attention mechanism method is based on the information extraction and compression method combined with the attention mechanism. It extracts redundant information according to the appearance law of redundant information in the information, which can improve the processing speed of information.

[0098] S3. The question data C(a) is converted into question representation f(a) through a convolutional neural network, and question representation f(b) similar to question representation f(a) is obtained from the trust domain retrieval library to facilitate the subsequent judgment of similarity values. The specific implementation steps are as follows:

[0099] S31. Setting analysis core D a , the analysis kernel D a and question data C a Perform convolution operation to obtain convolution data B a ;

[0100] S32, determine the pooling kernel H c , the pooling kernel H c Convolution data B a Perform pooling operation to obtain pooling data G b ;

[0101] S33, using sigmoid function as activation function, the pooled data G b Substitute the activation function for full connection calculation to obtain the question representation f(a), the calculation formula is as follows:

[0102] f(a)=K(w b G b +d)

[0103] Among them, f(a) represents the question representation, G b represents the bth pooled data, w b Represents the weight coefficient in the activation function, K represents the activation function, d represents the bias coefficient, and the weight coefficient can be obtained through the hierarchical analysis method;

[0104] S34, obtaining similar question data from the trust domain retrieval library and converting it into a question representation f(b) using the method of steps S31 to S33.

[0105] It should be noted that the convolutional neural network is a type of feedforward neural network that includes convolution calculations and has a deep structure. It is one of the representative algorithms of deep learning. The present invention uses the convolution operations, pooling operations, and full connection operations commonly used in convolutional neural networks to obtain the question representation f(a). There is no need to elaborate on this. In terms of the selection of activation functions, we can use the commonly used Tanh function, ReLU function, and sigmoid function. The present invention uses the sigmoid function as the activation function to obtain the question representation f(a).

[0106] S4. Calculate the content similarity value sim based on f(a) and f(b) cos and the average similarity Determine whether there is a matching answer in the trust domain search library. If If they are not similar, a request is sent to the bottom end. Similarly, the answer matching the question representation f(b) in the trust domain search library is sent to the review end. The specific implementation steps are as follows:

[0107] S41. Calculate the content similarity value sim between the question representation f(a) and the question representation f(b) cos ;

[0108] S42. Obtain the historical content similarity value sim in the past six months his , calculate the average similarity value by the average method The calculation formula is as follows:

[0109]

[0110] in, Represents the average similarity value, sim his Indicates the historical content similarity value, n represents sim his The number of similarity values ​​sim cos The calculation formula is as follows:

[0111]

[0112] Among them, sim cosIndicates the content similarity value, f(a) represents the question representation input by the customer, and f(b) represents the question representation obtained from the trust domain search library. When similar answers cannot be found in the trust domain search library, the bottom-up end searches for unpopular question templates based on keyword searches in the multi-dimensional review library and the unpopular question database. This can avoid the situation where the customer's questions cannot be answered. When similar answers are found in the trust domain search library, the review end further reviews the answers. Since the answers in the trust domain search library have certain legal risks, the review end can provide further compliance guarantees. The question representation f(a) containing the features is obtained in step S3, and their similarity is calculated through similarity value calculation, and then it is determined whether the answer matching the customer's question can be found in the trust domain search library.

[0113] S5. The audit end receives the answer and calculates the key value P i Get the keyword, compare the keyword with the vocabulary in the multi-dimensional audit library established based on network information, and determine whether it is a compliant word. If so, send an answer to the client. If not, return to S2 until the wording in the answer meets the regulations. The client refers to the program corresponding to the server that provides local services to customers. The specific implementation steps are as follows:

[0114] S51. A multi-dimensional audit database is established based on Internet dictionaries, news sensitive words and illegal words. The audit end obtains the question data C corresponding to the dissimilar question representation f(a) a The multi-dimensional audit database is an audit database built based on the trial results of legal cases in the Internet court and commonly used legal knowledge, which can ensure that the answers sent comply with legal regulations.

[0115] S52, obtaining a number of pre-processed question data C a The frequency e of word i in i , calculate the word frequency value fre of word i i , the calculation formula is as follows:

[0116]

[0117] Among them, fre i Indicates the word frequency value, e i Indicates the frequency of occurrence of word i, the word frequency value fre i The vocabulary can be evaluated based on the frequency. The higher the frequency, the more critical the vocabulary is and the more likely it is to be a keyword.

[0118] S53. Obtaining question data C a The frequency u of the sentence beginning of word i in i 、The frequency of the end of the sentence v i and the frequency in the sentence t i , calculate the question data Ca The position frequency of word i in add i , the calculation formula is as follows:

[0119]

[0120] Among them, add i Indicates the position frequency of the word, u i represents the frequency of word i appearing at the beginning of the sentence, v i represents the frequency of word i appearing at the end of the sentence, t i Indicates the frequency of word i appearing in the sentence, position frequency add i The evaluation is based on the frequency of the word's position. The beginning of a sentence can better reflect the importance of a word, the end of a sentence is weaker than the beginning, and the middle of a sentence is the weakest. The weights assigned to them are 5, 4, and 1 respectively.

[0121] S54, obtaining the total number of words tw of the Internet dictionary and the frequency O of the word i in the Internet dictionary i Sum frequency e i And calculate the frequency value select i , the calculation formula is as follows:

[0122]

[0123] Among them, select i represents the frequency value of word i, tw represents the total number of words in the Internet dictionary;

[0124] S55, according to the word frequency value fre i 、Position frequencyadd i and frequency value select i Calculate the key value P of word i i , the calculation formula is as follows:

[0125] P i =A×fre i +B×add i +C×select i

[0126] Among them, P i Indicates the key value, A, B and C represent the weight coefficients, and the frequency value select i The evaluation is carried out based on the frequency of the words in the dictionary. The Internet dictionary is a dictionary compiled based on the words on the Internet. The frequency of appearance in the dictionary can also reflect the criticality of the words. A, B and C can be obtained through hierarchical analysis.

[0127] S56, setting keyword threshold P based on adaptive threshold method g , determine whether word i is a keyword;

[0128] If P i ≥P g , then it is a keyword and is compared with the words in the multi-dimensional audit library;

[0129] If P i <P g , it is not a keyword and is ready to detect the next word.

[0130] It should be noted that keywords are obtained based on the characteristic calculation of keywords to facilitate comparison with the multi-dimensional review library. During the review process, we ensure the objectivity, political neutrality, accuracy and comprehensiveness of information transmission, and have established multiple review aspects, such as review of political content, review of illegal content, review of criminal content, review of bad values, review of prompt word injection and protection of user privacy. In order to make the review comprehensive during the review process, we can also conduct cross-platform support review with cooperative platforms.

[0131] S6. After receiving the request, the bottom end sends the answer template formed based on the unpopular question database to the client according to the customer's question information. The specific implementation steps are as follows:

[0132] S61. Establish answer templates based on unpopular questions in the unpopular question database; In order to improve the efficiency of answering unpopular questions, answer templates are constructed based on the answers to unpopular questions. For example, a customer asks: If high-risk and high-yield products appear in asset allocation, how to allocate assets? Answer template: In asset allocation, if high-risk and high-yield products appear, how to allocate assets mainly depends on the investor's risk preference, investment objectives, investment period and overall financial situation. The following are some specific reference directions for you. Risk tolerance: First of all, investors need to clarify their risk tolerance. If the risk tolerance is high, the allocation ratio of high-risk and high-yield products can be appropriately increased; otherwise, the ratio should be kept low or avoid allocating such products. Investment objectives: Investment objectives are also an important factor in determining asset allocation. If the investor's goal is to pursue high returns and is willing to bear the corresponding risks, then the allocation ratio of high-risk and high-yield products can be increased accordingly. Investment period: Long-term investors are usually more able to withstand short-term fluctuations, so the allocation ratio of high-risk and high-yield products can be appropriately increased. Short-term investors may need to pay more attention to the liquidity and security of funds and diversify their investments: Regardless of their risk tolerance, they should follow the principle of "don't put all your eggs in one basket" and reduce overall risk by diversifying their investments.

[0133] S62: After receiving the request, the back-end finds similar questions and corresponding answer templates in the unpopular question database according to the question information;

[0134] S63. Send the answer template to the client.

[0135] It should be noted that the answer template includes two template implementation methods: preset response templates, designing and maintaining catch-all response templates, covering common sensitive topics and issues, and ensuring that the output responses always meet compliance requirements; intelligent response strategies, combining machine learning algorithms and semantic analysis technology, dynamically adjust and optimize the content and logic of catch-all responses to adapt to different scenarios and user needs.

[0136] S7. Regularly scan the target for vulnerabilities and provide vulnerability detection reports. The specific implementation steps are as follows:

[0137] S71, constructing a timing mechanism in the startup program of the vulnerability scanning program through a time wheel algorithm;

[0138] S72. Determine regular scanning targets in the vulnerability scanning program;

[0139] S73. Give a vulnerability detection report after the scan is completed.

[0140] It should be noted that in order to achieve endogenous security enhancement, we have redesigned the product functions to support the provision of a closed service loop through "big model security assessment + security alignment + security protection + security knowledge enhancement". We use automated tools and manual review to systematically scan and evaluate the security vulnerabilities of big models. We use technical means such as stress testing and simulated attacks to evaluate the stress resistance and performance of big models in the face of external attacks, implement multi-level security protection measures, and effectively protect big models from external attacks and internal abuse threats. We enhance the security awareness and skills of teams and users so that they can actively participate in and support the security protection of big models, protect the security and privacy of user data, follow strict data management and usage specifications, and conduct comprehensive evaluations of big model services from multiple dimensions to discover potential security risks and vulnerabilities. The two implementation methods are data set coverage and diversified attack simulation. Among them, data set coverage covers hundreds of thousands of diversified evaluation data sets, including text, images, and multimedia content, to simulate various usage scenarios in the real world; diversified attack simulation uses advanced prompt word attack instructions and other attack simulation technologies to detect the response and processing capabilities of big models to malicious inputs.

[0141] See also Figure 2 , which is a structural block diagram of the risk compliance intelligent interactive system provided in this embodiment. The intelligent interactive system includes: a database module, a preprocessing module, an intelligent processing module, a matching module, a compliance review module, a bottom-up module, and a vulnerability scanning module.

[0142] The database module builds a trust domain retrieval library and a database of unpopular questions based on the questions and answers of customers in the past three years. The preprocessing module obtains several question information input by the client and preprocesses the question information to obtain several question data C a , the intelligent processing module converts the question data C(a) into question representation f(a) through the convolutional neural network, obtains similar question representation f(b) from the trust domain retrieval library, and the matching module calculates the content similarity value sim based on f(a) and f(b) cos and the average similarity Determine whether there is a matching answer in the trust domain search library. If If they are not similar, a request is sent to the bottom end. Similarly, the answer matching the question representation f(b) in the trust domain search library is sent to the audit end, the compliance review module, and the audit end receives the answer and calculates the key value P i Get the keywords, compare the keywords with the words in the multi-dimensional audit library established based on network information, and determine whether they are compliant words. If so, send the answer to the client. If not, return to S2 until the words in the answer meet the regulations. The bottom-up module, after receiving the request, sends the answer template formed based on the unpopular question database according to the customer's question information to the client. The vulnerability scanning module regularly performs vulnerability scans on the scan target and gives a vulnerability detection report.

[0143] Those skilled in the art can understand that all or part of the steps in the above-mentioned embodiment method can be completed by instructing the relevant hardware through a program, so the present application can be in the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present application can be in the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.

[0144] Each embodiment in this specification is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the above embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0145] The above implementation methods have been described in detail. Specific examples are used herein to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. 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 a limitation on the present invention.

Claims

1. A risk compliance intelligent interaction method, characterized in that: The method comprises the following steps: S1. Build a trust domain retrieval library and unpopular question database based on historical data; S2. Obtaining several question information input by the client, and preprocessing the question information to obtain several question data C a ; S3, convert the question data C(a) into question representation f(a) through a convolutional neural network, and obtain a question representation f(b) similar to the question representation f(a) from the trust domain retrieval library; S4. Calculate the content similarity value sim based on f(a) and f(b) cos and the average similarity Determine whether there is a matching answer in the trust domain search library; like If they are not similar, a request is sent to the bottom end; like Similarly, the answer matching the question representation f(b) in the trust domain retrieval library is sent to the review end; S5. The audit end receives the answer and calculates the key value P i Obtain keywords, compare the keywords with the words in the multi-dimensional audit library established based on network information, and determine whether they are compliant words; If yes, a response is sent to the client and the process goes to step S7; If not, return to S2 until the wording in the answer complies with the regulations; S6. After receiving the request, the back-end sends the answer template formed based on the unpopular question database to the client according to the customer's question information; S7. Regularly perform vulnerability scans on the scan targets and provide vulnerability detection reports.

2. The intelligent interaction method according to claim 1, characterized in that: In step S1, the specific implementation steps are as follows: S11. Obtain the number Z of all questions asked by customers in historical data and the number C of repeated questions, and calculate the question repetition rate P c , the calculation formula is as follows: Among them, P c represents the repetition rate of questions, C represents the number of repeated questions, and Z represents the number of all questions asked by customers; S12, determining whether the problem is a common problem; If P c ≥0.5, it is marked as a common problem; If P c <0.5, it is marked as an unpopular question and a database of unpopular questions is constructed; S13. Build a trust domain search library based on frequently asked questions, government websites, official media and encyclopedia knowledge. The trust domain search library includes questions and matching answers.

3. The intelligent interaction method according to claim 1, characterized in that: In step S2, the specific implementation steps are as follows: S21. Use the NLPIR word segmenter to segment and annotate the obtained question information and obtain the annotated information S22. Use the extractive attention mechanism method to compress the redundant information in the annotation information to obtain the question data C a .

4. The intelligent interaction method according to claim 1, characterized in that: In step S3, the specific implementation steps are as follows: S31. Setting the analysis core D a , the analysis kernel D a and question data C a Perform convolution operation to obtain convolution data B a ; S32, determine the pooling kernel H c , the pooling kernel H c Convolution data B a Perform pooling operation to obtain pooling data G b ; S33, using sigmoid function as activation function, the pooled data G b Substitute the activation function for full connection calculation to obtain the question representation f(a), the calculation formula is as follows: f(a)=K(w b G b +d) Among them, f(a) represents the question representation, G b represents the bth pooled data, w b Represents the weight coefficient in the activation function, K represents the activation function, and d represents the bias coefficient; S34, obtaining similar question data from the trust domain retrieval library and converting it into a question representation f(b) using the method of steps S31 to S33.

5. The intelligent interaction method according to claim 1, characterized in that: In step S4, the specific implementation steps are as follows: S41. Calculate the content similarity value sim between the question representation f(a) and the question representation f(b) c os; S42. Obtain the historical content similarity value sim in the past six months his , calculate the average similarity value by the average method The calculation formula is as follows: in, Represents the average similarity value, sim his Indicates the historical content similarity value, n represents sim his The number of 6. The intelligent interaction method according to claim 5, characterized in that: The content similarity value sim cos The calculation formula is as follows: Among them, sim cos represents the content similarity value, f(a) represents the question representation input by the customer, and f(b) represents the question representation obtained from the trust domain retrieval library.

7. The intelligent interaction method according to claim 1, characterized in that: In step S5, the specific implementation steps are as follows: S51. A multi-dimensional audit database is established based on Internet dictionaries, news sensitive words and illegal words. The audit end obtains the question data C corresponding to the dissimilar question representation f(a) a ; S52, obtaining a number of pre-processed question data C a The frequency e of word i in i , calculate the word frequency value fre of word i i , the calculation formula is as follows: Among them, fre i Indicates the word frequency value, e i Indicates the frequency of occurrence of word i; S53. Obtaining question data C a The frequency u of the sentence beginning of word i in i 、The frequency of the end of the sentence v i and the frequency in the sentence t i , calculate the question data C a The position frequency of word i in add i , the calculation formula is as follows: Among them, add i Indicates the position frequency of the word, u i represents the frequency of word i appearing at the beginning of the sentence, v i represents the frequency of word i appearing at the end of the sentence, t i Indicates the frequency of word i appearing in the sentence; S54, obtaining the total number of words tw of the Internet dictionary and the frequency O of the word i in the Internet dictionary i Sum frequency e i And calculate the frequency value select i , the calculation formula is as follows: Among them, select i represents the frequency value of word i, tw represents the total number of words in the Internet dictionary; S55, according to the word frequency value fre i 、Position frequencyadd i and frequency value select i Calculate the key value P of word i i , the calculation formula is as follows: P i =A×fre i +B×add i +C×select i Among them, P i represents the key value, A, B and C represent the weight coefficients respectively; S56, setting keyword threshold P based on adaptive threshold method g , determine whether word i is a keyword; If P i ≥P g , then it is a keyword and is compared with the words in the multi-dimensional audit library; If P i <P g , it is not a keyword and is ready to detect the next word.

8. The intelligent interaction method according to claim 1, characterized in that: In step S6, the specific implementation steps are as follows: S61. Establishing an answer template based on unpopular questions in the unpopular question database; S62: After receiving the request, the back-end finds similar questions and corresponding answer templates in the unpopular question database according to the question information; S63. Send the answer template to the client.

9. The intelligent interaction method according to claim 1, characterized in that: In step S7, the specific implementation steps are as follows: S71, constructing a timing mechanism in the startup program of the vulnerability scanning program through a time wheel algorithm; S72. Determine regular scanning targets in the vulnerability scanning program; S73. Give a vulnerability detection report after the scan is completed.

10. A system for implementing the risk compliance intelligent interaction method according to any one of claims 1 to 9, characterized in that: The system includes: Database module, which builds a trust domain search library and a database of unpopular questions based on the questions and answers of customers in the past three years; The preprocessing module obtains several question information input by the client and preprocesses the question information to obtain several question data C a ; The intelligent processing module converts the question data C(a) into question representation f(a) through a convolutional neural network, and obtains similar question representation f(b) from the trust domain retrieval library; Matching module, calculates the content similarity value sim based on f(a) and f(b) cos and the average similarity Determine whether there is a matching answer in the trust domain search library; like If they are not similar, a request is sent to the bottom end; like Similarly, the answer matching the question representation f(b) in the trust domain retrieval library is sent to the review end; In the compliance review module, the audit end receives the answer and calculates the key value P i Obtain keywords, compare the keywords with the words in the multi-dimensional audit library established based on network information, and determine whether they are compliant words; If yes, a reply is sent to the client; If not, return to S2 until the wording in the answer complies with the regulations; The backup module, after receiving the request, sends the answer template formed based on the database of unpopular questions to the client according to the customer's question information; The vulnerability scanning module regularly performs vulnerability scans on the scan targets and provides vulnerability detection reports.

Images