Natural resource data desensitization method and system based on NLP

Abstract

The embodiments of this application relate to the field of data processing technology, and specifically to a natural resource data desensitization method and system based on NLP, which accurately identifies and desensitizes private content by deeply mining the key semantics in the report, while maintaining the original intent and information integrity of the report. This technical solution is proposed to solve the efficiency and accuracy problems faced by traditional desensitization methods when processing natural resource survey reports, and to provide strong support for data security and sharing in the field of natural resource surveys.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and more specifically, to a natural resource data desensitization method and system based on NLP. Background Art

[0002] In the field of natural resource research, research reports are important documents that bring together a large number of field investigations and data collection results. However, these reports often contain sensitive information and private data, such as specific geographic locations, rare species data, etc. If they are directly disclosed or shared, it is difficult to ensure that the data is effectively protected, so desensitization is required.

[0003] Traditional desensitization methods mostly rely on manual modification or the use of simple replacement rules. This method is not only inefficient, but also difficult to ensure the thoroughness of desensitization and the consistency of data. At the same time, traditional methods have obvious deficiencies in processing complex semantics and maintaining the original meaning of reports, which often leads to serious information loss in desensitized reports and reduced readability. Summary of the invention

[0004] In view of this, the present application provides a natural resource data desensitization method and system based on NLP.

[0005] The embodiment of the present application provides a natural resource data desensitization method based on NLP, which is applied to an NLP data desensitization system. The method includes:

[0006] Using the basic NLP data desensitization network, perform key embedded semantic mining on the natural resources research report sample to obtain the key embedded semantic vector, and use the basic NLP data desensitization network to perform privacy semantic recognition on the key embedded semantic vector to obtain the privacy semantic representation vector;

[0007] According to the first attention coefficient of the key embedded semantic vector and the second attention coefficient of the privacy semantic representation vector, a semantic enhancement operation is performed on the key embedded semantic vector and the privacy semantic representation vector to obtain enhanced semantics of the research report;

[0008] Using the basic NLP data desensitization network, perform NLP text data paraphrase operation on the key embedded semantic vector, the privacy semantic representation vector and the research report enhanced semantics to obtain a natural resource research paraphrase report sample of the natural resource research report sample;

[0009] Perform network debugging quality analysis based on the difference between the natural resources survey retelling report sample and the natural resources survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network;

[0010] According to the error variable of the debugging error, the neuron weights of the basic NLP data desensitization network are optimized to obtain the debugged NLP data desensitization network.

[0011] In some technical solutions, the basic NLP data desensitization network includes a semantic feature mining branch and a semantic feature recognition branch, the semantic feature mining branch includes an X-order connected semantic feature mining subnet, the semantic feature recognition branch includes an X-order connected semantic feature recognition subnet, the semantic feature mining subnet and the semantic feature recognition subnet correspond one to one, and X is an integer greater than 0; the basic NLP data desensitization network is used to perform key embedded semantic mining on a natural resources survey report sample to obtain a key embedded semantic vector, including:

[0012] Using the first-order semantic feature mining subnet of the X-order semantic feature mining subnet, key embedding semantic mining is performed on the natural resources investigation report sample to obtain a key embedding semantic vector generated by the first-order semantic feature mining subnet;

[0013] Using the u-th order semantic feature mining subnet of the x-th order semantic feature mining subnet, performing key embedding semantic mining on the key embedding semantic vector generated at the u-1th order, to obtain the key embedding semantic vector generated by the u-th order semantic feature mining subnet;

[0014] Performing a self-incremental 1 walk process on u to obtain a key embedded semantic vector generated by each order semantic feature mining subnet in the X-order semantic feature mining subnet, wherein u is an integer greater than 0 and not greater than X; the self-incremental 1 walk process on u is to increase the value of u by 1 until u=X;

[0015] The method of using the basic NLP data desensitization network to perform privacy semantics recognition on the key embedded semantic vector to obtain a privacy semantic representation vector includes:

[0016] Using the Xth-order semantic feature recognition subnet of the Xth-order semantic feature recognition subnet, performing privacy semantic recognition on the key embedded semantic vector generated by the Xth-order semantic feature mining subnet, and obtaining the privacy semantic representation vector generated by the Xth-order semantic feature recognition subnet;

[0017] Using the u-th order semantic feature recognition subnet of the x-th order semantic feature recognition subnet, perform privacy semantic recognition on the privacy semantic representation vector generated at the u+1th order, to obtain the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet;

[0018] A self-incremental 1 walk process is performed on u to obtain a privacy semantic representation vector generated by each order semantic feature recognition subnet in the X-order semantic feature recognition subnet.

[0019] In some technical solutions, the key embedded semantic vector generated by each order semantic feature mining subnet in the X-order semantic feature mining subnet has a corresponding first attention coefficient, and the privacy semantic representation vector generated by each order semantic feature recognition subnet in the X-order semantic feature recognition subnet has a corresponding second attention coefficient; the semantic enhancement operation is performed on the key embedded semantic vector and the privacy semantic representation vector based on the first attention coefficient of the key embedded semantic vector and the second attention coefficient of the privacy semantic representation vector to obtain enhanced semantics of the research report, including:

[0020] Obtaining a first attention coefficient of a key embedded semantic vector generated by a u-th order semantic feature mining subnet in the X-th order semantic feature mining subnet, and obtaining a second attention coefficient of a privacy semantic representation vector generated by a u-th order semantic feature recognition subnet in the X-th order semantic feature recognition subnet;

[0021] According to the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet and the second attention coefficient of the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet, a semantic enhancement operation is performed on the key embedded semantic vector generated by the u-th order semantic feature mining subnet and the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet to obtain the u-th order research report enhanced semantics;

[0022] Perform a self-adding 1 walk process on u to obtain the enhanced semantics of the research report of order X.

[0023] In some technical solutions, the obtaining of the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet in the x-th order semantic feature mining subnet includes: obtaining the original attention coefficient and the attention trend adjustment feature of the key embedded semantic vector generated by the u-th order semantic feature mining subnet, and determining the current debugging cycle of the basic NLP data desensitization network; determining the attention trend adjustment feature and the feature operation result of the current debugging cycle, and combining the original attention coefficient and the feature operation result to obtain the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet, the original attention coefficient is 1, and the first attention coefficient is greater than or equal to 0 and less than or equal to 1;

[0024] The obtaining of the second attention coefficient of the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet in the x-th order semantic feature recognition subnet includes: combining a preset value with the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet to determine the second attention coefficient of the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet.

[0025] In some technical solutions, the attention trend adjustment features of the key embedded semantic vectors generated by each order semantic feature mining subnet in the X-order semantic feature mining subnet are different, and the attention trend adjustment features of the key embedded semantic vectors have a set quantitative relationship with the order of the semantic feature mining subnet that outputs the key embedded semantic vectors.

[0026] In some technical solutions, the NLP text data paraphrase operation is performed on the key embedded semantic vector, the privacy semantic representation vector, and the research report enhanced semantics to obtain a natural resource research paraphrase report sample of the natural resource research report sample, including:

[0027] Integrate the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet in the X-th order semantic feature recognition subnet with the u-th order research report enhanced semantics in the X-th order research report enhanced semantics to obtain the u-th order text semantic integrated vector; perform a self-incremental 1 walk process on u to obtain the X-th order text semantic integrated vector;

[0028] Integrate the key embedded semantic vector generated by the X-th order semantic feature mining subnet in the X-order semantic feature mining subnet and the X-order text semantic integration vector to obtain the natural resources investigation report retelling semantics;

[0029] A text paraphrase operation is performed on the paraphrase semantics of the natural resource investigation report to obtain a natural resource investigation paraphrase report sample of the natural resource investigation report sample.

[0030] In some technical solutions, before performing network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network, the method further includes:

[0031] Performing a privacy decision analysis on the natural resource survey report sample through the text privacy decision branch to obtain a first privacy decision analysis viewpoint, and performing a privacy decision analysis on the natural resource survey retelling report sample through the text privacy decision branch to obtain a second privacy decision analysis viewpoint;

[0032] The first privacy decision analysis viewpoint is used to characterize the probability value of the natural resource survey report sample being generated by the basic NLP data desensitization network, and the second privacy decision analysis viewpoint is used to characterize the probability value of the natural resource survey retelling report sample being generated by the basic NLP data desensitization network;

[0033] The network debugging quality analysis is performed based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network, including:

[0034] Performing a natural resource survey report network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample, and obtaining a first error variable of the natural resource survey report debugging error of the basic NLP data desensitization network;

[0035] Performing a privacy decision analysis network debugging quality analysis based on the first privacy decision analysis viewpoint and the second privacy decision analysis viewpoint to obtain a first decision offset weight of a first decision analysis debugging error of the basic NLP data desensitization network;

[0036] The sum of the first error variable and the first decision offset weight is used as the error variable of the debugging error of the basic NLP data desensitization network.

[0037] In some technical solutions, before performing network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network, the method further includes:

[0038] Extracting the semantic vector of the comments on the sample of the natural resources investigation report to obtain the semantic vector of the comments on the sample of the natural resources investigation report;

[0039] The basic NLP data desensitization network is used to perform semantic paraphrase optimization on the semantic vector of the comment opinion, thereby obtaining the semantic paraphrase vector of the natural resources research report sample.

[0040] In some independent technical solutions, the network debugging quality analysis is performed based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network, including:

[0041] Performing a natural resource survey report network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample, and obtaining a first error variable of the natural resource survey report debugging error of the basic NLP data desensitization network;

[0042] Performing feature network debugging quality analysis based on the difference between the semantic paraphrase vector and the comment view semantic vector to obtain a second error variable of the semantic debugging error of the basic NLP data desensitization network;

[0043] The sum of the first error variable and the second error variable is used as the error variable of the debugging error of the basic NLP data desensitization network.

[0044] In some independent technical solutions, the method further includes:

[0045] A privacy decision analysis is performed on the natural resource survey report sample through a text privacy decision branch to obtain a first privacy decision analysis viewpoint, and a privacy decision analysis is performed on the natural resource survey retelling report sample through the text privacy decision branch to obtain a second privacy decision analysis viewpoint; wherein the first privacy decision analysis viewpoint is used to characterize that the natural resource survey report sample is a probability value generated by the basic NLP data desensitization network, and the second privacy decision analysis viewpoint is used to characterize that the natural resource survey retelling report sample is a probability value generated by the basic NLP data desensitization network;

[0046] A privacy decision analysis is performed on the comment opinion semantic vector through the text privacy decision branch to obtain a third privacy decision analysis opinion, and a privacy decision analysis is performed on the semantic paraphrase vector through the text privacy decision branch to obtain a fourth privacy decision analysis opinion; wherein the third privacy decision analysis opinion is used to characterize that the comment opinion semantic vector is a probability value generated by the basic NLP data desensitization network, and the fourth privacy decision analysis opinion is used to characterize that the semantic paraphrase vector is a probability value generated by the basic NLP data desensitization network;

[0047] The network debugging quality analysis is performed based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network, including:

[0048] Performing a natural resource survey report network debugging quality analysis based on the difference between the natural resource survey restatement report sample and the natural resource survey report sample to obtain a first error variable of the natural resource survey report debugging error of the basic NLP data desensitization network, and performing a feature network debugging quality analysis based on the difference between the semantic restatement vector and the comment view semantic vector to obtain a second error variable of the semantic debugging error of the basic NLP data desensitization network;

[0049] Performing a privacy decision analysis network debugging quality analysis based on the first privacy decision analysis viewpoint and the second privacy decision analysis viewpoint to obtain a first decision offset weight of a first decision analysis debugging error of the basic NLP data desensitizing network, and performing a privacy decision analysis network debugging quality analysis based on the third privacy decision analysis viewpoint and the fourth privacy decision analysis viewpoint to obtain a second decision offset weight of a second decision analysis debugging error of the basic NLP data desensitizing network;

[0050] The sum of the first error variable, the second error variable, the first decision offset weight and the second decision offset weight is used as the error variable of the debugging error of the basic NLP data desensitization network.

[0051] The present application also provides an NLP data desensitization system, comprising: a memory for storing program instructions and data; a processor for coupling with the memory and executing instructions in the memory to implement the above method.

[0052] The present application also provides a computer storage medium, comprising instructions, which implement the above method when executed on a processor.

[0053] The embodiment of this application deeply mines the key semantics in the report, accurately identifies and desensitizes the private content, while maintaining the original intention and information integrity of the report. This technical solution is proposed to solve the efficiency and accuracy problems faced by traditional desensitization methods when processing natural resource survey reports, and provide strong support for data security and sharing in the field of natural resource surveys. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] The accompanying drawings herein are incorporated in and constitute a part of the specification, illustrate embodiments consistent with the present application, and together with the description, are used to explain the principles of the present application.

[0055] Figure 1 A schematic flowchart of the steps of a natural resource data desensitization method based on NLP provided in an embodiment of the present application.

[0056] Figure 2 A structural block diagram of an NLP data desensitization system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0057] The technical solution in this application will be described below in conjunction with the accompanying drawings.

[0058] When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. Instead, they are merely examples of devices and methods consistent with some aspects of the present application. It should be noted that the terms "first", "second", etc. in the specification of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0059] The method embodiments provided in the embodiments of the present application can be executed in an NLP data desensitization system, a computer device or a similar computing device. Taking the operation on the NLP data desensitization system as an example, the NLP data desensitization system may include one or more processors (the processor may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory for storing data. Optionally, the above-mentioned NLP data desensitization system may also include a transmission device for communication functions. It can be understood by those of ordinary skill in the art that the above-mentioned structure is only illustrative and does not limit the structure of the above-mentioned NLP data desensitization system. For example, the NLP data desensitization system may also include more or fewer components than those shown above, or have a configuration different from that shown above.

[0060] The memory can be used to store computer programs, for example, software programs and modules of application software, such as a computer program corresponding to a natural resource data desensitization method based on NLP in an embodiment of the present application. The processor executes various functional applications and data processing by running the computer program stored in the memory, that is, implementing the above method. The memory may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the NLP data desensitization system via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0061] The transmission device is used to receive or send data via a network. Specific examples of the above-mentioned network may include a wireless network provided by a communication provider of the NLP data desensitization system. In one example, the transmission device includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device can be a radio frequency (Radio Frequency, referred to as RF) module, which is used to communicate with the Internet wirelessly.

[0062] In conjunction with the above, please refer to Figure 1 , Figure 1 It is a flow chart of a natural resource data desensitization method based on NLP provided in an embodiment of the present application. The method is applied to an NLP data desensitization system and may further include steps 110 to 150.

[0063] Step 110: Use the basic NLP data desensitizing network to perform key embedded semantic mining on the natural resources survey report sample to obtain the key embedded semantic vector, and use the basic NLP data desensitizing network to perform privacy semantic identification on the key embedded semantic vector to obtain the privacy semantic representation vector.

[0064] Step 120: Based on the first attention coefficient of the key embedded semantic vector and the second attention coefficient of the privacy semantic representation vector, a semantic enhancement operation is performed on the key embedded semantic vector and the privacy semantic representation vector to obtain enhanced semantics of the research report.

[0065] Step 130: Utilize the basic NLP data desensitizing network to perform NLP text data paraphrase operation on the key embedded semantic vector, the privacy semantic representation vector, and the enhanced semantics of the survey report to obtain a natural resource survey paraphrase report sample of the natural resource survey report sample.

[0066] Step 140: Perform network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network.

[0067] Step 150: Optimize the neuron weights of the basic NLP data desensitization network according to the error variable of the debugging error to obtain the debugged NLP data desensitization network.

[0068] In the field of natural resource management (such as marine natural resources), research reports often contain a lot of sensitive information, such as geographic location, species distribution, etc. These non-desensitized information may have adverse effects on the ecological environment. Therefore, the application of NLP data desensitization system in this field is particularly important.

[0069] First, the NLP data desensitization system uses the basic NLP data desensitization network to conduct an in-depth analysis of a sample natural resource survey report. Through key embedded semantic mining technology, the system can accurately identify the key information in the report and convert this information into key embedded semantic vectors. Subsequently, the system performs privacy semantic recognition on these key embedded semantic vectors and further extracts privacy semantic representation vectors.

[0070] Next, the NLP data desensitization system performs semantic enhancement operations on these two types of vectors based on the first attention coefficient of the key embedded semantic vector and the second attention coefficient of the privacy semantic representation vector. This process ensures that key information and privacy information receive sufficient attention in subsequent processing, thereby generating more accurate desensitization results. The enhanced semantics are called research report enhanced semantics.

[0071] Then, the NLP data desensitization system uses the basic NLP data desensitization network to perform NLP text data retelling operations on key embedded semantic vectors, privacy semantic representation vectors, and research report enhanced semantics. The purpose of this step is to desensitize the original report while retaining key information and generate a new natural resource research retelling report sample. This retelling report not only retains the core content of the original report, but also effectively removes sensitive information.

[0072] In order to verify the desensitization effect, the NLP data desensitization system compares the original natural resource survey report sample and the desensitized retelling report sample and analyzes the difference between the two. Through this comparison, the system can calculate the debugging error generated by the basic NLP data desensitization network during the processing process, which is expressed in the form of an error variable.

[0073] Finally, based on the error variable of this debugging error, the NLP data desensitization system optimizes the neuron weights of the basic NLP data desensitization network. This process is automatically completed through machine learning algorithms, aiming to improve the accuracy and efficiency of desensitization processing. The optimized NLP data desensitization network will be able to more accurately process sensitive information in natural resource survey reports and provide more secure and reliable data support for natural resource management.

[0074] It can be seen that the application of NLP data desensitizing system in the desensitizing processing of natural resource survey reports not only improves the security of data processing, but also provides strong technical support for data analysis in the field of natural resource management.

[0075] The above steps 110 to 150 are further described in detail below in conjunction with specific marine natural resources application scenarios.

[0076] In step 110, the basic NLP data desensitization network is a network model built using natural language processing (NLP) technology, which is specially designed to identify and desensitize sensitive information in research reports. This network learns how to accurately identify key information and private data in texts through a large amount of training data, and can automatically desensitize these data during processing to protect the confidentiality and security of information.

[0077] In the context of marine natural resources, a sample natural resource survey report can refer to a report that records in detail key information such as the distribution of marine resources, the current state of the ecological environment, species diversity, etc. This report is written by researchers after field investigations and data collection, and contains a large amount of field observation data, analysis results, and possible sensitive information, such as the habitats and breeding areas of specific marine organisms.

[0078] Key embedding semantic mining is an important step in the NLP data desensitization system. It refers to the use of NLP technology to deeply analyze the report text and extract semantic information that is crucial to the understanding and desensitization process. In the marine natural resources scenario, this can include identifying key words and phrases related to marine ecology, resource distribution, etc., as well as their associations and contextual meanings.

[0079] Key semantic embedding vectors are an expression that converts key semantic information in text into numerical form through NLP technology. In the context of marine natural resources, these vectors may represent key concepts such as specific sea areas, species names, and ecological and environmental characteristics mentioned in the report. Each vector is a point in a high-dimensional space, and its numerical features reflect the semantic content and contextual relationship of the corresponding word or phrase. These vectors not only help understand the meaning of the text, but also serve as the basis for subsequent desensitization processing.

[0080] Privacy semantic recognition is one of the core functions of the NLP data desensitization system, which focuses on identifying and marking sensitive or private information in text. In a survey report on marine natural resources, this may involve identifying specific geographical locations, detailed information on rare species, and vulnerable areas of the marine ecological environment. If this information is not properly desensitized, it may have an adverse impact on the marine ecology.

[0081] The privacy semantic representation vector is the output of the privacy semantic recognition process, which converts the identified sensitive information into a numerical vector form. Similar to the key embedded semantic vector, the privacy semantic representation vector is also a point in a high-dimensional space, but its features focus more on reflecting the privacy or sensitive content in the text. The numerical features of these vectors are carefully designed to ensure that these sensitive information can be effectively shielded or replaced in the subsequent desensitization process, thereby protecting the privacy of the data.

[0082] In detail, the NLP data desensitization system first loads a sample survey report on marine natural resources based on step 110. This report may contain detailed information on marine ecology, species distribution, marine resources, etc. The system then uses the basic NLP data desensitization network to conduct an in-depth analysis of the report. First, the system performs key embedding semantic mining, which is an automated process that extracts key information from the report through NLP technology, such as specific marine areas, important species names, key features of the ecological environment, etc. This information is essential for understanding the overall content of the report and subsequent desensitization processing. After the key embedding semantic mining is completed, the system converts the extracted key information into key embedding semantic vectors. These vectors are numerical representations in high-dimensional space that can capture the deep semantic structure and associations in the text. For example, one vector may represent "coral reef area" while another vector may represent "target fish". Next, the system performs privacy semantic recognition, focusing on identifying and marking sensitive or private information in the text. In the survey report on marine resources, this can include specific latitude and longitude coordinates, detailed descriptions of rare marine organisms, specific locations of marine protected areas, etc. If this information is improperly disclosed, it may pose a potential threat to the marine ecological environment. After the privacy semantics identification is completed, the system will also convert the sensitive information into privacy semantic representation vectors. Similar to the key embedding semantic vectors, these vectors are also numerical representations, but they focus more on reflecting the privacy content in the text. For example, a privacy semantic representation vector may represent the location information of a specific marine protected area. Through this step of processing, the NLP data desensitization system successfully converts the original text report into a series of numerical key embedding semantic vectors and privacy semantic representation vectors, laying a solid foundation for subsequent data desensitization work.

[0083] In step 120, the first attention coefficient is a numerical indicator used to measure the importance of the key embedded semantic vector in the overall semantics. This coefficient is calculated by the deep learning algorithm of the NLP data desensitization network, which reflects the importance of each key embedded semantic vector for understanding and desensitizing the entire research report. Simply put, the first attention coefficient is the "weight" or "attention" of the key information in the report. Similar to the first attention coefficient, the second attention coefficient is an importance indicator for the privacy semantic representation vector. In the context of marine natural resources, it indicates the criticality of privacy or sensitive information in the overall report. This coefficient helps the NLP data desensitization system determine which privacy information should be given more attention and protection during the processing process to ensure the security and confidentiality of this information.

[0084] The semantic enhancement operation is a key link in the NLP data desensitization system. It uses the first attention coefficient and the second attention coefficient to perform weighted processing on the key embedded semantic vector and the privacy semantic representation vector. The purpose of this operation is to enhance the expression of important information and privacy information in the report, so that these key contents can be more accurately and effectively protected in the subsequent desensitization process. Through semantic enhancement, the system can more accurately identify and process sensitive data in the text.

[0085] The enhanced semantics of the research report is the output result obtained after the semantic enhancement operation. In the application scenario of marine natural resources, it represents a collection of weighted key information and privacy information, which will be the focus of subsequent NLP text data retelling operations. The enhanced semantics not only retains the core content of the original research report, but also provides clearer and more precise guidance for subsequent desensitization work.

[0086] In detail, the NLP data desensitization system performs semantic enhancement operations on the key embedded semantic vector and the privacy semantic representation vector based on the first attention coefficient and the second attention coefficient calculated previously based on step 120. For example, the system first reviews the key embedded semantic vector and the privacy semantic representation vector generated in step 110, as well as their corresponding attention coefficients. These coefficients are essentially the importance indicators of each vector in the overall semantics, which are automatically calculated by the NLP data desensitization network through a deep learning algorithm. Next, the system performs a semantic enhancement operation. The core of this operation is to weight the vector according to the attention coefficient to enhance the expression of important information and privacy information. For the key embedded semantic vector, the system uses the first attention coefficient for weighting to ensure that the core content in the report receives sufficient attention in the subsequent desensitization process. Similarly, for the privacy semantic representation vector, the system uses the second attention coefficient for weighting to strengthen the protection of sensitive information. Through the weighted processing of this step, the original key embedded semantic vector and the privacy semantic representation vector are converted into enhanced semantics for the research report. This enhanced semantics not only highlights the key information and privacy information in the report, but also provides more clear and precise guidance for the subsequent NLP text data retelling operation. In short, step 120 is a key link in the NLP data desensitization process. It ensures the effective protection and accurate processing of important information and privacy information during the desensitization process through semantic enhancement operations.

[0087] In step 130, the NLP text data retelling operation is an advanced technology in the field of natural language processing (NLP), which aims to retell the content of the original text by reorganizing, converting or generating text, and may change the text's expression, sentence structure or vocabulary. In the application scenario of marine natural resources, this operation is particularly used for privacy anonymity protection and desensitization processing, that is, in the retelling process, special attention will be paid to and modified for those parts that may disclose sensitive information, such as the precise location of a specific marine area, the specific name of a rare marine species, etc., to ensure that the retelling text retains the main information and intention of the original research report, and effectively removes details that may cause privacy or security issues.

[0088] The sample natural resource survey paraphrase report is a new report text generated after the NLP text data paraphrase operation. This report is similar to the original natural resource survey report in content, but has been desensitized, especially in terms of key information and sensitive data related to marine natural resources, which have been anonymized or blurred. For example, the name of a specific marine protected area, longitude and latitude coordinates, or detailed data of a specific species that may be mentioned in the original report may be replaced with a more general description or a vague data range in the paraphrase report to achieve the purpose of privacy anonymity protection. Such a paraphrase report sample can provide valuable natural resource survey data and analysis results while protecting sensitive information.

[0089] In detail, the NLP data desensitization system will perform the NLP text data retelling operation based on step 130, which is a key link in the entire desensitization process. The system will first call the basic NLP data desensitization network, which is a trained deep learning model with powerful text processing and analysis capabilities. Next, the system will input three key components: key embedding semantic vectors, privacy semantic representation vectors, and research report enhanced semantics. These components are obtained by in-depth analysis and processing of the original natural resource research report samples in the previous steps, and together they constitute a comprehensive and accurate expression of the original report content. Then, the NLP data desensitization network will use these components to perform a retelling operation. For example, the network will refer to the key embedding semantic vector to retain the core content and structure of the report, and at the same time desensitize and anonymize sensitive information based on the privacy semantic representation vector and the research report enhanced semantics. In this process, the network will intelligently replace or blur specific geographical locations, species names and other sensitive information to ensure that the desensitized report meets the requirements of privacy protection without affecting the transmission and understanding of the overall information. Finally, through the NLP text data retelling operation, the system will generate a new natural resource survey retelling report sample. This report is highly relevant to the original report in terms of content, but all sensitive information has been effectively desensitized and anonymized, so that it can provide safe and reliable data support for natural resource management, scientific research analysis or public information release while protecting privacy.

[0090] In step 130, network debugging quality analysis is a process of evaluating the performance of the NLP data desensitization network. In the application scenario of marine natural resources, it specifically refers to the in-depth study and evaluation of the differences between the retelling report and the original report after the network processing. This process aims to discover possible problems or deficiencies in the network when processing marine natural resource data, and then optimize the performance and accuracy of the network. By comparing the content differences between the retelling report and the original report, the degree of semantic retention, the desensitization effect and other aspects, the quality of the network can be comprehensively and carefully analyzed.

[0091] Debugging error refers to the difference or deviation between the repetition report output by the network and the expected output (i.e., the ideal desensitization effect) during the debugging and optimization process of the NLP data desensitization network. In the application scenario of marine natural resources, this error may manifest as incomplete desensitization, excessive information loss, and excessive semantic changes. Debugging error is an important indicator for measuring network performance. By reducing debugging error, the desensitization effect and accuracy of the network can be improved.

[0092] Error variables are specific indicators or parameters used to quantify and describe debugging errors. In the application scenario of marine natural resources, error variables can include information loss rate, semantic retention, desensitization accuracy, and other aspects. These variables can objectively reflect the performance of the NLP data desensitization network when processing marine natural resource data, and provide strong data support for further optimization of the network. By analyzing these error variables, we can find out the problems in the network and make targeted improvements and adjustments.

[0093] In detail, based on step 140, a network debugging quality analysis will be performed, and the debugging error and its error variables of the NLP data desensitization network will be calculated. First, two reports need to be prepared: one is a sample of the natural resource survey retelling report processed by the NLP data desensitization network, and the other is a sample of the original natural resource survey report. By carefully comparing the two reports, the specific differences between the report processed by the network and the original report can be observed. Next, a network debugging quality analysis will be performed. This process mainly includes the following aspects: checking whether the desensitization effect is thorough, that is, whether the sensitive information in the original report is effectively removed or replaced; evaluating the degree of semantic retention, that is, whether the retelling report accurately conveys the main content and intention of the original report; and analyzing the information loss, that is, whether important information is mistakenly deleted or omitted during the network processing process. In order to quantify these analysis results, a series of error variables of the debugging error will be calculated. For example, the information loss rate can be used to measure the completeness of information during network processing; the semantic retention can reflect the similarity between the retelling report and the original report in content; and the desensitization accuracy rate can evaluate the network's processing effect on sensitive information. Finally, these error variables are summarized and analyzed to more comprehensively understand the performance of the NLP data desensitization network in the marine natural resources application scenario. Through these data, problems and deficiencies in the network can be discovered, providing a strong basis and guidance for subsequent optimization work.

[0094] In step 150, the connections between neurons are represented by weights, which determine the degree of influence of the input signal on the neuron output. In the application scenario of marine natural resources, neuron weights specifically refer to the connection parameters between neurons in the NLP data desensitization network. These weights are constantly adjusted during the network training process so that the network can better learn and adapt to the characteristics of marine natural resource data, thereby improving the accuracy and efficiency of desensitization processing. In short, neuron weights are key parameters for transmitting and processing information in the network, and their optimization is crucial to improving network performance.

[0095] The debugged NLP data desensitization network refers to a neural network model that has been improved for the marine natural resource data desensitization task after a series of debugging and optimization processes. This network improves its ability to process marine natural resource data, especially in terms of privacy protection and desensitization, by adjusting its internal neuron weights and other parameters. The debugged network can more accurately identify and remove sensitive information while retaining the key content of the original data, thereby providing safer and more efficient technical support for the investigation, analysis and sharing of marine natural resources.

[0096] In detail, step 150 is the key link of the entire NLP data desensitization network optimization process. The main purpose is to optimize the neuron weights of the basic NLP data desensitization network according to the error variables of the debugging error calculated in the previous step, and then obtain the debugged NLP data desensitization network. First, the system carefully analyzes the error variables of the debugging error, which reflect the shortcomings of the network in processing marine natural resource data, such as incomplete desensitization and excessive information loss. Through in-depth research on these error variables, it can be determined which neuron weights need to be adjusted and the direction and amplitude of the adjustment. Next, the system uses advanced optimization algorithms, such as gradient descent method, back propagation algorithm, etc., to fine-tune the weights of neurons. This process is iterative, that is, the performance of the network will be re-evaluated after each adjustment until a satisfactory desensitization effect and information retention degree are achieved. During the optimization process, the system will also pay special attention to maintaining the stability and generalization ability of the network to avoid overfitting or underfitting. Through continuous adjustment and optimization, a debugged NLP data desensitization network with better performance and more adapted to the characteristics of marine natural resource data can be obtained. This debugged network will have a higher desensitization accuracy and better information retention capabilities, providing more reliable technical support for subsequent processing and sharing of marine natural resource data.

[0097] In the embodiment of the present application, in the application scenario of marine natural resources, the key embedded semantic vector, the privacy semantic representation vector and the research report enhanced semantics are the core invention points based on the feature vector, each of which carries different information and is specifically expressed in the form of a numerical feature vector. The following is a detailed description and example of these three vectors.

[0098] (1) Key semantic embedding vector

[0099] The key semantic embedding vector is a key feature vector that captures the main content and structure of the research report. It converts the keywords or phrases in the text into vector representations in a high-dimensional space through deep learning models (such as Word2Vec, BERT, etc.), thereby retaining the semantic information of the text. In the application scenario of marine natural resources, these keywords can include marine species names, geographical locations, ecological environment descriptions, etc.

[0100] For example, there is a research report on "coral reef ecosystem", in which the keywords include "coral reef", "fish diversity", "water temperature", "salinity", etc. Through model training, these keywords will be embedded in a high-dimensional space to form key embedded semantic vectors. For example, "coral reef" can be represented as a vector similar to [0.23, -0.45, 0.67, ...], while "water temperature" can be represented as [0.12, 0.78, -0.34, ...]. The similarity or distance between these vectors can reflect the semantic relationship between words.

[0101] (2) Privacy Semantic Representation Vector

[0102] The privacy semantic representation vector is a feature vector used to identify and represent sensitive information in text. In the application scenario of marine natural resources, this sensitive information may include specific geographic location coordinates, detailed data of rare species, etc. Through specific model training, the privacy semantic representation vector can accurately capture this information that needs to be desensitized.

[0103] In a research report, if the coordinates of a specific marine protected area are mentioned as "23°12′N, 113°28′E", this coordinate information may be highlighted in the privacy semantic representation vector, such as the vector [0.89, -0.23, 0.45, ...]. This vector will be clearly different from the vectors of other non-sensitive information in the report, making it easier to perform desensitization processing later.

[0104] (3) Research reports strengthen semantics

[0105] Research report enhanced semantics is the overall semantic feature vector of the research report extracted through deep learning technology. It does not only focus on a single word or phrase, but captures its core meaning and theme from the perspective of the entire report. This vector helps to keep the overall content and intent of the report intact during the desensitization process.

[0106] Continuing with the research report on "coral reef ecosystems" as an example, the enhanced semantics of the research report will be a vector that can represent the theme and content of the entire report. For example, it may be a high-dimensional vector similar to [0.56, -0.34, 0.78, ...]. When performing the desensitization operation, the system refers to this vector to ensure that the desensitized text remains consistent with the original report in terms of theme and content.

[0107] To facilitate understanding of the embodiments of the present application, a natural resource survey report and its corresponding natural resource survey retelling report are provided below. The natural resource survey retelling report is a report obtained after the natural resource survey report is processed by the debugged NLP data desensitization network, and has completed privacy anonymization and desensitization processing.

[0108] 1. Natural Resources Research Report

[0109] This report aims to investigate the marine natural resources in the XXX sea area. After field investigation and data collection, it was found that the sea area is rich in marine biological resources, including a variety of rare fish and coral reef ecosystems. The specific survey results are as follows:

[0110] Marine biodiversity: The sea area is rich in biodiversity, with more than 150 fish species recorded, including many fish with high economic value, such as tuna and mackerel;

[0111] Coral reef ecosystem: Coral reefs are widely distributed and in good condition, forming a unique marine ecosystem. The corals are diverse and colorful, providing a good habitat for marine life;

[0112] Water quality and environment: During the survey, seawater samples were collected and analyzed. The results showed that the water quality in this sea area is clear and the environmental quality is high, which is conducive to the growth and reproduction of marine life;

[0113] Geographical location and climate: The sea area is located in the southeast of XXX, affected by the monsoon, with a suitable climate, providing a good living environment for marine life;

[0114] To sum up, the sea area is rich in marine natural resources. It is recommended to strengthen ecological monitoring and protection of the sea area to ensure the sustainable use of marine resources.

[0115] II. Natural Resources Research Report

[0116] This report provides a comprehensive survey of marine resources in a specific sea area. The results show that the area has significant biodiversity, including a variety of marine life, especially a variety of fish resources. At the same time, the coral ecosystem in the area is in good condition, which has built an ideal habitat for many marine life. Water quality analysis shows that the environment in the sea area is clean and extremely beneficial for the reproduction and survival of marine life. In addition, the geographical location and climatic conditions of the survey area are suitable for the survival of organisms. Overall, the natural resources in this sea area are very rich and deserve further attention and rational use. In order to protect these valuable resources, it is recommended to implement long-term ecological monitoring measures.

[0117] In the comparison between the natural resource survey report and the natural resource survey retelling report given above, we can see that the retelling report successfully achieved privacy anonymity and desensitization. For example:

[0118] (1) Privacy and anonymity processing:

[0119] The original report explicitly mentioned "XXX sea area", but the re-report changed it to "a certain sea area", successfully removing the specific reference and achieving the privacy protection of the geographical location;

[0120] Specific fish names in the report, such as "tuna, mackerel", etc., were generalized in the re-report as "various marine organisms" and "various types of fish resources", without mentioning any specific species names, thereby protecting possible ecologically sensitive information or commercial secrets involved.

[0121] (2) Desensitization treatment:

[0122] The original report stated that "more than 150 fish species were recorded" but the re-report did not mention the specific number. Instead, it used "significant biodiversity" to avoid disclosing specific data.

[0123] Regarding the description of coral reefs, the original report's "corals are diverse and colorful" was annotated to "the coral ecosystem is in good condition", reducing the exposure of detailed information;

[0124] The description of the water quality environment changed from “clear water and high environmental quality” to “clean environment”, which also reduced the disclosure of specific environmental parameters;

[0125] The description of the geographical location and climate has also been desensitized from the specific "located in the southeast of XXX, affected by the monsoon" to the more vague "the geographical location and climatic conditions are suitable for the survival of organisms."

[0126] Through these processes, the natural resource survey re-report successfully removed or blurred details that may involve privacy or sensitive information while retaining the core content and meaning of the original report, achieving the purpose of privacy anonymity and desensitization. Such processing is of great significance for protecting the ecological information of the survey area and preventing the data from being abused or used for illegal purposes, and it also complies with standard practices for data protection and information security.

[0127] It can be seen that the embodiment of the present application, by utilizing the basic NLP data desensitization network, deeply mines the key semantic information in the natural resources survey report, not only accurately identifies and characterizes the privacy content in the report, but also further enhances the core significance of the survey report through innovative semantic enhancement operations. This technical solution brings many significant benefits. First, by mining the key embedded semantic vectors, the embodiment of the present application can accurately capture the main points and key points in the survey report, laying a solid foundation for subsequent data processing and privacy protection. Secondly, the introduction of the privacy semantic representation vector makes the identification and desensitization of sensitive information more accurate, effectively realizes data anonymity desensitization, and ensures the security and compliance of the data. In addition, the embodiment of the present application cleverly combines the first attention coefficient and the second attention coefficient to semantically enhance the key embedded semantic vector and the privacy semantic representation vector, thereby generating a more concise and information-rich survey report enhanced semantics. This not only improves the readability and information transmission efficiency of the report, but also enables the desensitized report to more accurately reflect the core intent of the original report. Finally, through the NLP text data retelling operation, the embodiment of the present application generated a sample of a natural resource survey retelling report, which not only retained the important information of the original report, but also achieved effective desensitization of the privacy content. The network debugging quality analysis based on the difference between the retelling report and the original report further optimized the performance of the NLP data desensitization network and improved the accuracy and efficiency of the desensitization process. In summary, the technical solution of the embodiment of the present application fully retains the key information of the survey report while protecting privacy, achieves high efficiency and security of data processing, and is of great significance to data management and application in the field of natural resource surveys.

[0128] In some optional embodiments, the basic NLP data desensitization network includes a semantic feature mining branch and a semantic feature recognition branch, the semantic feature mining branch includes an X-order connected semantic feature mining subnet, the semantic feature recognition branch includes an X-order connected semantic feature recognition subnet, the semantic feature mining subnet and the semantic feature recognition subnet correspond one to one, and X is an integer greater than 0.

[0129] The method of utilizing the basic NLP data desensitizing network to perform key embedding semantic mining on the natural resources survey report sample to obtain the key embedding semantic vector includes: utilizing the 1st-order semantic feature mining subnet of the X-order semantic feature mining subnet to perform key embedding semantic mining on the natural resources survey report sample to obtain the key embedding semantic vector generated by the 1st-order semantic feature mining subnet; utilizing the u-th-order semantic feature mining subnet of the X-order semantic feature mining subnet to perform key embedding semantic mining on the key embedding semantic vector generated by the u-1th-order semantic feature mining subnet to obtain the key embedding semantic vector generated by the u-th-order semantic feature mining subnet; performing a self-incrementing 1 walk process on u to obtain the key embedding semantic vector generated by each-order semantic feature mining subnet in the X-order semantic feature mining subnet, wherein u is an integer greater than 0 and not greater than X; the self-incrementing 1 walk process on u is to increase the value of u by 1 until u=X.

[0130] The method of utilizing the basic NLP data desensitizing network to perform privacy semantic recognition on the key embedded semantic vector to obtain a privacy semantic representation vector includes: utilizing the Xth order semantic feature recognition subnet of the Xth order semantic feature recognition subnet to perform privacy semantic recognition on the key embedded semantic vector generated by the Xth order semantic feature mining subnet to obtain a privacy semantic representation vector generated by the Xth order semantic feature recognition subnet; utilizing the uth order semantic feature recognition subnet of the Xth order semantic feature recognition subnet to perform privacy semantic recognition on the privacy semantic representation vector generated by the u+1th order to obtain a privacy semantic representation vector generated by the uth order semantic feature recognition subnet; and performing a self-incremental 1 walk process on u to obtain a privacy semantic representation vector generated by each order semantic feature recognition subnet in the Xth order semantic feature recognition subnet.

[0131] In some optional embodiments, the basic NLP data desensitization network that the NLP data desensitization system relies on is designed to be quite sophisticated and has deep learning capabilities. This network is mainly composed of two parts: a semantic feature mining branch and a semantic feature recognition branch. Both branches use a multi-order connected sub-network structure to improve the depth and accuracy of data processing.

[0132] For example, the semantic feature mining branch contains X-order connected semantic feature mining subnets, which are used to perform layer-by-layer key embedding semantic mining on text data. Similarly, the semantic feature recognition branch is also composed of X-order connected semantic feature recognition subnets, which focus on the recognition and representation of privacy semantics. It is worth mentioning that the subnets of these two branches are one-to-one corresponding. This structure enables the system to accurately identify and process the privacy information while mining key semantics.

[0133] When performing key embedding semantic mining tasks, the NLP data desensitization system will first use the first-order semantic feature mining subnet to perform preliminary key semantic mining on the natural resources research report sample. The purpose of this stage is to extract the most basic and core semantic information from the report to form the first-order key embedding semantic vector. Subsequently, the system passes these vectors to the next-order subnet for deeper mining. This process will be carried out step by step, and each order subnet will further refine and mine the semantic vectors generated by the previous order until all X orders of processing are completed. Finally, the system obtains a series of key embedding semantic vectors generated by subnets of different orders, which together constitute a comprehensive and in-depth representation of the key semantics of the report.

[0134] When performing privacy semantic recognition, the NLP data desensitization system uses the opposite processing order of key embedding semantic mining. The system first uses the X-th order semantic feature recognition subnet to perform privacy semantic recognition on the key embedding semantic vectors generated by the last order mining subnet. The purpose of this stage is to accurately identify privacy information from the key semantics that have been mined and represent it as a privacy semantic representation vector. Then, the system passes these representation vectors to the previous order subnet for more detailed privacy recognition. Similarly, this process will be carried out step by step until all X-order processing is completed. In this way, the system can peel off non-privacy information layer by layer and finally obtain a semantic representation vector that is accurate to each privacy point.

[0135] Through this multi-order connected sub-network structure and step-by-step in-depth processing method, the NLP data desensitization system can not only fully mine the key semantic information in the natural resource survey report, but also accurately identify and characterize the privacy content therein. This technical solution significantly improves the accuracy and efficiency of data desensitization, and provides a strong guarantee for data security and sharing in the field of natural resource surveys. At the same time, through step-by-step in-depth processing, the system can also generate more concise and information-rich semantic representation vectors, providing more convenient and efficient support for subsequent data analysis and application. In general, this deep learning-based NLP data desensitization system has brought revolutionary changes to the processing and application of natural resource survey reports.

[0136] In the next step, the key embedded semantic vector generated by each order semantic feature mining subnet in the X-order semantic feature mining subnet has a corresponding first attention coefficient, and the privacy semantic representation vector generated by each order semantic feature recognition subnet in the X-order semantic feature recognition subnet has a corresponding second attention coefficient; according to the first attention coefficient of the key embedded semantic vector and the second attention coefficient of the privacy semantic representation vector, the key embedded semantic vector and the privacy semantic representation vector are subjected to semantic enhancement operation to obtain enhanced semantics of the research report, including: obtaining the key embedded semantic vector generated by the u-th order semantic feature mining subnet in the X-order semantic feature mining subnet The first attention coefficient of the vector is obtained, and the second attention coefficient of the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet in the x-th order semantic feature recognition subnet is obtained; according to the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet and the second attention coefficient of the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet, a semantic enhancement operation is performed on the key embedded semantic vector generated by the u-th order semantic feature mining subnet and the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet to obtain the u-th order research report enhanced semantics; u is processed by self-incrementing 1 to obtain the x-th order research report enhanced semantics.

[0137] Based on this embodiment, the NLP data desensitization system will further use the attention mechanism to strengthen the importance of key embedded semantic vectors and privacy semantic representation vectors. In this process, the system pays attention to the key embedded semantic vectors and privacy semantic representation vectors generated by each subnet, and assigns them the first attention coefficient and the second attention coefficient respectively.

[0138] For example, each key embedding semantic vector generated by the semantic feature mining subnet has a corresponding first attention coefficient. This coefficient reflects the importance of the vector in the overall semantics. The larger the coefficient, the more key information the vector contains and the greater its contribution to the overall semantics. Similarly, each privacy semantic representation vector generated by the semantic feature recognition subnet also has a second attention coefficient, which is used to measure the importance of the vector in identifying privacy information.

[0139] When the NLP data desensitization system performs semantic enhancement, it first obtains the first attention coefficient of the key embedded semantic vector and the second attention coefficient of the privacy semantic representation vector generated by each subnet. Then, the system weights the corresponding vectors according to these coefficients to enhance the important semantic information. This process is achieved by multiplying the vector with the corresponding attention coefficient to obtain the enhanced key embedded semantic vector and privacy semantic representation vector.

[0140] Next, the NLP data desensitization system fuses the enhanced key embedding semantic vector and the privacy semantic representation vector to generate enhanced semantics for each level of the research report. This step is achieved through a specific fusion algorithm that comprehensively considers the importance of key embedding semantics and privacy semantics and generates a more comprehensive and accurate enhanced semantics for the research report.

[0141] By performing a self-adding 1 walk process on u, the NLP data desensitization system processes the vectors generated by each subnet in turn, and finally obtains the enhanced semantics of the X-order research report. These enhanced semantics not only retain the key information in the original report, but also highlight the importance of privacy information, providing a more accurate and reliable basis for subsequent data processing and privacy protection.

[0142] In this way, by introducing the attention mechanism for semantic enhancement, the NLP data desensitization system can more accurately identify and enhance key information and privacy information in natural resource research reports. This technical solution not only improves the accuracy and efficiency of data desensitization, but also enables the desensitized report to better reflect the core intent and key information of the original report. At the same time, the enhanced semantics also provide richer and more valuable input for subsequent data analysis, mining and application.

[0143] In some preferred embodiments, the obtaining of the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet in the X-order semantic feature mining subnet includes: obtaining the original attention coefficient and the attention trend adjustment feature of the key embedded semantic vector generated by the u-th order semantic feature mining subnet, and determining the current debugging cycle of the basic NLP data desensitization network; determining the feature operation result of the attention trend adjustment feature and the current debugging cycle, and combining the original attention coefficient and the feature operation result to obtain the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet, the original attention coefficient is 1, and the first attention coefficient is greater than or equal to 0 and less than or equal to 1; the obtaining of the second attention coefficient of the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet in the X-order semantic feature recognition subnet includes: combining a preset value with the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet, and determining the second attention coefficient of the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet.

[0144] Based on this embodiment, the NLP data desensitization system adopts a more sophisticated and dynamic method when obtaining the first attention coefficient of the key embedded semantic vector and the second attention coefficient of the privacy semantic representation vector.

[0145] First, for the acquisition of the first attention coefficient, the system not only considers the original attention coefficient, but also introduces the attention trend adjustment feature and the current debugging cycle. For example, the system obtains the original attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet. This coefficient is usually set to 1 as the initial weight. Then, the system obtains the attention trend adjustment feature of the vector, which reflects the changing trend of the key embedded semantics in different debugging cycles. At the same time, the system also determines the current debugging cycle of the basic NLP data desensitization network to understand the progress of model training.

[0146] Next, the system performs feature operations based on the attention trend adjustment features and the current debugging cycle. This operation can include addition, multiplication, division or other complex mathematical operations, with the aim of obtaining an adjustment value that reflects the importance of the key embedded semantics at the current stage. The system then combines this adjustment value with the original attention coefficient, and obtains the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet through certain algorithms (such as weighted average, multiplication adjustment, etc.). This coefficient is a value between 0 and 1, and the closer it is to 1, the higher the importance of the vector at the current stage.

[0147] As for the acquisition of the second attention coefficient, the NLP data desensitization system adopts a method associated with the first attention coefficient. The system combines a preset value with the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet to determine the second attention coefficient. This preset value may be an empirical value or an optimal parameter obtained through experiments, which is used to adjust the importance of the privacy semantic representation vector in the overall semantics. In this way, the system can ensure that the recognition of privacy semantics and the mining of key embedded semantics maintain a dynamic balance, thereby improving the accuracy and efficiency of data desensitization.

[0148] Through these preferred embodiments, the NLP data desensitization system can more accurately obtain the attention coefficients of key embedded semantic vectors and privacy semantic representation vectors. This not only helps to improve the accuracy of data desensitization, but also can dynamically adjust the value of the coefficient according to the progress of model training and the changing trend of key embedded semantics, so that the desensitization results are more in line with actual needs. At the same time, this method can also enhance the flexibility and adaptability of the system, enabling it to cope with more complex and changeable data processing scenarios.

[0149] In some examples, the attention trend adjustment features of the key embedded semantic vectors generated by each order semantic feature mining subnet in the X-order semantic feature mining subnet are different, and the attention trend adjustment features of the key embedded semantic vectors have a set quantitative relationship (such as a negative correlation) with the order of the semantic feature mining subnet that outputs the key embedded semantic vector.

[0150] Based on this example, the X-order semantic feature mining subnet in the NLP data desensitization system shows unique characteristics. It should be mentioned here that each order of semantic feature mining subnet has an associated attention trend adjustment feature when generating key embedding semantic vectors. This feature is not static, but changes dynamically according to the order of the semantic feature mining subnet.

[0151] For example, there is a set quantitative relationship between the attention trend regulation feature of the key semantic embedding vector and the order of the semantic feature mining subnet that outputs the vector. Taking the negative correlation as an example, as the order of the semantic feature mining subnet increases, the attention trend regulation feature of the generated key semantic embedding vector may gradually decrease. This means that at higher orders, the system may pay more attention to the semantic features of details or specific aspects rather than the overall or general semantics.

[0152] To explain this more vividly, you can imagine a multi-level text analysis process. In the initial stage, the system may pay more attention to the overall meaning and main concepts of the text, so the attention trend regulation feature will be relatively high. However, as the analysis deepens and the order increases, the system begins to focus on more specific and subtle semantic features, and the attention trend regulation feature will decrease accordingly to highlight these specific aspects of information.

[0153] This setting not only helps the NLP data desensitization system process and analyze text data more accurately, but also improves the flexibility and adaptability of the system. By dynamically adjusting the attention trend adjustment feature, the system can better capture the key information in the text and emphasize or downplay certain aspects of the semantic features as needed.

[0154] In this way, the key embedded semantic vector generated by each order semantic feature mining subnet in the NLP data desensitization system has a unique attention trend regulation feature, and there is a set quantitative relationship between this feature and the order of the subnet. This design not only improves the accuracy and efficiency of the system in processing text data, but also enables it to more flexibly respond to various complex semantic analysis needs.

[0155] In some other possible embodiments, the NLP text data repetition operation is performed on the key embedded semantic vector, the privacy semantic representation vector and the survey report enhanced semantics to obtain a natural resource survey repetition report sample of the natural resource survey report sample, including: integrating the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet in the X-order semantic feature recognition subnet with the u-th order survey report enhanced semantics in the X-order survey report enhanced semantics to obtain a u-th order text semantic integration vector; performing a self-incremental 1 walk process on u to obtain an X-order text semantic integration vector; integrating the key embedded semantic vector generated by the X-th order semantic feature mining subnet in the X-order semantic feature mining subnet and the X-order text semantic integration vector to obtain the natural resource survey report repetition semantics; performing a text repetition operation on the natural resource survey report repetition semantics to obtain the natural resource survey repetition report sample of the natural resource survey report sample.

[0156] In this embodiment, the NLP data desensitization system further performs an NLP text data retelling operation to generate a sample natural resource survey retelling report.

[0157] First, the system integrates the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet in the X-th order semantic feature recognition subnet with the corresponding u-th order research report enhanced semantics in the X-th order research report enhanced semantics. This step is done through specific algorithms, such as vector concatenation, weighted averaging, or fusion through neural networks, to obtain a new vector, namely the u-th order text semantic integration vector. This vector contains both privacy semantic information and the enhanced research report semantics, so it can more comprehensively represent the semantic content of this order.

[0158] Next, the system performs a self-adding 1 walk on the variable u, i.e. u=u+1, and then repeats the above integration steps until all orders are processed. In this way, the system can obtain a series of text semantic integration vectors, from the 1st order to the Xth order, which together constitute the Xth order text semantic integration vector.

[0159] Then, the system further integrates the key embedded semantic vector generated by the highest order, that is, the Xth order semantic feature mining subnet, with the Xth order text semantic integration vector obtained previously. The purpose of this step is to integrate the key embedded semantics with the text semantics that have been integrated with privacy and enhancement semantics, so as to obtain a more complete and rich natural resources research report retelling semantics.

[0160] Finally, the system performs a text paraphrase operation on the paraphrase semantics. This step is usually achieved through text generation technology in NLP, such as using a sequence-to-sequence (Seq2Seq) model or other advanced text generation models to convert the paraphrase semantics into natural language text. The final generated text is the natural resource survey paraphrase report example of the natural resource survey report example.

[0161] In this way, through multi-level semantic integration and paraphrase operations, the NLP data desensitization system can generate a paraphrase report that retains the key information of the original report and effectively desensitizes the privacy content. Such a report not only meets the dual requirements of data sharing and privacy protection, but also improves the availability and security of data. At the same time, since the enhanced semantic information is incorporated into the paraphrase process, the generated paraphrase report is more accurate and rich in semantic expression, which helps to improve the effect of subsequent data analysis and mining.

[0162] In an alternative embodiment, before performing network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitizing network, the method also includes: performing privacy decision analysis on the natural resource survey report sample through a text privacy decision branch to obtain a first privacy decision analysis point of view, and performing privacy decision analysis on the natural resource survey retelling report sample through the text privacy decision branch to obtain a second privacy decision analysis point of view; wherein the first privacy decision analysis point of view is used to characterize the probability value that the natural resource survey report sample is generated by the basic NLP data desensitizing network, and the second privacy decision analysis point of view is used to characterize the probability value that the natural resource survey retelling report sample is generated by the basic NLP data desensitizing network.

[0163] Then, the network debugging quality analysis is performed based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitizing network, including: performing natural resource survey report network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the first error variable of the natural resource survey report debugging error of the basic NLP data desensitizing network; performing privacy decision analysis network debugging quality analysis based on the first privacy decision analysis viewpoint and the second privacy decision analysis viewpoint to obtain the first decision offset weight of the first decision analysis debugging error of the basic NLP data desensitizing network; and taking the sum of the first error variable and the first decision offset weight as the error variable of the debugging error of the basic NLP data desensitizing network.

[0164] Through this embodiment, the NLP data desensitization system adds a privacy decision analysis step before calculating the debugging error of the basic NLP data desensitization network. The purpose of this step is to more comprehensively evaluate the performance of the desensitization network when processing natural resource research reports.

[0165] For example, the NLP data desensitization system first performs a privacy decision analysis on the original natural resource research report sample through the text privacy decision branch to obtain a first privacy decision analysis point of view. This point of view is actually a probability value, which is used to characterize the possibility that the original report is generated by the basic NLP data desensitization network. In other words, it reflects the ability of the desensitization network to retain the key information of the original report.

[0166] Similarly, the system will also perform privacy decision analysis on the desensitized natural resource survey retelling report sample and obtain the second privacy decision analysis opinion. This opinion is also a probability value, but it is used to characterize the possibility that the retelling report is generated by the desensitized network. This can be seen as a measure of the desensitized network in protecting privacy information while maintaining report coherence.

[0167] Next, the NLP data desensitization system conducts a quality analysis of the natural resource survey report network debugging based on the differences between the natural resource survey retelling report sample and the original report sample. This analysis is mainly to find out the errors or deficiencies that may occur when the desensitization network processes the report, thereby obtaining a first error variable. This error variable reflects the performance of the desensitization network in the text retelling process.

[0168] In addition, the system will also perform a privacy decision analysis network debugging quality analysis based on the first and second privacy decision analysis viewpoints. The purpose of this step is to evaluate the accuracy of the desensitized network in terms of privacy decisions and obtain a first decision deviation weight of the first decision analysis debugging error. This weight actually reflects the degree of deviation of the desensitized network in terms of privacy protection.

[0169] Finally, the NLP data desensitization system uses the sum of the first error variable and the first decision offset weight as the error variable of the debugging error of the basic NLP data desensitization network. This error variable comprehensively considers the performance of the desensitization network in terms of text repetition and privacy protection, so it can more comprehensively evaluate the performance of the desensitization network.

[0170] Through this implementation method, the NLP data desensitization system can more accurately identify the problems that may exist when the desensitization network processes natural resource research reports, and thus make targeted optimizations and improvements. This not only helps to improve the performance of the desensitization network, but also better meet the needs of data sharing and privacy protection.

[0171] In some alternative technical solutions, before performing network debugging quality analysis based on the difference between the natural resources survey retelling report sample and the natural resources survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitizing network, the method also includes: extracting the comment opinion semantic vector of the natural resources survey report sample to obtain the comment opinion semantic vector of the natural resources survey report sample; using the basic NLP data desensitizing network to perform semantic paraphrase optimization on the comment opinion semantic vector to obtain the semantic paraphrase vector of the natural resources survey report sample.

[0172] Based on this technical solution, the NLP data desensitization system adds a step of extracting the semantic vector of the comments in the natural resource survey report sample before performing the network debugging quality analysis. The purpose of this step is to more deeply understand and analyze the comments in the report, thereby further improving the performance and accuracy of the desensitization network.

[0173] First, the NLP data desensitization system extracts the semantic vectors of comments and opinions from the sample natural resource research report. In this process, the system uses natural language processing technologies, such as word embedding, named entity recognition, and sentiment analysis, to identify and extract the comments and opinions in the report and their related semantic information. These comments and opinions can include evaluations of natural resources, views on research results, and predictions of future development trends, which are all important information components in the report.

[0174] Through refinement, the system obtains a semantic vector representing the comments and opinions of the sample natural resources research report. This vector not only contains the content of the comments and opinions, but also contains the position and relationship of these opinions in the semantic space, so as to more comprehensively reflect the comments and opinions in the report.

[0175] Next, the NLP data desensitization system uses the basic NLP data desensitization network to perform semantic paraphrase optimization on the semantic vector of the comment. The purpose of this step is to desensitize and optimize the original comment on the basis of retaining the semantics of the original comment to obtain a new semantic paraphrase vector. This paraphrase vector can not only reflect the core content of the original comment, but also achieve anonymous desensitization protection of sensitive information to a certain extent.

[0176] In the process of semantic paraphrase optimization, the desensitization network will comprehensively consider multiple factors, such as semantic coherence, information integrity, privacy protection, etc., to ensure that the generated semantic paraphrase vector not only meets the desensitization requirements but also maintains the main meaning of the original comment.

[0177] Through this technical solution, the NLP data desensitization system can more deeply understand and process the comments and opinions in the natural resource research report, thereby further improving the performance and accuracy of the desensitization network. This not only helps to protect the privacy information in the report, but also provides richer and more valuable information resources for subsequent data analysis and mining.

[0178] In this way, by refining the semantic vectors of comments and optimizing semantic repetitions of natural resource research reports, the NLP data desensitization system can more comprehensively understand and analyze the comments in the report, and improve the accuracy and effectiveness of desensitization processing. This technical solution not only enhances the security and availability of data, but also provides stronger support for subsequent data processing and analysis.

[0179] In an optional technical solution, the network debugging quality analysis is performed based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitizing network, including: performing a natural resource survey report network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the first error variable of the natural resource survey report debugging error of the basic NLP data desensitizing network; performing a feature network debugging quality analysis based on the difference between the semantic retelling vector and the comment opinion semantic vector to obtain the second error variable of the semantic debugging error of the basic NLP data desensitizing network; and taking the sum of the first error variable and the second error variable as the error variable of the debugging error of the basic NLP data desensitizing network.

[0180] Based on this technical solution, the NLP data desensitization system will comprehensively consider multiple factors when performing network debugging quality analysis to obtain a more comprehensive and accurate debugging error.

[0181] First, the system analyzes the debugging quality of the natural resource survey report network based on the differences between the natural resource survey paraphrase report sample and the natural resource survey report sample. The purpose of this step is to evaluate the performance of the desensitized network in processing natural resource survey reports, especially its accuracy in text paraphrasing. By comparing and analyzing the differences between the paraphrased report sample and the original report sample, the system can calculate a first error variable that reflects the debugging error of the desensitized network in report paraphrasing.

[0182] Secondly, the system analyzes the debugging quality of the feature network based on the difference between the semantic paraphrase vector and the comment view semantic vector. In this step, the system focuses on the performance of the desensitized network in processing and optimizing the comment view semantics. By comparing the difference between the semantic paraphrase vector (the comment view semantics after being processed by the desensitized network) and the original comment view semantic vector, the system can calculate a second error variable that reflects the debugging error of the desensitized network in semantic processing.

[0183] Finally, the system uses the sum of the first error variable and the second error variable as the error variable of the debugging error of the basic NLP data desensitization network. This comprehensive error variable can more comprehensively reflect the overall performance of the desensitization network in processing the semantics of natural resource research reports and comment opinions.

[0184] Through this technical solution, the NLP data desensitization system can more accurately evaluate and optimize the performance of the desensitization network. This not only helps to improve the accuracy of the desensitization network in text retelling and semantic processing, but also provides more reliable and higher-quality data support for subsequent data processing and analysis.

[0185] In summary, the NLP data desensitization system obtains a more comprehensive and accurate debugging error by comprehensively considering the differences between the natural resource survey retelling report and the original report, as well as the differences between the semantic retelling vector and the semantic vector of the original comment opinion. This technical solution not only improves the performance evaluation accuracy of the desensitization network, but also provides a strong basis for optimizing the network, thereby further improving the effect and quality of data desensitization.

[0186] In some other possible embodiments, the method also includes: performing a privacy decision analysis on the natural resources survey report sample through a text privacy decision branch to obtain a first privacy decision analysis opinion, and performing a privacy decision analysis on the natural resources survey retelling report sample through the text privacy decision branch to obtain a second privacy decision analysis opinion; wherein the first privacy decision analysis opinion is used to characterize that the natural resources survey report sample is a probability value generated by the basic NLP data desensitization network, and the second privacy decision analysis opinion is used to characterize that the natural resources survey retelling report sample is a probability value generated by the basic NLP data desensitization network; performing a privacy decision analysis on the comment opinion semantic vector through the text privacy decision branch to obtain a third privacy decision analysis opinion, and performing a privacy decision analysis on the semantic retelling vector through the text privacy decision branch to obtain a fourth privacy decision analysis opinion; wherein the third privacy decision analysis opinion is used to characterize that the comment opinion semantic vector is a probability value generated by the basic NLP data desensitization network, and the fourth privacy decision analysis opinion is used to characterize that the semantic retelling vector is a probability value generated by the basic NLP data desensitization network.

[0187] Then, the network debugging quality analysis is performed based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitizing network, including: performing the natural resource survey report network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the first error variable of the natural resource survey report debugging error of the basic NLP data desensitizing network, and performing the feature network debugging quality analysis based on the difference between the semantic retelling vector and the comment view semantic vector to obtain the second error variable of the semantic debugging error of the basic NLP data desensitizing network; A privacy decision analysis network debugging quality analysis is performed based on the first privacy decision analysis viewpoint and the second privacy decision analysis viewpoint to obtain a first decision offset weight of the first decision analysis debugging error of the basic NLP data desensitizing network, and a privacy decision analysis network debugging quality analysis is performed based on the third privacy decision analysis viewpoint and the fourth privacy decision analysis viewpoint to obtain a second decision offset weight of the second decision analysis debugging error of the basic NLP data desensitizing network; the sum of the first error variable, the second error variable, the first decision offset weight and the second decision offset weight is used as the error variable of the debugging error of the basic NLP data desensitizing network.

[0188] Based on this embodiment, the NLP data desensitization system introduces more detailed privacy decision analysis steps when processing marine natural resources survey reports to ensure the comprehensiveness and accuracy of the desensitization processing.

[0189] First, the system performs a privacy decision analysis on the original marine natural resources survey report sample through the text privacy decision branch, and obtains a first privacy decision analysis view. This view is actually a probability value, indicating the possibility that the report sample is generated by the basic NLP data desensitization network. Similarly, the system will also perform a similar privacy decision analysis on the desensitized marine natural resources survey retelling report sample to obtain a second privacy decision analysis view.

[0190] In addition, the system will also perform privacy decision analysis on the comment semantic vector extracted from the report and the semantic paraphrase vector optimized by the desensitizing network. The purpose of this step is to evaluate the privacy protection ability of the desensitizing network when processing these key semantic information. Through the analysis of the text privacy decision branch, the system obtains the third privacy decision analysis point of view and the fourth privacy decision analysis point of view, which respectively represent the probability values ​​of the comment semantic vector and the semantic paraphrase vector generated by the desensitizing network.

[0191] Next, the NLP data desensitization system conducts a network debugging quality analysis to calculate the debugging error of the desensitized network. In this process, the system comprehensively considers multiple factors. First, the system conducts a natural resource survey report network debugging quality analysis based on the difference between the marine natural resource survey retelling report sample and the original report sample, and obtains a first error variable that reflects the accuracy of the report retelling. At the same time, the system also conducts a feature network debugging quality analysis based on the difference between the semantic retelling vector and the comment view semantic vector, and obtains a second error variable that reflects the accuracy of semantic processing.

[0192] In order to more comprehensively evaluate the performance of the desensitized network, the system will also perform a privacy decision analysis network debugging quality analysis based on the first privacy decision analysis viewpoint and the second privacy decision analysis viewpoint, and obtain a first decision offset weight that reflects the privacy protection performance of the desensitized network. Similarly, the system will also perform similar analysis based on the third privacy decision analysis viewpoint and the fourth privacy decision analysis viewpoint, and obtain a second decision offset weight that reflects the privacy protection ability of the desensitized network when processing key semantic information.

[0193] Finally, the NLP data desensitization system uses the sum of the first error variable, the second error variable, the first decision offset weight, and the second decision offset weight as the error variable of the debugging error of the basic NLP data desensitization network. This comprehensive error variable can more comprehensively reflect the overall performance of the desensitization network when processing marine natural resources research reports.

[0194] Through this technical solution, the NLP data desensitization system can more accurately evaluate and optimize the performance of the desensitization network, thereby ensuring the availability and accuracy of data while protecting privacy. This is of great significance for the processing of sensitive data such as marine natural resources survey reports. It can not only improve data security, but also provide more reliable and higher-quality data support for subsequent data analysis and mining.

[0195] In this way, the NLP data desensitization system obtains a more comprehensive and accurate debugging error by introducing detailed privacy decision analysis steps and network debugging quality analysis. This technical solution has shown significant advantages in protecting the privacy of sensitive data such as marine natural resources research reports, and provides a strong guarantee for the safe sharing and efficient use of data.

[0196] It is worth mentioning that in the above embodiment, the NLP data desensitization system performs a series of complex analysis and calculation steps to determine the debugging error of the basic NLP data desensitization network. These steps can be summarized and introduced by the following formula:

[0197] The first privacy decision analysis perspective (P1) and the second privacy decision analysis perspective (P2):

[0198] (P1=text{Privacy decision analysis}(Sample of natural resources research report));

[0199] (P2=text{Privacy decision analysis}(Sample of natural resource research report));

[0200] Among them, (P1) and (P2) are probability values, indicating the possibility that the corresponding report sample is generated by the NLP data desensitization network.

[0201] The third privacy decision analysis perspective (P3) and the fourth privacy decision analysis perspective (P4):

[0202] (P3=text{Privacy decision analysis}(Comment opinion semantic vector));

[0203] (P4=text{Privacy decision analysis}(semantic paraphrase vector));

[0204] (P3) and (P4) are also probability values, which respectively represent the probability that the comment opinion semantic vector and semantic paraphrase vector are generated by the NLP data desensitization network.

[0205] The first error variable (E1) and the second error variable (E2):

[0206] (E1=text{Network debugging quality analysis}(Natural resources research report sample, Natural resources research retelling report sample));

[0207] (E2=text{network debugging quality analysis}(comment view semantic vector, semantic paraphrase vector));

[0208] (E1) represents the network debugging error of natural resource survey reports, while (E2) represents the network debugging error of semantic features.

[0209] The first decision bias weight (W1) and the second decision bias weight (W2):

[0210] (W1=text{Privacy Decision Analysis Network Debugging Quality Analysis}(P1,P2));

[0211] (W2=text{Privacy Decision Analysis Network Debugging Quality Analysis}(P3,P4));

[0212] (W1) and (W2) are the decision bias weights derived from the network debugging quality analysis based on the privacy decision analysis perspective.

[0213] Error variable for debugging error (E):

[0214] (E=E1+E2+W1+W2);

[0215] The final debugging error (E) is the sum of the first error variable, the second error variable, the first decision offset weight, and the second decision offset weight.

[0216] The above formula is used to quantify and evaluate the performance of the NLP data desensitization network when processing natural resource survey reports and their related semantic vectors. In this way, the desensitization network can be systematically debugged and optimized to improve the accuracy and efficiency of data desensitization.

[0217] Further, Figure 2 The structural block diagram of the NLP data desensitization system 300 is shown, including: a memory 310 for storing program instructions and data; a processor 320 for coupling with the memory 310 and executing the instructions in the memory 310 to implement the above method.

[0218] Furthermore, a computer storage medium is provided, comprising instructions, which implement the above method when executed on a processor.

[0219] In several embodiments provided in the embodiments of the present application, it should be understood that the disclosed device and method can also be implemented in other ways. The device and method embodiments described above are merely schematic, for example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the device, method and computer program product according to multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram can represent a module, a program segment or a part of a code, and the module, a program segment or a part of a code contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented with a dedicated hardware-based system that performs a specified function or action, or can be implemented with a combination of dedicated hardware and computer instructions.

[0220] In addition, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.

[0221] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium, including several instructions to enable a computer device (which can be a personal computer, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program code. It should be noted that in this article, the term "include", "include" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements that are not explicitly listed, or also includes elements inherent to such process, method, article or device. Without more constraints, an element defined by the phrase "comprising a..." does not exclude the existence of other identical elements in the process, method, article or apparatus comprising the element.

[0222] The above description is only the preferred embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A natural resource data desensitization method based on NLP, characterized in that: Applied to the NLP data desensitization system, the method includes: Using the basic NLP data desensitization network, perform key embedded semantic mining on the natural resources research report sample to obtain the key embedded semantic vector, and use the basic NLP data desensitization network to perform privacy semantic recognition on the key embedded semantic vector to obtain the privacy semantic representation vector; According to the first attention coefficient of the key embedded semantic vector and the second attention coefficient of the privacy semantic representation vector, a semantic enhancement operation is performed on the key embedded semantic vector and the privacy semantic representation vector to obtain enhanced semantics of the research report; Using the basic NLP data desensitization network, perform NLP text data paraphrase operation on the key embedded semantic vector, the privacy semantic representation vector and the research report enhanced semantics to obtain a natural resource research paraphrase report sample of the natural resource research report sample; Perform network debugging quality analysis based on the difference between the natural resources survey retelling report sample and the natural resources survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network; According to the error variable of the debugging error, the neuron weights of the basic NLP data desensitization network are optimized to obtain the debugged NLP data desensitization network; The basic NLP data desensitization network includes a semantic feature mining branch and a semantic feature recognition branch, the semantic feature mining branch includes an X-order connected semantic feature mining subnet, the semantic feature recognition branch includes an X-order connected semantic feature recognition subnet, the semantic feature mining subnet and the semantic feature recognition subnet correspond one to one, and X is an integer greater than 0; the basic NLP data desensitization network is used to perform key embedded semantic mining on a natural resources investigation report sample to obtain a key embedded semantic vector, including: Using the first-order semantic feature mining subnet of the X-order semantic feature mining subnet, key embedding semantic mining is performed on the natural resources investigation report sample to obtain a key embedding semantic vector generated by the first-order semantic feature mining subnet; Using the u-th order semantic feature mining subnet of the x-th order semantic feature mining subnet, perform key embedding semantic mining on the key embedding semantic vector generated at the u-1th order, and obtain the key embedding semantic vector generated by the u-th order semantic feature mining subnet; Performing a self-incremental 1 walk process on u to obtain a key embedded semantic vector generated by each order semantic feature mining subnet in the X-order semantic feature mining subnet, wherein u is an integer greater than 0 and not greater than X, and the self-incremental 1 walk process on u is to increase the value of u by 1 until u=X; The method of using the basic NLP data desensitization network to perform privacy semantics recognition on the key embedded semantic vector to obtain a privacy semantic representation vector includes: Using the Xth-order semantic feature recognition subnet of the Xth-order semantic feature recognition subnet, privacy semantic recognition is performed on the key embedded semantic vector generated by the Xth-order semantic feature mining subnet to obtain the privacy semantic representation vector generated by the Xth-order semantic feature recognition subnet; Using the u-th order semantic feature recognition subnet of the x-th order semantic feature recognition subnet, perform privacy semantic recognition on the privacy semantic representation vector generated at the u+1th order, to obtain the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet; Perform a self-incremental 1 walk process on u to obtain a privacy semantic representation vector generated by each order semantic feature recognition subnet in the X-order semantic feature recognition subnet; The key embedded semantic vector generated by each order semantic feature mining subnet in the X-order semantic feature mining subnet has a corresponding first attention coefficient, and the privacy semantic representation vector generated by each order semantic feature recognition subnet in the X-order semantic feature recognition subnet has a corresponding second attention coefficient; the semantic enhancement operation is performed on the key embedded semantic vector and the privacy semantic representation vector according to the first attention coefficient of the key embedded semantic vector and the second attention coefficient of the privacy semantic representation vector to obtain enhanced semantics of the research report, including: Obtaining a first attention coefficient of a key embedded semantic vector generated by a u-th order semantic feature mining subnet in the X-th order semantic feature mining subnet, and obtaining a second attention coefficient of a privacy semantic representation vector generated by a u-th order semantic feature recognition subnet in the X-th order semantic feature recognition subnet; According to the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet and the second attention coefficient of the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet, a semantic enhancement operation is performed on the key embedded semantic vector generated by the u-th order semantic feature mining subnet and the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet to obtain the u-th order research report enhanced semantics; Perform a self-adding 1 walk process on u to obtain the enhanced semantics of the research report of order X; The obtaining of the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet in the x-th order semantic feature mining subnet includes: obtaining the original attention coefficient and the attention trend adjustment feature of the key embedded semantic vector generated by the u-th order semantic feature mining subnet, and determining the current debugging cycle of the basic NLP data desensitization network; determining the attention trend adjustment feature and the feature operation result of the current debugging cycle, and combining the original attention coefficient and the feature operation result to obtain the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet, wherein the original attention coefficient is 1, and the first attention coefficient is greater than or equal to 0 and less than or equal to 1; The obtaining of the second attention coefficient of the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet in the x-th order semantic feature recognition subnet includes: combining a preset value with the first attention coefficient of the key embedded semantic vector generated by the u-th order semantic feature mining subnet to determine the second attention coefficient of the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet; The attention trend adjustment features of the key embedded semantic vectors generated by each order semantic feature mining subnet in the X-order semantic feature mining subnet are different, and the attention trend adjustment features of the key embedded semantic vectors and the order of the semantic feature mining subnet that outputs the key embedded semantic vector have a set quantitative relationship; The NLP text data retelling operation is performed on the key embedded semantic vector, the privacy semantic representation vector and the research report enhanced semantics to obtain a natural resource research retelling report sample of the natural resource research report sample, including: Integrate the privacy semantic representation vector generated by the u-th order semantic feature recognition subnet in the X-th order semantic feature recognition subnet with the u-th order research report enhanced semantics in the X-th order research report enhanced semantics to obtain the u-th order text semantic integrated vector; perform a self-incremental 1 walk process on u to obtain the X-th order text semantic integrated vector; Integrate the key embedded semantic vector generated by the X-th order semantic feature mining subnet in the X-order semantic feature mining subnet and the X-order text semantic integration vector to obtain the natural resources investigation report retelling semantics; Performing a text paraphrase operation on the paraphrase semantics of the natural resources investigation report to obtain a natural resources investigation paraphrase report sample of the natural resources investigation report sample; Before performing network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network, the method further includes: Performing a privacy decision analysis on the natural resource survey report sample through the text privacy decision branch to obtain a first privacy decision analysis viewpoint, and performing a privacy decision analysis on the natural resource survey retelling report sample through the text privacy decision branch to obtain a second privacy decision analysis viewpoint; The first privacy decision analysis viewpoint is used to characterize the probability value of the natural resource survey report sample being generated by the basic NLP data desensitization network, and the second privacy decision analysis viewpoint is used to characterize the probability value of the natural resource survey retelling report sample being generated by the basic NLP data desensitization network; The network debugging quality analysis is performed based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network, including: Performing a natural resource survey report network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample, and obtaining a first error variable of the natural resource survey report debugging error of the basic NLP data desensitization network; Performing a privacy decision analysis network debugging quality analysis based on the first privacy decision analysis viewpoint and the second privacy decision analysis viewpoint to obtain a first decision offset weight of a first decision analysis debugging error of the basic NLP data desensitization network; The sum of the first error variable and the first decision offset weight is used as the error variable of the debugging error of the basic NLP data desensitization network.

2. The method according to claim 1, characterized in that Before performing network debugging quality analysis based on the difference between the natural resource survey retelling report sample and the natural resource survey report sample to obtain the error variable of the debugging error of the basic NLP data desensitization network, the method further includes: Extracting the semantic vector of the comments on the sample of the natural resources investigation report to obtain the semantic vector of the comments on the sample of the natural resources investigation report; The basic NLP data desensitization network is used to perform semantic paraphrase optimization on the semantic vector of the comment opinion, thereby obtaining the semantic paraphrase vector of the natural resources research report sample.

3. A NLP data desensitization system, characterized in that: include: Memory, used to store program instructions and data; A processor, coupled to a memory, and configured to execute instructions in the memory to implement the method according to any one of claims 1 to 2.

4. A computer storage medium, characterized in that: Contains instructions, which, when executed on a processor, implement the method according to any one of claims 1 to 2.

Images