Method for searching for intranet confidential standard text, electronic device and medium

Abstract

The application relates to the technical field of text retrieval, in particular to a method for searching internal network confidential standard text, an electronic device and a medium. The method comprises the following steps: obtaining a first vector on the browser side through a first pre-training model; performing first dimension reduction on the first vector to obtain a second vector; performing second dimension reduction on the first vector to obtain a third vector; generating a first request message based on the second vector and the third vector; the first request message is sent to a server, the server obtains a query result based on the first request message, and the query result is sent to the browser. The application solves the problems of low accuracy of the search process and the search result of the confidential file search.

Description

Technical Field

[0001] This application relates to the field of classified text retrieval technology, and in particular to a method, electronic device, and medium for retrieving classified standard texts on an intranet. Background Technology

[0002] Classified documents refer to documents containing sensitive information related to government, military, scientific research, finance, and other fields. With the increasing number of classified documents, how to conduct secure retrieval of these documents has become a key research focus. The retrieval behavior of personnel in classified institutions is highly sensitive and can reflect the institution's current research priorities, technological bottlenecks, or operational intentions. Therefore, improving the confidentiality of retrieval activities has become a current research priority.

[0003] Since most current text retrieval technologies are based on keyword matching or basic semantic embedding, this retrieval method suffers from problems such as easy vulnerability to attackers leading to data loss during the retrieval process and low accuracy of retrieval results. Summary of the Invention

[0004] At least one aspect and advantage of this application will be set forth in part in the description which follows, or may be apparent from the description, or may be obtained by practicing the subject matter of this disclosure.

[0005] According to the first aspect of this application, a method for retrieving classified standard texts on an intranet includes: The first vector is obtained on the browser side using the first pre-trained model; The first dimension reduction is performed on the first vector to obtain the second vector; The third vector is obtained by performing a second dimensionality reduction on the first vector. The first request message is generated based on the second and third vectors; The first request message is sent to the server so that the server can obtain the query result based on the first request message and then send the query result to the browser.

[0006] According to one embodiment of this application, the process of obtaining the second vector includes: reducing the dimensionality of the first vector using a first matrix to obtain a second vector with reduced dimensionality; The process of obtaining the third vector includes: using the second matrix to reduce the dimensionality of the first vector to obtain the third vector with reduced dimensionality.

[0007] According to one embodiment of this application, noise is added to the second vector to change the value of a second percentage within the second vector, where the second percentage is 5% to 15%. Noise is added to the second vector to change the value of the second percentage within the second vector, which is between 5% and 15%.

[0008] According to one embodiment of this application, before acquiring the second or third vector, noise is added to the first vector to change the value of a first percentage within the first vector, where the first percentage is 5% to 15%.

[0009] According to one embodiment of this application, the ratio of the second or third vector dimension to the first vector dimension is 0.2:1 to 0.3:1.

[0010] According to one embodiment of this application, the process of obtaining query results includes: The first request message is used to obtain a set of vectors for retrieval. Based on the vectors in the set, the first and second vector data tables are searched to obtain the first and second retrieval results. The query results are obtained based on the first and second search results; The matching records in the first and second search results include document identifiers, similarity scores, and document vectors.

[0011] According to one embodiment of this application, the process of obtaining query results includes: The first result set in the first search result and the second result set in the second search result are determined based on the document identifiers shared by the first search result and the second search result, respectively. The fourth vector is determined based on the first result set and the first search result; The fifth vector is determined based on the second result set and the second search result; Based on the fourth vector, a third search result is obtained by searching within the first vector data table if the similarity to the fourth vector is greater than the first threshold. Based on the fifth vector, a fourth search result is obtained from the first vector data table if the similarity to the fourth vector is greater than the first threshold. The query results are obtained by merging the third and fourth search results.

[0012] According to one embodiment of this application, the search results also include documents corresponding to document identifiers common to both the first and second search results.

[0013] According to a second aspect of this application, a device for retrieving classified standard texts to prevent information leakage is provided, the device comprising: The query vector determination module is used to obtain the first vector on the browser side through the first pre-trained model; The first dimension reduction processing module is used to perform a first dimension reduction on the first vector to obtain the second vector; The second dimension reduction module is used to perform a second dimension reduction on the first vector to obtain the third vector; The request message generation module is used to generate a first request message based on the second vector and the third vector. The request message processing module is used to send the first request message to the server so that the server can obtain the query result based on the first request message and then send the query result to the browser.

[0014] According to a third aspect of this application, an electronic device is provided, comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the steps of the above-described method.

[0015] According to a fourth aspect of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described method.

[0016] In this embodiment, a first vector is obtained on the browser side through a first pre-trained model. The first vector is then processed using different dimensionality reduction algorithms to obtain a second vector and a third vector. Based on the second and third vectors, a first request message is generated and sent to the server. The server then obtains the query results based on the first request message and sends the query results back to the browser. Since the first pre-trained model used is based on a technology in the field of classified information, the first vector generated in this way has higher confidentiality compared to vectors generated by general models. This increases the difficulty for attackers to crack the first vector, thus initially enhancing the security of the first vector. By generating two vectors from the first vector through dimensionality reduction processing to generate the request message, the sensitive information in the original query is reduced, avoiding the leakage problem caused by an attacker intercepting a vector in the message. Furthermore, it enables retrieval from different semantic features, thereby improving the retrieval accuracy on the server side. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments of this application will be briefly introduced below.

[0018] Figure 1 A flowchart illustrating a method for retrieving classified standard texts on an intranet, provided as an embodiment of this application; Figure 2 This is a block diagram of a device for retrieving classified standard texts on an intranet, provided as an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0020] It should be noted that although the functional modules are divided in the device diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than the module division in the device or the order in the flowchart.

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0022] According to one embodiment of this application, a method for retrieving classified standard texts on an intranet is provided, such as... Figure 1 As shown, the method includes steps S101 to S105.

[0023] Step S101: Obtain the first vector on the browser side through the first pre-trained model.

[0024] It should be noted that the first pre-trained model used in this application utilizes a corpus containing several sensitive words, covering various classified technical fields. Therefore, compared to general pre-trained models using public corpora (such as the bidirectional encoder representation model BERT), their vector spaces are not the same. This is mainly manifested in the fact that the position of the same query text in the vector space of the first pre-trained model and its distance from the vector space may differ. Therefore, this embodiment uses the first pre-trained model to increase the difficulty for attackers to crack the first vector, thereby enhancing the security of the first vector.

[0025] Step S102: Perform a first dimensionality reduction on the first vector to obtain the second vector.

[0026] It should be noted that the second vector refers to a vector segment with some sensitive information removed. In practice, a preset dimensionality reduction algorithm can be used to process the first vector to reduce its information and achieve the dimensionality reduction objective. For example, a segment of a predetermined length can be extracted from the first vector as the second vector.

[0027] Step S103: Perform a second dimensionality reduction on the first vector to obtain the third vector.

[0028] It should be noted that the third vector refers to a vector segment that differs from the second vector after removing some sensitive information. Since the second and third vectors point to different semantic features, this reduces the amount of information in the original query and allows for retrieval from different dimensions, thereby improving retrieval accuracy. For example, if the semantic feature of the second vector is time-related, and the semantic feature of the third vector is technology-related, then subsequent query results can be retrieved through both technology and time dimensions.

[0029] In practice, a dimensionality reduction method for obtaining the second vector can be used to process the first vector to obtain a third vector that is different from the second vector. For example, the second vector can be extracted from the starting position of the first vector, and the third vector can be extracted from the middle position.

[0030] It should be noted that the second and third vectors can be determined based on the security level of the first vector. The security level can be determined based on the query text obtained from the browser. For example, after obtaining the query text from the browser, the Euclidean distance between the query text and several preset confidential texts can be calculated, and the security level of the first vector can be determined by classifying the range of Euclidean distances it falls within.

[0031] In practice, a target dimensionality reduction algorithm can be selected from a pre-configured set of algorithms based on the security level of the first vector, and then used to process the first vector. For example, assuming the pre-configured dimensionality reduction algorithms include continuous subvector extraction (corresponding to a confidential security level), random projection (corresponding to a top-secret security level), and hash-based feature selection (corresponding to a secret security level), if the security level of the input query text is top-secret, then random projection will be used as the target dimensionality reduction algorithm.

[0032] In practice, a time-rotation mechanism can be used to select the target dimensionality reduction algorithm from a pre-configured set of algorithms. For example, if multiple dimensionality reduction algorithms are pre-configured, and the runtime of the current target dimensionality reduction algorithm exceeds a preset time threshold, a dimensionality reduction algorithm switching process is executed to update the target dimensionality reduction algorithm. As another example, if multiple dimensionality reduction algorithms correspond to different security levels, and the runtime of the dimensionality reduction algorithm selected according to the security level of the input query text exceeds a preset time threshold, the next dimensionality reduction algorithm in the group corresponding to the security level of the input query text is switched as the target dimensionality reduction algorithm.

[0033] Step S104: Generate a first request message based on the second vector and the third vector.

[0034] It should be noted that the metadata of the first request message generally includes relevant parameters of the second and third vectors, such as the number of vector dimensions, the starting position of the second vector, and the starting position of the third vector.

[0035] Step S105: Send the first request message to the server so that the server obtains the query result based on the first request message and then sends the query result to the browser.

[0036] It should be noted that the server stores classified text vectors. After parsing the second and third vectors, the server can calculate the similarity with the classified text vectors. Based on the text corresponding to the classified text vectors with similarity greater than the similarity threshold, the query results can be determined.

[0037] In this embodiment, a first vector is obtained on the browser side through a first pre-trained model. The first vector is then processed using different dimensionality reduction algorithms to obtain a second vector and a third vector. Based on the second and third vectors, a first request message is generated and sent to the server. The server then obtains the query results based on the first request message and sends the query results back to the browser. Since the first pre-trained model used is based on a technology in the field of classified information, the first vector generated in this way has higher confidentiality compared to vectors generated by general models. This increases the difficulty for attackers to crack the first vector, thus initially enhancing the security of the first vector. By generating two vectors from the first vector through dimensionality reduction processing to generate the request message, the sensitive information in the original query is reduced, avoiding the leakage problem caused by an attacker intercepting a vector in the message. Furthermore, it enables retrieval from different semantic features, thereby improving the retrieval accuracy on the server side.

[0038] According to one embodiment of this application, the process of obtaining the second vector includes: reducing the dimensionality of the first vector using a first matrix to obtain a second vector with reduced dimensionality; The process of obtaining the third vector includes: using the second matrix to reduce the dimensionality of the first vector to obtain the third vector with reduced dimensionality.

[0039] It should be noted that the vector space of the first pre-trained model is high-dimensional (e.g., 768-dimensional), generated using a corpus containing a small number of sensitive words. Therefore, its vector space points to different semantic features in different dimensions or ranges of dimensions. For example, dimensions 0-30 are time-related, dimensions 31-70 are law-related, and dimensions 71-110 are technology-related. Therefore, the second and third vectors can be selected based on the semantic features of the vector space of the first pre-trained model. For example, the first vector can be dimensionality-reduced to extract the second vector containing law-related information, and the first vector can be dimensionality-reduced to extract the third vector containing technology-related information.

[0040] In practical applications, the dimensionality reduction method for the first vector can be determined based on the query text obtained from the browser. For example, the distance between the query text obtained from the browser and different semantic features can be calculated, the distance values ​​can be sorted, and the first vector can be dimensionality reduced based on the two semantic features with the largest distance values. For example, if the semantic feature with the largest distance value is time-related or technology-related, then it can be processed according to the dimension or dimensional range of the corresponding semantic feature in the first vector.

[0041] According to one embodiment of this application, noise is added to the second vector to change the value of a second percentage within the second vector, where the second percentage is 5% to 15%. Noise is added to the second vector to change the value of the second percentage within the second vector, which is between 5% and 15%.

[0042] It should be noted that in this embodiment, the second percentage is used to represent the ratio of the modified dimension number in the second vector. The purpose of adding noise is to change the numerical values ​​included in the vector, so that even if an attacker captures the noisy vector, it will be difficult to reconstruct the original vector, which can reduce the probability of leakage incidents during the retrieval of classified text; at the same time, it ensures the accuracy of subsequent retrieval results.

[0043] In this embodiment, the second percentage is 5% to 15%, but in actual applications, it can be randomly given by the browser side according to 5% to 15%.

[0044] In practice, after calculating the number of dimensions to be changed based on the second percentage and the number of dimensions of the second vector, the location for adding noise is randomly determined, and then the preset noise addition algorithm is used for processing.

[0045] In practical applications, noise is added by setting the value at the location where noise is to be added to zero or other random values.

[0046] In practical applications, noise is added by reversing the direction of the value at the location where noise is added. For example, if the original value is -0.823, the value is changed to 0.823.

[0047] In practical applications, noise is added by multiplying the value at the location where noise is added with a random factor to obtain the modified value.

[0048] It should be noted that the above method of adding noise is an example. In actual applications, other methods of adding noise can also be used, which will not be described here.

[0049] It should be noted that in some embodiments, multiple noise-adding methods can be used simultaneously to add noise to the second vector. For example, assuming the second vector has a dimension of 300 and the second percentage is 15%, that is, 45 positions are randomly selected in the second vector for noise addition. In this case, some of the values ​​at these 45 positions can be set to zero or other random values, and the values ​​of the other part can be multiplied by a random factor. The use of multiple noise-adding methods further increases the difficulty of restoring the original vector, thereby improving the confidentiality of text retrieval.

[0050] According to one embodiment of this application, before acquiring the second or third vector, noise is added to the first vector to change the value of a first percentage within the first vector, where the first percentage is 5% to 15%.

[0051] It should be noted that the process of adding noise to the first vector can be similar to the process of adding noise to the second vector described above. In this way, the vector that has been reduced in dimensionality changes. Even if an attacker captures the second and third vectors at the same time, they will still not be able to reconstruct the original vector. This serves as a double protection for the original query by first-level encryption of the original text vector and then adding noise to the second and third vectors after dimensionality reduction. This increases the difficulty for attackers to reconstruct the original query and further improves the security of text retrieval.

[0052] According to one embodiment of this application, the ratio of the second or third vector dimension to the first vector dimension is 0.2:1 to 0.3:1.

[0053] It should be noted that if the ratio of the second or third vector dimension to the first vector dimension is too low, it will affect the accuracy of classified text retrieval; if the ratio is too high, it will be easy for attackers to reconstruct the original query. Therefore, this embodiment sets the ratio of the second or third vector dimension to the first vector dimension to be between 0.2:1 and 0.3:1. This ensures that the retrieval accuracy on the server side is not affected, while also preventing attackers from cracking the original query.

[0054] In practical applications, the ratio of the dimension of the second or third vector to the dimension of the first vector can be randomly provided by the browser. For example, assuming the dimension of the first vector is 1000 and the ratio provided by the browser is 2:1, then the dimension of the second or third vector is 200, which means that the first vector is dimensionalized to obtain a second and third vector with 200 dimensions.

[0055] According to one embodiment of this application, the process of obtaining query results includes: The first request message is used to obtain a set of vectors for retrieval. Based on the vectors in the set, the first and second vector data tables are searched to obtain the first and second retrieval results. The query results are obtained based on the first and second search results; The matching records in the first and second search results include document identifiers, similarity scores, and document vectors.

[0056] It should be noted that the first request message is generated by the browser based on the second and third vectors. Therefore, the vector set obtained by the server based on the first request message specifically includes the second and third vectors. The server pre-stores a first vector data table and a second vector data table. The first vector data table pre-stores several second vectors corresponding to different classified texts, and the second vector data table pre-stores several third vectors corresponding to different classified texts. In application, the first pre-trained model can be used to represent the classified text using different vectors, thereby obtaining the first vector data table and the second vector data table.

[0057] In practice, the first and second search results are obtained by calculating the approximation degree of each vector in the vector set with any vector in the first vector data table and the approximation degree of any vector in the second vector data table. In application, if the number of texts in the first or second search results is too large (e.g., exceeding a preset text threshold), the top predetermined number of texts with the highest approximation (e.g., 5) can be used as search results; if the number of texts in the first or second search results is too small (e.g., not exceeding the preset text threshold), all texts can be used as search results.

[0058] According to one embodiment of this application, the process of obtaining query results includes: The first result set in the first search result and the second result set in the second search result are determined based on the document identifiers shared by the first search result and the second search result, respectively. The fourth vector is determined based on the first result set and the first search result; The fifth vector is determined based on the second result set and the second search result; Based on the fourth vector, a third search result is obtained by searching within the first vector data table if the similarity to the fourth vector is greater than the first threshold. Based on the fifth vector, a fourth search result is obtained from the first vector data table if the similarity to the fourth vector is greater than the first threshold. The query results are obtained by merging the third and fourth search results.

[0059] It should be noted that the fourth and fifth vectors are obtained based on a preset vector generation strategy.

[0060] Specifically, the fourth and fifth vectors can be determined according to a preset vector generation strategy. In practical applications, taking the first result set as an example, the vector generation strategy can be as follows: if the number of documents in the first result set is 0, the fourth vector is determined based on the document vectors ranked N (e.g., 3) in the first search results for similarity, or the fourth vector is determined by the average value of the document vectors in the first search results; if the number of documents in the first result set is not zero, the average value of the vectors in the first result set and the first search results can be calculated as the fourth vector, or a weighted sum can be calculated for the vectors in the first result set and the first search results, and the calculated weighted sum is used as the fourth vector.

[0061] According to one embodiment of this application, the search results also include documents corresponding to document identifiers common to both the first and second search results.

[0062] Specifically, the final search result can be determined based on the number of document identifiers shared by the first and second search results. For example, when the number of document identifiers shared by the first and second search results is greater than a preset threshold, the document pointed to by the document identifier with the highest similarity between the first and second search results is determined as the search result; when the number of document identifiers shared by the first and second search results is not greater than the preset threshold, all documents corresponding to the document identifiers shared by the first and second search results are determined as the search result.

[0063] One embodiment of this application provides a retrieval device for classified standard texts on an intranet, such as... Figure 2 As shown, the device 20 includes: a query vector determination module 201, a first dimensionality reduction processing module 202, a second dimensionality reduction processing module 203, a request message generation module 204, and a request message processing module 205.

[0064] The query vector determination module 201 is used to obtain the first vector on the browser side through the first pre-trained model; The first dimension reduction processing module 202 is used to perform a first dimension reduction on the first vector to obtain a second vector; The second dimension reduction processing module 203 is used to perform a second dimension reduction on the first vector to obtain a third vector; The request message generation module 204 is used to generate a first request message based on the second vector and the third vector; The request message processing module 205 is used to send the first request message to the server so that the server obtains the query result based on the first request message and then sends the query result to the browser.

[0065] In this embodiment, a first vector is obtained on the browser side through a first pre-trained model. The first vector is then processed using different dimensionality reduction algorithms to obtain a second vector and a third vector. Based on the second and third vectors, a first request message is generated and sent to the server. The server then obtains the query results based on the first request message and sends the query results back to the browser. Since the first pre-trained model used is based on a technology in the field of classified information, the first vector generated in this way has higher confidentiality compared to vectors generated by general models. This increases the difficulty for attackers to crack the first vector, thus initially enhancing the security of the first vector. By generating two vectors from the first vector through dimensionality reduction processing to generate the request message, the sensitive information in the original query is reduced, avoiding the leakage problem caused by an attacker intercepting a vector in the message. Furthermore, it enables retrieval from different semantic features, thereby improving the retrieval accuracy on the server side.

[0066] Furthermore, the process of obtaining the second vector includes: using the first matrix to reduce the dimensionality of the first vector to obtain a second vector with reduced dimensionality; The process of obtaining the third vector includes: using the second matrix to reduce the dimensionality of the first vector to obtain the third vector with reduced dimensionality.

[0067] Furthermore, noise is added to the second vector to change the value of the second percentage within the second vector, which is 5% to 15%. Noise is added to the second vector to change the value of the second percentage within the second vector, which is between 5% and 15%.

[0068] Furthermore, before acquiring the second or third vector, noise is added to the first vector to change the value of a first percentage within the first vector, which is 5% to 15%.

[0069] Furthermore, the ratio of the second or third vector dimension to the first vector dimension is 0.2:1 to 0.3:1.

[0070] Furthermore, the process of obtaining query results includes: The first request message is used to obtain a set of vectors for retrieval. Based on the vectors in the set, the first and second vector data tables are searched to obtain the first and second retrieval results. The query results are obtained based on the first and second search results; The matching records in the first and second search results include document identifiers, similarity scores, and document vectors.

[0071] Furthermore, the process of obtaining query results includes: The first result set in the first search result and the second result set in the second search result are determined based on the document identifiers shared by the first search result and the second search result, respectively. The fourth vector is determined based on the first result set and the first search result; The fifth vector is determined based on the second result set and the second search result; Based on the fourth vector, a third search result is obtained by searching within the first vector data table if the similarity to the fourth vector is greater than the first threshold. Based on the fifth vector, a fourth search result is obtained from the first vector data table if the similarity to the fourth vector is greater than the first threshold. The query results are obtained by merging the third and fourth search results.

[0072] Furthermore, the search results also include documents corresponding to document identifiers shared by the first and second search results.

[0073] The retrieval device for classified standard texts on the intranet in this embodiment can execute the retrieval method for classified standard texts on the intranet shown in the embodiment of this application. The implementation principle is similar, and will not be described again here.

[0074] Another embodiment of this application provides a computer device including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0075] Specifically, the processor can be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc.

[0076] Specifically, the processor connects to the memory via a bus, which may include a path for transmitting information. The bus can be a PCI bus or an EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc.

[0077] The memory may be ROM or other types of static storage devices that can store static information and instructions, RAM or other types of dynamic storage devices that can store information and instructions, or EEPROM, CD-ROM or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer, but is not limited thereto.

[0078] Optionally, the memory stores the code of a computer program that executes the scheme of this application, and the execution is controlled by a processor. The processor executes the application code stored in the memory to implement the operation of the apparatus provided in the above embodiments.

[0079] Another embodiment of this application provides a computer-readable storage medium storing computer-executable instructions for performing the methods provided in the above embodiments.

[0080] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0081] It will be understood by those skilled in the art that all or some of the steps and systems in the methods disclosed above can be implemented as software, firmware, hardware, and suitable combinations thereof. Some or all of the physical components can be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software can be distributed on a computer-readable medium, which can include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0082] The above is a detailed description of the preferred embodiments of this application. However, this application is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A method for retrieving classified standard texts on an intranet, characterized in that, include: The first vector is obtained on the browser side using the first pre-trained model; The first dimension reduction is performed on the first vector to obtain the second vector; The third vector is obtained by performing a second dimensionality reduction on the first vector. The first request message is generated based on the second and third vectors; The first request message is sent to the server so that the server obtains the query result based on the first request message and then sends the query result to the browser.

2. The method for retrieving classified standard texts on an intranet as described in claim 1, characterized in that, The process of obtaining the second vector includes: using the first matrix to reduce the dimensionality of the first vector to obtain a second vector with reduced dimensionality; The process of obtaining the third vector includes: using the second matrix to reduce the dimensionality of the first vector to obtain the third vector with reduced dimensionality.

3. The method for retrieving classified standard texts on an intranet as described in claim 2, characterized in that, Noise is added to the second vector to change the value of a second percentage within the second vector, where the second percentage is 5% to 15%. Noise is added to the second vector to change the value of a second percentage within the second vector, which is 5% to 15%.

4. The method for retrieving classified standard texts on an intranet as described in claim 1, characterized in that, Before obtaining the second or third vector, noise is added to the first vector to change the value of a first percentage within the first vector, where the first percentage is 5% to 15%.

5. The method for retrieving classified standard texts on an intranet as described in claim 1, characterized in that, The ratio of the second or third vector dimension to the first vector dimension is 0.2:1 to 0.3:

1.

6. The method for retrieving classified standard texts on an intranet as described in claim 1, characterized in that, The process of obtaining the query results includes: The first request message is used to obtain a set of vectors for retrieval. Based on the vectors in the set, the first and second vector data tables are searched to obtain the first and second retrieval results. The query results are obtained based on the first and second search results; The hit records in the first and second search results include document identifiers, similarity scores, and document vectors.

7. The method for retrieving classified standard texts on an intranet as described in claim 6, characterized in that, The process of obtaining the query results includes: The first result set in the first search result and the second result set in the second search result are determined based on the document identifiers shared by the first search result and the second search result, respectively. The fourth vector is determined based on the first result set and the first search result; The fifth vector is determined based on the second result set and the second search result; Based on the fourth vector, a third search result is obtained by searching within the first vector data table if the similarity to the fourth vector is greater than the first threshold. Based on the fifth vector, a fourth search result is obtained from the first vector data table if the similarity to the fourth vector is greater than the first threshold. The query results are obtained by merging the third and fourth search results.

8. The method for retrieving classified standard texts on an intranet as described in claim 7, characterized in that, The search results also include documents corresponding to document identifiers shared by the first and second search results.

9. An electronic device, characterized in that, The method includes a memory and a processor, the memory being coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the method according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-8.

Images