A corpus analysis method and a corpus analysis model training method and device

Abstract

The application provides a corpus analysis method and device, and relates to the technical field of artificial intelligence. The application aims to improve the corpus analysis accuracy and scene adaptability of a localized corpus analysis model through a dynamic scene adaptation and supplementary information fusion mechanism, so as to obtain an accurate corpus analysis result. The method specifically comprises: analyzing input corpus information to obtain corresponding semantic information; matching a preset application scene according to the semantic information, wherein different application scenes correspond to different supplementary information; collecting corresponding supplementary information according to the application scene; combining the corpus information and the supplementary information into target corpus information that meets the application scene; inputting the target corpus information into a localized corpus analysis model, and outputting a corpus analysis result corresponding to the corpus information.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a corpus analysis method, a corpus analysis model training method and apparatus. Background Technology

[0002] With the development of artificial intelligence technology, generative large models have been widely used in the field of corpus analysis. Traditional generative large models require uploading corpus information to a cloud server for processing and feedback, which can easily lead to the leakage of corpus information. To prevent the leakage of corpus information, current methods utilize local deployment technology to deploy generative large models to local devices for corpus analysis.

[0003] However, local deployment of generative large models has introduced new technical challenges. This is because the performance of generative large models relies on a large number of parameters; more parameters mean the model can learn more complex language patterns and semantic relationships, resulting in more accurate analysis results. However, local devices (such as personal computers and mobile devices) have limited hardware resources, including memory capacity and computing power, making it difficult to support the original generative large models running on cloud servers. To adapt to the local environment, the number of parameters in the generative large model must be reduced. This reduction in parameters makes the thought process of the locally deployed generative large model less refined, leading to inaccurate output results.

[0004] Therefore, a new method is urgently needed to obtain accurate corpus analysis results. Summary of the Invention

[0005] This application provides a corpus analysis method, a corpus analysis model training method, and an apparatus. The purpose is to improve the corpus analysis accuracy of the localized corpus analysis model through a dynamic scene adaptation and supplementary information fusion mechanism, so as to obtain accurate corpus analysis results.

[0006] To address the aforementioned technical problems, this application provides the following technical solutions:

[0007] Firstly, this application provides a corpus analysis method, the method comprising:

[0008] The input corpus information is analyzed to obtain the corresponding semantic information;

[0009] The semantic information is used to match a preset application scenario, where different application scenarios correspond to different supplementary information;

[0010] Collect corresponding supplementary information based on the application scenario;

[0011] The corpus information is combined with the supplementary information to form target corpus information that conforms to the application scenario;

[0012] The target corpus information is input into a locally deployed corpus analysis model, which outputs the corpus analysis results corresponding to the corpus information.

[0013] Secondly, this application provides a corpus analysis apparatus, the apparatus comprising:

[0014] The unit is used to analyze the input corpus information and obtain the corresponding semantic information;

[0015] A matching unit is used to match a preset application scenario based on the obtained unit semantic information, wherein different application scenarios correspond to different supplementary information;

[0016] The acquisition unit is used to acquire corresponding supplementary information according to the application scenario of the matching unit;

[0017] A combination unit is used to combine the corpus information with the supplementary information from the acquisition unit into target corpus information that conforms to the application scenario.

[0018] The output result unit is used to input the target corpus information of the combined unit into the locally deployed corpus analysis model and output the corpus analysis result corresponding to the corpus information.

[0019] Thirdly, this application provides a corpus analysis model training method, the method comprising: acquiring target training corpus information, wherein the target training corpus information is determined based on initial training corpus information and training supplementary information corresponding to the initial training corpus information, and the training supplementary information is obtained based on the application scenario matched by the initial training corpus information;

[0020] The target training corpus information is vectorized to obtain a training vector matrix of the target training corpus information. The elements of the training vector matrix are used to represent the position information of each word in the target training corpus information.

[0021] Obtain a training weight matrix, the elements of which are used to characterize the semantic influence between row-dimensional elements and column-dimensional elements in the training vector matrix;

[0022] Based on the number of dimensions to be decomposed, the training weight matrix is ​​decomposed to obtain a low-rank matrix after decomposition. The number of dimensions to be decomposed represents the number of lexical units to be analyzed for semantic influence.

[0023] Based on the decomposed low-rank matrix, the change in the weight matrix is ​​obtained;

[0024] Based on the changes in the weight matrix and the training weight matrix, the final training weight matrix is ​​determined, and the corpus analysis model is determined according to the final training weight matrix.

[0025] Fourthly, this application provides a corpus analysis model training device, the device comprising:

[0026] An information acquisition unit is used to acquire target training corpus information, wherein the target training corpus information is determined based on initial training corpus information and training supplementary information corresponding to the initial training corpus information, and the training supplementary information is obtained based on the application scenario matched by the initial training corpus information;

[0027] The information acquisition unit is used to vectorize the target training corpus information to obtain a training vector matrix of the target training corpus information, wherein the elements of the training vector matrix are used to characterize the position information of each word in the target training corpus information.

[0028] The information acquisition unit is used to acquire a training weight matrix, wherein the elements of the training weight matrix are used to characterize the semantic influence between the row dimension elements and the column dimension elements in the training vector matrix.

[0029] The matrix unit is used to decompose the training weight matrix in the information acquisition unit based on the number of dimensions to be decomposed, and to obtain the decomposed low-rank matrix. The number of dimensions to be decomposed represents the number of lexical units to be analyzed for semantic influence.

[0030] The matrix acquisition unit is used to obtain the change in the weight matrix based on the decomposed low-rank matrix.

[0031] A model unit is defined to determine the final training weight matrix based on the change in the weight matrix in the acquisition matrix unit and the training weight matrix in the acquisition matrix unit, and to determine the corpus analysis model based on the final training weight matrix.

[0032] Fifthly, this application provides a computing device, the computing device comprising: at least one processor, and a memory, wherein the memory stores instructions executable by the processor, the instructions being executed by the processor, thereby enabling the processor to perform the above-described corpus analysis method or the above-described corpus analysis model training method.

[0033] Sixthly, this application provides a readable storage medium for storing a computer program, wherein the computer program, when running, controls the device where the storage medium is located to execute the corpus analysis method or the corpus analysis model training method described above.

[0034] Compared to existing technologies, this application provides a corpus analysis method, a corpus analysis model training method, and an apparatus. This method improves the accuracy and scene adaptability of locally deployed corpus analysis models through dynamic scene adaptation and supplementary information fusion mechanisms, thereby obtaining accurate corpus analysis results. Specifically, this scheme analyzes the input corpus information to obtain semantic information, then matches the semantic information with a preset application scenario and collects supplementary information corresponding to that scenario. This supplementary information provides more semantic details to the locally deployed corpus analysis model. When the corpus information and supplementary information are combined to form target corpus information and input into the locally deployed corpus analysis model, the model can perform analysis based on a richer information foundation, effectively avoiding semantic understanding bias caused by a reduction in the number of parameters, and thus outputting accurate corpus analysis results. Attached Figure Description

[0035] The above and other objects, features, and advantages of exemplary embodiments of this application will become readily understood by reading the following detailed description with reference to the accompanying drawings. In the drawings, several embodiments of this application are illustrated by way of example and not limitation, with the same or corresponding reference numerals denoteing the same or corresponding parts, wherein:

[0036] Figure 1 A flowchart illustrating a corpus analysis method proposed in an embodiment of this application is shown schematically.

[0037] Figure 2 A flowchart illustrating a corpus analysis method proposed in an embodiment of this application is shown schematically.

[0038] Figure 3 A schematic diagram illustrating a weight matrix decomposition proposed in an embodiment of this application is shown.

[0039] Figure 4 The schematic diagram illustrates the results of a corpus analysis device proposed in an embodiment of this application.

[0040] Figure 5 A schematic diagram illustrating the results of another corpus analysis device proposed in an embodiment of this application is shown.

[0041] Figure 6 A schematic diagram of a corpus analysis model training device proposed in an embodiment of this application is shown. Detailed Implementation

[0042] Exemplary embodiments of this application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of this application and to fully convey the scope of this application to those skilled in the art.

[0043] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains.

[0044] Currently, local deployment technology is used to deploy generative large models to local devices for corpus analysis. However, local deployment of generative large models has introduced new technical challenges. This is because the performance of generative large models depends on a large number of parameters; more parameters mean that the generative large model can learn more complex language patterns and semantic relationships, thus outputting accurate analysis results. However, local devices (such as personal computers and mobile terminals) have limited hardware resources such as memory capacity and computing power, making it difficult to support the original generative large models running on cloud servers. To adapt to the local environment, it is necessary to reduce the number of parameters in the generative large model. This reduction in the number of parameters makes the thought process of the locally deployed generative large model less refined, leading to inaccurate output results.

[0045] To address this issue, the inventors of this application propose a corpus analysis method. This method analyzes input corpus information to obtain corresponding semantic information; matches the semantic information with preset application scenarios, where different application scenarios correspond to different supplementary information; collects corresponding supplementary information based on the application scenarios; combines the corpus information and the supplementary information to form target corpus information conforming to the application scenario; and inputs the target corpus information into the corpus analysis model, outputting the corpus analysis results corresponding to the corpus information. This solution improves the corpus analysis accuracy of the localized corpus analysis model through dynamic scenario adaptation and supplementary information fusion mechanisms, thereby obtaining accurate corpus analysis results.

[0046] like Figure 1 As shown, an embodiment of this application provides a corpus analysis method, including:

[0047] Step 101: Analyze the input corpus information to obtain the corresponding semantic information.

[0048] In this step, the corpus information can be the original user request information in the form of text, speech, images, or a combination thereof. For example, the corpus information can be text content, such as "Someone T wants to travel from point A to point B on June 6th, XX year, and needs to depart in the morning. Please plan the route from A to B." It can also be the speech expression of the text content. Analyzing the input corpus information refers to extracting semantic information from the text using natural language processing algorithms. This can be implemented using a Transformer-based encoder to identify the subject's actions, object attributes, and logical relationships in the sentences. Semantic information refers to the information about the meaning extracted from the corpus information after analysis using natural language processing techniques. Natural language processing techniques can include lexical analysis of the corpus information, breaking it down into multiple words, then syntactic analysis to determine the sentence structure, clarifying the grammatical relationships between words, and finally using semantic understanding algorithms to derive the semantic information. For example, if the corpus information is "I want to go to the beach for tourism, are there any recommended places?", the corpus analysis model deployed locally on the mobile phone uses natural language processing technology to first perform lexical analysis, breaking the sentence down into words such as "I", "want to", "beach tourism", "are there", "recommend", and "place". Then, it uses syntactic analysis to determine the sentence structure and clarify the grammatical relationships between the words. Finally, it uses semantic understanding algorithms to derive the semantic information "the user expressed a need to go to the beach for tourism and seek recommended places".

[0049] Step 102: Match the preset application scenario based on the semantic information.

[0050] In this step, different application scenarios in the preset application scenarios correspond to different supplementary information. Supplementary information collection refers to retrieving auxiliary data related to the current scenario from the local knowledge base. Auxiliary data is data that is adapted to the current scenario and can enhance scenario understanding, such as product sales and user reviews in shopping recommendation scenarios; real-time visitor flow and weather at tourist attractions in travel planning scenarios; and authoritative information and historical cases in knowledge question-and-answer scenarios. Supplementary information provides further data support for functions such as accurate semantic understanding and providing adaptation services. The preset application scenarios are context categories with preset keyword information and needs that are pre-set before corpus analysis. They are used to classify the input semantic information so that targeted supplementary information can be obtained and more accurate corpus analysis can be performed later. These can be shopping recommendation scenarios, travel planning scenarios, knowledge question-and-answer scenarios, and life service scenarios. Knowledge question-and-answer scenarios can also include specific scenarios such as information proofreading, email reply, and information summarization, which are not further limited here.

[0051] After obtaining semantic information in step 101, the semantic information can be matched with preset application scenarios to obtain the matching degree of the preset application scenarios. The application scenario corresponding to the highest matching degree of the preset application scenarios is taken as the application scenario corresponding to the semantic information. Specifically, this involves: extracting keyword information from the semantic information; matching the keyword information with preset keyword information corresponding to different application scenarios to obtain the keyword information matching degree under different application scenarios; determining the matching degree of different scenarios based on the keyword information matching degree under different application scenarios, and determining the application scenario corresponding to the maximum matching degree of the different scenarios. By extracting keyword information from the semantic information, matching it with preset keyword information, and selecting the maximum matching degree, the application scenario is determined. This method avoids misjudgment of application scenarios by quantifying the keyword information matching degree, ensuring that the most accurate supplementary information is matched for the corpus, thereby laying a correct foundation for subsequent corpus combination and model analysis. It effectively solves the problem of invalid supplementary information due to scene matching errors, which in turn affects the accuracy of localized model analysis. It is worth noting that the extraction of keyword information from semantic information employs named entity recognition technology (which can accurately identify specific entity nouns as keyword information) and keyword extraction algorithms (such as rule-based extraction and machine learning-based extraction, with models like Naive Bayes and Support Vector Machines used for keyword classification and extraction). These techniques enhance the accuracy and comprehensiveness of keyword information extraction, thereby providing more effective support for subsequent application scenario matching. Keyword matching utilizes either edit distance-based (e.g., Levenshtein distance) or semantic vector space-based (e.g., cosine similarity calculation) matching methods. The maximum matching degree for different scenarios can be the sum of the matching degrees of all keywords in the semantic information and the preset keyword.

[0052] Step 103: Collect corresponding supplementary information according to the application scenario.

[0053] In this step, after obtaining the application scenario in step 102, the preset keyword information in the application scenario is compared with the keyword information in the semantic information. Keyword information in the preset keyword information that does not match the keyword information in the semantic information is identified. This identified keyword information is designated as supplementary information, while the keyword information in the semantic information is considered basic information. For example, if the specific features preset in the application scenario are A, B, C, and G, and the keyword information in the semantic information is C and G, then A and B are supplementary information, and C and G are basic information.

[0054] Depending on the application scenario, the supplementary information collected can be obtained through keyword retrieval or by searching through category directories.

[0055] In some embodiments, the keyword-based data collection method specifically involves: acquiring preset key information corresponding to the application scenario, where the preset keyword information is pre-defined; comparing this key information with the keyword information in the corpus information to determine supplementary information in the preset keyword information other than the keyword information in the corpus information; and searching for data matching the supplementary information in a local information database using keyword retrieval. The local information database can be a database composed of information from multiple local mobile device applications. For example, in a clothing Q&A scenario, the preset keyword information includes: time, location, temperature, and clothing type. If the user inputs the semantic information: "What type of clothing should be worn in location A on a certain day?", the obtained semantic information will contain the keyword information: "Location (location A), Time (a certain day), Clothing type". After obtaining the keyword information, it is compared with the preset keyword information in the clothing Q&A scenario, i.e., "Location (location A), Time (a certain day), Clothing type" is compared with "Time, Location, Temperature, Clothing type", and temperature is determined as supplementary information. After acquiring and supplementing the information, data matching the temperature information is retrieved from the local database using keyword searches and used as supplementary data for the temperature information.

[0056] In other embodiments, the local information database also includes a category directory, which includes at least one layer of application scenario classification information and corresponding supplementary information. The category directory can be a multi-layered storage structure, with different fields stored in each layer. Specifically, after determining the application scenario, the supplementary information is determined by searching the corresponding category directory based on the application scenario. For example, an existing category directory stores a scenario called "Mobile Phone Performance Inquiry." Under this application scenario, there is a first-layer storage structure such as "Brand A," "Brand B," etc. Under the first-layer storage structure "Brand A," there is a second-layer storage structure such as "Model 1 of A," "Model 2 of A," etc. Under the second-layer storage structure "Model 1 of A," information such as "User Reviews" and "Model Parameters" is stored. Similarly, under the first-layer storage structure "Brand B," there is a second-layer storage structure such as "Model 1 of B," "Model 2 of B," etc. Under the second-layer storage structure "Model 1 of B," information such as "User Reviews" and "Model Parameters" is stored. The existing semantic information is about the performance of brand A mobile phones. The application scenario for this semantic information is the purchase scenario of brand A model 1 mobile phone. Based on this purchase scenario, the category directory is searched to obtain the first-level storage structure: "Brand A" - "Model 1 of A" - "User Reviews, Model Parameters". At this point, the supplementary information "User Reviews, Model Parameters" can be obtained.

[0057] Step 104: Combine the corpus information with the supplementary information to form target corpus information that conforms to the application scenario.

[0058] In this step, to ensure the accuracy of the target corpus information, the target corpus information combination refers to the structured integration of corpus information and supplementary information according to a predefined scenario template. The field positions and formats of the predefined scenario template can be defined using XML markup language, or a predefined scenario template in JSON format can be used to form enhanced data that conforms to the model input specifications. The predefined scenario template defines the field mapping relationship and combination order between the corpus information and supplementary information.

[0059] In this embodiment, the key content required by the corpus information and supplementary information in different application scenarios is defined in advance, and corresponding fields are set for these contents. The arrangement order of each field in the target corpus information is also planned in detail, and a corresponding predefined scenario template is constructed. When processing the actual corpus information and supplementary information, this information is first parsed to filter out the key information specified by the template. Then, according to the requirements of the predefined scenario template, the extracted key information is embedded into the corresponding XML tags, and a complete XML structure document is constructed in a predetermined order.

[0060] During the construction process, it is necessary to repeatedly check whether each piece of information accurately matches the corresponding field and whether the order is correct. For example, with the help of automated verification tools, field matching rules and order sequences are pre-configured for each scenario template, and a rule verification library is built. After extracting the key information XML, the tool calls the verification library, uses regular expression matching to verify the compliance of field values, and uses sequence comparison algorithms to check the arrangement order, automatically determines whether the match is accurate, and outputs a verification report.

[0061] After the initial construction is completed, the generated XML document is thoroughly reviewed again to check for proper syntax, missing information, or incorrect embedding, ensuring it fully conforms to the model input specifications. If any issues are found, they are promptly modified and improved until the requirements are met.

[0062] An example of the above content is as follows: In a movie recommendation scenario, for example, a predefined scenario template includes fields such as "user preference type", "movie-watching time", and "number of movie-watchers". Suppose the corpus information is "I want to watch a science fiction movie with friends on the weekend", and the supplementary information is "science fiction movie session information". From the corpus information, the "user preference type" is extracted as "science fiction", the "movie-watching time" is "weekend", and although the "number of movie-watchers" is not explicitly stated, it can be set as "multiple people". An XML document is constructed according to the template: <Movie Recommendation><User Preference Type>Science Fiction< / User Preference Type><Movie-Watching Time>Weekend< / Movie-Watching Time><Number of Movie-Watchers>Multiple People< / Number of Movie-Watchers>< / Movie Recommendation>. After checking and ensuring it is correct, the target corpus information that meets the requirements is obtained.

[0063] Step 105: Input the target corpus information into a corpus analysis model deployed locally, and output the corpus analysis result corresponding to the corpus information.

[0064] In this step, the corpus analysis model can be implemented by training a trimmed language model on a local device. It can directly analyze the corpus information on the terminal device and give the corresponding response information (corpus analysis result) without relying on the cloud or network. That is, the corpus analysis model has a data model for processing data offline and providing feedback.

[0065] After obtaining the target corpus information in step 104, it is input into the corpus analysis model. Multiple calculation models in the corpus analysis model respectively obtain the target corpus information. Each calculation model outputs a corpus analysis result based on the target corpus information, calculates the matching degree between each corpus analysis result and the target corpus analysis, and takes the result output by the calculation model with a high matching degree as the corpus analysis result. Among them, the corpus analysis model includes multiple calculation models. The calculation model uses the Lora algorithm for calculation. The calculation model can be a summary Lora calculation model, a proofreading Lora calculation model, an email reply Lora calculation model, etc. The summary Lora calculation model is used to focus on the induction, refinement, and summary of the corpus content. That is, it uses the Lora algorithm to capture key information and aggregate core viewpoints from the input corpus, and quickly generates a concise summary output that covers the main content, adapting to scenarios that require extracting key points and summarizing conclusions, such as document summarization and meeting minutes extraction. The proofreading Lora calculation model is used to standardize the grammar of the corpus based on the Lora technology. The email reply Lora calculation model is used for the email communication scenario. It analyzes the email subject and text intention using the Lora algorithm, and combines the preset reply style (such as formal, concise) and common phrases to automatically generate reply content that fits the email communication context and meets the communication requirements.

[0066] This application also proposes a specific training method for a corpus analysis model, as follows: Obtaining target training corpus information, which is determined based on initial training corpus information and corresponding supplementary training information, wherein the supplementary training information is obtained based on application scenarios matched with the initial training corpus information; vectorizing the target training corpus information to obtain a training vector matrix, wherein the elements of the training vector matrix represent the positional information of each word in the target training corpus information; obtaining a training weight matrix, wherein the elements of the training weight matrix represent the semantic influence between row and column elements in the training vector matrix; decomposing the training weight matrix based on the number of dimensions to be decomposed to obtain a low-rank matrix after decomposition, wherein the number of dimensions to be decomposed represents the number of words to be analyzed for semantic influence; obtaining the weight matrix change based on the low-rank matrix after decomposition; determining the final training weight matrix based on the weight matrix change and the training weight matrix, and determining the corpus analysis model based on the final training weight matrix.

[0067] It is worth noting that when there are multiple computational models, each model is trained using the above training method to obtain an accurate corpus analysis model that applies to each computational model. For example, when the computational model is a summary LoRa computational model: the target training corpus information in the summary scenario is trained using the above model training method; when the computational model is a proofreading LoRa computational model: the target training corpus information in the proofreading scenario is obtained and trained using the above model training method; when the computational model is an email LoRa computational model: the target training corpus information in the email reply scenario is obtained and trained using the above model training method to obtain the corresponding corpus analysis model.

[0068] The training process of the corpus analysis model provided in this application embodiment is to construct an intelligent model with dynamic weight adjustment capabilities by integrating contextual information and semantic association analysis. Specifically, firstly, supplementary training information is obtained based on the initial training corpus and its matched application scenarios. The two are combined to form target training corpus information to ensure that the corpus covers the vocabulary and semantic logic specific to the scenario. Subsequently, the target training corpus information is vectorized into a training vector matrix, where each element corresponds to the position information of the word in the target training corpus information, providing a structured foundation for semantic analysis. Next, the training weight matrix captures the semantic influence of the row and column dimension elements (i.e., different words) in the vector matrix, quantifying the strength of the association between words. To simplify complex semantic relationships and focus on key influencing factors, the model performs low-rank decomposition of the weight matrix based on a preset number of dimensions to be decomposed (i.e., the number of core words to be analyzed), eliminating redundant associations and obtaining a simpler low-rank matrix. By calculating the change in the weight matrix before and after decomposition, the model can dynamically adjust the influence weight of each word, ultimately forming a final training weight matrix that integrates scenario characteristics and semantic core associations, thereby constructing a corpus analysis model that can accurately capture scenario semantics.

[0069] It is worth noting that when the weight matrix is ​​a decomposed low-rank matrix, and the decomposed low-rank matrix includes a first low-rank matrix and a second low-rank matrix, the step of decomposing the weight matrix based on the number of dimensions to be decomposed to obtain the decomposed low-rank matrix includes: obtaining a training first low-rank matrix and a training second low-rank matrix; calculating a training first semantic influence based on the training first low-rank matrix, wherein the training first semantic influence is the influence of row dimension elements on column dimension elements; calculating a training second semantic influence based on the training second low-rank matrix, wherein the training second semantic influence is the influence of column dimension elements on row dimension elements; if the training first semantic influence and / or the training second semantic influence do not conform to the corresponding training preset range, then based on the number of dimensions to be decomposed, the training first low-rank matrix and / or the training second low-rank matrix that do not conform to the corresponding training preset range are decomposed to obtain a re-decomposed training low-rank matrix.

[0070] Specifically, in the training of the corpus analysis model, the process of decomposing the weight matrix to obtain the low-rank matrix is ​​as follows: First, a matrix factorization algorithm (such as Singular Value Decomposition (SVD) or Non-negative Matrix Factorization (NMF)) is used to decompose the training weight matrix into a first training low-rank matrix and a second training low-rank matrix. These two matrices capture semantic relationships between words from different directions or dimensions. Based on the first training low-rank matrix, the weighted contribution of each row dimension element to each column dimension element is calculated to quantify the semantic influence of row dimension words on column dimension words, thus obtaining the first training semantic influence. Similarly, based on the second training low-rank matrix, the influence of each column dimension element on each row dimension element is calculated to obtain the second training semantic influence. Subsequently, these two semantic influence values ​​are compared with a pre-set training range. If any influence value exceeds a reasonable range (e.g., excessively high unidirectional semantic association leading to information imbalance), the low-rank matrix that does not meet the requirements is decomposed again, using the number of dimensions to be decomposed as a constraint, to optimize the matrix structure. This process corrects semantic association biases through bidirectional semantic verification and iterative decomposition, ensuring that the low-rank matrix can effectively compress information and accurately reflect the bidirectional semantic influence between lexical units, thereby improving the model's ability to capture complex semantic relationships.

[0071] To provide a more detailed explanation of the above embodiments, this application also provides another corpus analysis method, such as... Figure 2 As shown, the following specific steps are provided in this embodiment of the application:

[0072] Step 201: Combine the corpus information with the supplementary information to form target corpus information that conforms to the application scenario.

[0073] The system invokes a predefined scenario template bound to the application scenario. This predefined scenario template defines the field relationships and combination order of the corpus information and supplementary information. Based on these field relationships and combination order, the corpus information and supplementary information are filled into the corresponding fields of the predefined scenario template to obtain the target corpus information. Specifically, after determining the supplementary information, the system quickly locates and invokes the predefined scenario template bound to the previously matched application scenario. Taking a JSON-formatted predefined scenario template as an example, it clearly lists the correspondence between the corpus information, supplementary information, and each field, as well as the combination order of these fields in the final target corpus information. Then, after extracting the corresponding information from the corpus information and supplementary information according to the predefined scenario template, the extracted information is filled into the predefined scenario template according to the field information in the predefined scenario. For example, key content corresponding to predefined scene template fields can be identified from the corpus information, and corresponding supplementary data can be obtained from supplementary information. After information extraction, the extracted content is accurately filled into the corresponding fields in the predefined scene template according to the field rules, thereby generating target corpus information that meets the requirements. By embedding corpus information and supplementary information into the predefined scene template, the method of information integration is standardized, making the information obtained by the corpus analysis model more structured and organized. The model can understand and process data more accurately, thereby improving the accuracy of the analysis results.

[0074] For example, the corpus information is "I plan to book a double room with a view of West Lake in Hangzhou from August 15th to August 17th, preferably with a bathtub." Supplementary information includes a detailed list of facilities at various hotels in Hangzhou (including whether they have a view of West Lake and whether they have a bathtub) and past user reviews. The predefined scenario template clearly specifies the mapping relationship and combination order between each field and the corpus information and supplementary information. For example, the "Check-in Time" field maps to "August 15th" in the corpus information; "Check-out Time" maps to "August 17th"; "City" corresponds to "Hangzhou"; and "Room Type" corresponds to "a double room with a view of West Lake and a bathtub." After accurately extracting the relevant information from the corpus information, the supplementary information is processed according to the requirements of the predefined scenario template. First, hotels located in Hangzhou that offer double rooms with a view of West Lake and a bathtub are selected from the hotel facilities list. Then, past user reviews of these hotels that meet the criteria are collected, especially feedback regarding the view, bathtub facilities, and overall stay experience. Then, following the combination order set in the template, the check-in time, check-out time, city, room type (including special requests), hotel price range, facilities (highlighting West Lake view and bathtub), and user reviews are sequentially filled into the corresponding fields. After completing the construction of the target corpus information, it is input into the corpus analysis model. In some embodiments, the target corpus information is constructed in JSON format, for example as follows: "Check-in time": "August 15", "Check-out time": "August 17", "City": "Hangzhou", "Room type": "Double room with West Lake view and bathtub", "Facilities": "West Lake view, bathtub", "User reviews": "The hotel has an excellent location, the West Lake view is breathtaking, the bathtub is clean and tidy, but the breakfast selection is limited...".

[0075] Step 202: Input the target corpus information into the corpus analysis model and output the corpus analysis results corresponding to the corpus information.

[0076] In this step, the corpus analysis model contains multiple computational models, which can obtain corresponding results from the corpus information. The computational model can be a LoRa computational model.

[0077] When the target corpus information is input into the corpus analysis model, the accuracy of the corpus analysis result output by any computational model is calculated. If the accuracy of the corpus analysis result is greater than or equal to a preset accuracy threshold, the output of the computational model corresponding to the accuracy of the corpus analysis result is determined as the corpus analysis result corresponding to the corpus information. That is, after the target corpus information is input into the corpus analysis model, multiple computational models in the corpus analysis model output results based on the target corpus information. The corpus analysis model calculates the accuracy of the corpus analysis result based on the multiple output results. When the accuracy of the corpus analysis result is greater than or equal to a preset accuracy threshold, the output of the computational model corresponding to the accuracy of the corpus analysis result is determined as the corpus analysis result corresponding to the corpus information. Here, the accuracy of the corpus analysis result refers to the degree to which the computational model outputs accurately respond to the target corpus information. The preset accuracy threshold is set according to the actual situation, and its specific value is not limited here. This multi-model integration method utilizes the advantages of different models in different scenarios, avoids analysis errors caused by limited parameters, and improves the accuracy and robustness of the overall corpus analysis results without increasing the number of parameters in a single model.

[0078] When the corpus analysis results do not meet the expected results, the target corpus information is vectorized to obtain a vector matrix of the target corpus information. The elements of the vector matrix are used to represent the positional information of each word in the target corpus information. A weight matrix is ​​determined, and the elements of the weight matrix are used to represent the semantic influence between the row and column elements in the vector matrix. Based on the number of dimensions to be decomposed, the weight matrix is ​​decomposed to obtain a low-rank matrix after decomposition. The number of dimensions to be decomposed represents the number of words to be analyzed for semantic influence. The change in the weight matrix is ​​calculated based on the low-rank matrix after decomposition. Based on the change in the weight matrix and the weight matrix, the final weight matrix is ​​determined and the corpus analysis model is updated according to the final weight matrix. Specifically, when the corpus analysis results do not meet the expected results, the corpus analysis model fine-tunes the calculation model. At this time, the fine-tuned calculation model is the calculation model with a high degree of matching between the target corpus analysis and the calculation results. Specifically, when the corpus analysis results fail to meet expectations, the computational model outputting the corpus analysis results is fine-tuned. The fine-tuning process involves: vectorizing the target corpus information to obtain a vector matrix; determining the weight matrix based on this matrix; and performing low-rank decomposition on the weight matrix based on the number of dimensions to be decomposed, simplifying the complex semantic relationships into a set of core influencing factors and eliminating redundant associations. Next, the change in the weight matrix is ​​calculated based on the decomposed low-rank matrix, quantifying the adjustment magnitude of the semantic influence between word units. Finally, based on this change and the original weight matrix, the final weight matrix is ​​determined. This approach strengthens the association weights of key semantic word units and weakens the influence of noise information without increasing computational resource consumption, effectively improving the model's accuracy and adaptability in subsequent corpus analysis and ensuring the accuracy of the corpus analysis results.

[0079] During the weight matrix decomposition process, if the parameter change of the decomposed low-rank matrix exceeds a preset parameter change, the decomposed low-rank matrix is ​​further decomposed as follows: Based on the first low-rank matrix, a first semantic influence is calculated, which is the influence of row-dimensional elements on column-dimensional elements; based on the second low-rank matrix, a second semantic influence is calculated, which is the influence of column-dimensional elements on row-dimensional elements; if the first semantic influence and / or the second semantic influence do not conform to the corresponding preset range, then based on the number of dimensions to be decomposed, the first low-rank matrix and / or the second low-rank matrix that do not conform to the corresponding preset range are decomposed to obtain a re-decomposed low-rank matrix. During the weight matrix decomposition process, if the parameter change of the decomposed low-rank matrix exceeds a preset threshold, it means that the current decomposition has failed to accurately capture the semantic relationships in the corpus, and the decomposition effect needs to be optimized through a two-way semantic influence verification mechanism. Specifically, the weight matrix is ​​first decomposed into a first low-rank matrix and a second low-rank matrix, corresponding to the semantic influence of row-dimensional elements on column-dimensional elements (first semantic influence) and the semantic influence of column-dimensional elements on row-dimensional elements (second semantic influence), respectively.

[0080] Taking medical record analysis as an example, the first low-rank matrix quantifies the influence of the "symptom presentation" row dimension elements on the "treatment plan" column dimension elements, while the second low-rank matrix quantifies the inverse influence of the "treatment plan" column dimension elements on the "symptom presentation" row dimension elements, such as the feedback on the alleviating effect of the "antiviral treatment" treatment plan on symptoms like "fever and fatigue." If the semantic influence in a certain direction exceeds the preset range, the corresponding low-rank matrix is ​​further decomposed based on the number of dimensions to be decomposed, eliminating unreasonable weight biases that correct the association between treatment plans and symptoms, making the association between treatment decisions and symptoms in medical record analysis more accurate and reasonable.

[0081] Taking the intelligent transportation - urban road traffic optimization scenario as an example, this paper describes the innovation of bidirectional correlation quantization based on bidirectional semantic influence and low-rank matrix quantization logic. Bidirectional semantic influence refers to the mutual semantic influence between row-dimensional elements and column-dimensional elements; that is, the semantic influence of row-dimensional elements on column-dimensional elements, and the semantic influence between column-dimensional elements and row-dimensional elements. An example is shown below:

[0082] In the intelligent traffic flow analysis system, a bidirectional semantic influence matrix (such as the aforementioned weight matrix) is constructed to quantify the bidirectional semantic influence between the row-dimensional elements of the "real-time traffic status" and the column-dimensional elements of the "control strategy." The first low-rank matrix quantifies the semantic influence strength (first semantic influence degree) of the row-dimensional elements of the "real-time traffic status" (such as "distribution of congested road sections, peak traffic density") on the column-dimensional elements of the "control strategy" (such as "traffic light timing adjustment, activation of tidal lanes"), i.e., analyzing the correlation logic of congestion characteristics driving strategy decisions. The second low-rank matrix quantifies the reverse semantic influence (second semantic influence degree) of the column-dimensional elements of the "control strategy" on the row-dimensional elements of the "real-time traffic status," such as the improvement feedback on status indicators like "peak traffic density of road sections, and traffic delay duration" after the implementation of the "dynamic tidal lane" strategy. Through quantitative modeling of bidirectional semantic influence, the dynamic correlation between traffic status and control strategies is accurately matched, improving the efficiency of intelligent decision-making for urban road traffic optimization and providing technical support for strategy iteration and precise control of intelligent transportation systems.

[0083] As can be seen, based on the decomposition of the weight matrix into first and second low-rank matrices, the semantic influence between row and column elements is calculated separately, and low-rank matrices that do not meet the preset range are further decomposed. This method, through the precise calculation and verification of bidirectional semantic influence, compensates for the shortcomings of single weight matrix decomposition in characterizing semantic relationships. Especially when dealing with long texts or complex sentences, it can more finely adjust semantic associations, solve the problem of inaccurate understanding of complex semantic relationships by localized models, and significantly improve the accuracy of corpus analysis.

[0084] like Figure 3 The content shown illustrates the matrix decomposition process of the computational model as follows: Each computational model has an initial weight matrix of size 1000*2000. Through a certain operation (represented by "||"), this is transformed into a 1000*2 matrix and then processed by a 2*1000 matrix (using "||"). It is worth noting that there are no restrictions on the size of the initial weight matrix and the decomposed matrix of the calculation model; they can be set according to the actual situation.

[0085] This embodiment also provides a detailed description of the weight matrix decomposition of the computational model, as follows:

[0086] Now assume that the weight matrix (Z) of the computational model H can be decomposed into two optimal low-rank matrices (X) and (Y), i.e., (Z approx X times Y). The optimal low-rank matrix (X) is determined when the first semantic influence of the low-rank matrix (X), where the first semantic influence is the influence of row-dimensional elements on column-dimensional elements, falls within its corresponding preset range; similarly, the optimal low-rank matrix (Y) is determined when the second semantic influence of the low-rank matrix (Y), where the second semantic influence is the influence of column-dimensional elements on row-dimensional elements, falls within its corresponding preset range. The elements of the weight matrix Z can be represented by a series of dimension element pairs (Xi, Yi) from i=1 to i=N. N is a positive integer representing the index range of the matrix elements. The ranks of matrices (X) and (Y) are much smaller than the rank of (Z).

[0087] To obtain matrices (X) and (Y), the decomposition of the weight matrix needs to be fine-tuned. The process is as follows: The weight matrix W of the pre-trained model (assuming dimension d×k) is decomposed into the product of two low-rank matrices A (dimension d×r) and B (dimension r×k), i.e., W = A*B, where the rank r is much smaller than d and k. This significantly reduces the number of parameters that need to be trained, lowering computational complexity and memory usage. During the fine-tuning process, for the weight matrix Z decomposed into two low-rank matrices (A) and (B), only the matrix elements of the two low-rank matrices (A) and (B) are updated. The optimal two low-rank matrices (X) and (Y) can be achieved through this fine-tuning process.

[0088] After obtaining the corpus analysis results, the user's satisfaction with the corpus analysis results corresponding to the corpus information is also obtained. When the satisfaction is less than or equal to a preset satisfaction, after obtaining the cloud model access instruction, the target corpus information is divided into a first security level information and a second security level information using a preset information security level. The first security level information is encrypted using an asymmetric algorithm to obtain first encrypted information. The second security level information is encrypted using a symmetric algorithm to obtain second encrypted information. The first encrypted information and the second encrypted information are sent to a private cloud in a progressive manner. There are multiple private clouds, to obtain the first corpus analysis results.

[0089] The progressive transmission refers to sending encrypted information in batches based on the security level of the first and second encrypted information or the amount of data in the encrypted information. For example, encrypted information can be sent in descending order of security level, i.e., high-security-level encrypted information is sent first, followed by low-security-level encrypted information; or the amount of data in the first and second encrypted information can be determined, the data amounts sorted, and the encrypted information sent in batches according to the order of data amounts. The preset information security level refers to a classification rule based on the sensitivity of the data, which can be implemented using keyword filtering or regular expression matching to distinguish between privacy-related data and non-sensitive data. The first security level information refers to sensitive data, such as identity information or privacy information. The second security level information refers to non-sensitive data, involving conventional data. Asymmetric algorithms refer to encryption methods that use a public key for encryption and a private key for decryption, specifically implemented using RSA or ECC algorithms, suitable for protecting the transmission of high-security-level information. Symmetric algorithms refer to methods that use the same key for encryption and decryption, specifically implemented using AES or DES algorithms, suitable for quickly processing low-security-level information.

[0090] In some embodiments, when the satisfaction level of the corpus analysis result is greater than the preset satisfaction level, it indicates that the user is satisfied with the results of the corpus analysis model, and the corpus analysis ends at this time. When the satisfaction level of the corpus analysis result is less than or equal to the preset satisfaction level, it indicates that the user is not satisfied with the results of the corpus analysis model, and the corpus analysis model encrypts the target corpus information and transmits it to the private cloud to obtain the first corpus analysis result. The cloud server needs to set up a key pairing mechanism in advance to ensure that the cloud server can decrypt the encrypted target corpus information. If the first security level information is asymmetric encrypted, a public key and private key pairing mechanism needs to be set up in advance. The specific steps are as follows: (1) Key pair generation: Generate a key pair using an asymmetric encryption algorithm (such as RSA, ECC, etc.), including a public key and a private key. The public key can be publicly distributed, while the private key must be strictly kept confidential and can only be used by the private cloud. (2) Public key storage: Store the generated public key in the local device that needs to be encrypted. The public key can be distributed through a secure channel or made public in a public directory. (3) Encryption Operation: The local device uses the public key to encrypt the first-level security information. The encrypted ciphertext can only be decrypted using the corresponding private key. Specifically, the local device uses its own stored public key to encrypt the first-level security information and convert it into ciphertext. (4) Private Key Storage and Use: The private key is stored by multiple private clouds. When decryption is required, the multiple private clouds use the private key to decrypt the encrypted first-level security information and restore the original user request information.

[0091] Therefore, through the above technical solutions, this application achieves hierarchical and dynamic protection of target corpus information. It processes data of different sensitivity levels using a hybrid mode of asymmetric and symmetric encryption, ensuring the anti-cracking capability of core privacy fields while reducing the computational overhead of conventional data encryption. The progressive transmission method effectively avoids the network detection risk caused by transmitting the entire ciphertext in a single transmission by sending data packets of different encryption types in stages, making it impossible for attackers to locate the complete ciphertext structure through traffic analysis. The separation mechanism between key management and data transmission channels further blocks the correlation between encrypted information and decrypted elements, forming multi-dimensional protection at the transmission link level.

[0092] Step 203: Input the encrypted target corpus information into the private cloud to obtain the first corpus analysis results.

[0093] After encrypting the target corpus information in step 202, this step involves decrypting the encrypted target corpus information to obtain the data volume of the decrypted target corpus information. Based on the data volume and a multi-attribute decision algorithm, a comprehensive score for the private cloud is obtained. The target private cloud is determined according to the comprehensive score. The target corpus information is then input into the target private cloud to obtain the first corpus analysis result. Here, "data volume" refers to the size of the target corpus information. The multi-attribute decision algorithm is a computational model that weights candidate private clouds by setting multiple evaluation dimensions. Specifically, it can be implemented using the analytic hierarchy process (AHP) or grey relational analysis, generating a comprehensive score based on indicators such as computing power, response speed, data storage capacity, and network bandwidth. The comprehensive score is the ranking result after quantitatively evaluating the performance of each private cloud based on the multi-attribute decision algorithm. Specifically, it can be generated using normalized weighted summation or fuzzy comprehensive evaluation methods, used to dynamically select the private cloud with the highest matching degree to the current needs, which is the target private cloud.

[0094] Specifically, when acquiring the target corpus information and considering indicators such as computing power, response speed, data storage capacity, and network bandwidth of each candidate private cloud, if the data storage capacity is less than the data volume, the private cloud with the lower data storage capacity is removed from the candidate private clouds. A comprehensive score is calculated based on indicators such as computing power, response speed, data storage capacity, and network bandwidth, and the comprehensive scores are sorted from high to low. The private cloud with the highest comprehensive score is selected as the target private cloud, and the target corpus information is input into the target private cloud to obtain the first corpus analysis result. It can be seen that the above technical solution, by decrypting the encrypted target corpus information to obtain the data volume and combining it with a multi-attribute decision algorithm to calculate the comprehensive score of the candidate private clouds, can accurately select the most suitable target private cloud from multiple private clouds to process the request, ensuring efficient processing of the target corpus information and improving the system resource utilization efficiency and processing performance. Inputting the target corpus information into the target private cloud to obtain the target output result avoids the possibility of corpus information leakage caused by centralized processing of corpus information. In this embodiment, the specific implementation of inputting target corpus information into the target private cloud to obtain the first corpus analysis result is as follows: Feature information of the target corpus information is extracted using natural language processing technology; a query statement adapted to the target private cloud is constructed based on the feature information; the query statement is input into the generative large model of the target private cloud to obtain multiple sets of candidate outputs; structured data alignment processing is performed on the multiple sets of candidate outputs to generate intermediate results; the intermediate results are filtered and integrated according to the constraints in the user request information to generate the first corpus analysis result. The generative large model is used to upload corpus information to the generative large model in the cloud server for corpus analysis to obtain corpus analysis results.

[0095] In some embodiments, the step of extracting feature information from the target corpus using natural language processing (NLP) techniques can be as follows: based on NLP technology, the target corpus can be processed by word segmentation, part-of-speech tagging, named entity recognition, etc. In one example, a word vector model (such as Word2Vec or BERT) can be used to convert text into computer-understandable feature vectors, mining key feature information such as semantics, sentiment, and entities within the corpus. For instance, if the target corpus is "Recommended interactive attractions in Hangzhou West Lake suitable for family trips on weekends in June," word segmentation yields the feature information "Hangzhou West Lake," "June weekends," "family trips," and "attraction recommendations."

[0096] In some embodiments, the step of constructing a query statement adapted to the target private cloud based on feature information can be: analyzing the target private cloud's interface protocol and query syntax specifications (such as SQL-like queries and natural language commands), and combining the extracted corpus features to convert them into query commands executable by the private cloud. For example, assuming the target private cloud requires the query statement to be in the format of "[functional command] + [entity feature] + [requirement constraint]", based on the previous features, the query statement is constructed as: "Retrieve recommended scenic spots in the Hangzhou West Lake area, during weekends in June, suitable for family travel scenarios", ensuring that the private cloud can recognize and execute this command.

[0097] After obtaining the query statement, the query statement is input into the generative model of the target private cloud to obtain multiple sets of candidate outputs. For example, after inputting the above query statement, the model outputs the following candidates: ① "Family attractions around West Lake: Children's Palace, Natural History Museum, including interactive experiences"; ② "Family route around West Lake: Broken Bridge - Bai Causeway - Pinghu Qiuyue, suitable for family walks".

[0098] In some embodiments, the step of performing structured data alignment processing on the multiple sets of candidate outputs to generate intermediate results can be implemented by defining a unified data structure. For example, the data in ①-② above can be aligned according to the following data format:

[0099] Attraction Name: Location 1 / Location 2 / Location 3; Reasons for Recommendation: Reasons for Recommendation of Location 1 / Reasons for Recommendation of Location 2 / Reasons for Recommendation of Location 3; Suitable Scenarios: Suitable Scenarios for Location 1 / Suitable Scenarios for Location 2 / Suitable Scenarios for Location 3.

[0100] The processed data is as follows: "Attraction Name: Children's Palace / Natural History Museum / Broken Bridge-Bai Causeway-Autumn Moon over the Calm Lake, Recommendation Reason: Rich Interactive Experience / Route Suitable for Family Stroll, Applicable Scene: Parent-Child Interaction / Parent-Child Stroll, to generate a well-organized intermediate result dataset." The constraint "High Interaction" in the user request is extracted and filtered from ①-② to determine the results of the first corpus analysis.

[0101] In summary, by selecting a private cloud that matches the target language corpus information in the cloud server, it is possible to avoid allocating the target language corpus information to a private cloud with insufficient resources, thus ensuring the smoothness of the corpus analysis process, reducing analysis errors caused by resource mismatch, and improving the accuracy of the first corpus analysis results.

[0102] Step 204: Input the target corpus information into the generative large model to obtain the second corpus analysis results.

[0103] After outputting the first corpus analysis result in step 203, the user's satisfaction with the first corpus analysis result is obtained. If the satisfaction is less than or equal to a preset satisfaction level, a preset information security level is used to determine whether the target corpus information is information of the first security level. If yes, an instruction is sent based on the received user information to input the target corpus information into the generative large-scale model to obtain the second corpus analysis result. If no, the target corpus information is input into the generative large-scale model to obtain the second corpus analysis result. The generative large-scale model differs from locally deployed corpus analysis models; it is a cloud-based corpus analysis model with a large number of parameters.

[0104] Specifically, when users are still dissatisfied with the initial corpus analysis results output by the cloud server, the system automatically detects the security attributes of the target corpus information. If the target corpus information is determined to be at the first security level (i.e., sensitive information), it must wait for explicit user authorization before being input into the generative large-scale model for corpus analysis. If the target corpus information is determined to be at the second security level (i.e., non-sensitive information), the generative large-scale model is directly invoked for processing to obtain the second corpus analysis results. This application effectively solves the problem of insufficient analysis accuracy caused by parameter reduction in corpus analysis models. By establishing a multi-level feedback mechanism, the analysis path is dynamically adjusted when users remain dissatisfied, continuously outputting accurate second corpus analysis results while ensuring data security.

[0105] Furthermore, as a response to the above Figures 1-2 To implement the method shown, this application provides a corpus analysis device. This device embodiment corresponds to the foregoing method embodiment. For ease of reading, this device embodiment will not repeat the details of the foregoing method embodiment, but it should be understood that the device in this embodiment can implement all the content of the foregoing method embodiment. This device can obtain accurate corpus analysis results, specifically as follows... Figure 4 As shown, the device includes:

[0106] Unit 41 is used to analyze the input corpus information and obtain the corresponding semantic information;

[0107] The matching unit 42 is used to match a preset application scenario based on the semantic information obtained by the unit 41, wherein different application scenarios correspond to different supplementary information;

[0108] The acquisition unit 43 is used to acquire corresponding supplementary information according to the application scenario of the matching unit 42;

[0109] The combination unit 44 is used to combine the corpus information with the supplementary information of the acquisition unit 43 into target corpus information that conforms to the application scenario.

[0110] The output result unit 45 is used to input the target corpus information of the combination unit 44 into the locally deployed corpus analysis model and output the corpus analysis result corresponding to the corpus information.

[0111] Furthermore, such as Figure 5 As shown, the corpus analysis model contains multiple computational models, and the output result unit 45 includes...

[0112] The matrix acquisition module 451 is used to vectorize the target corpus information and obtain a vector matrix of the target corpus information when the corpus analysis result does not meet the expected analysis result. The elements of the vector matrix are used to represent the position information of each word in the target corpus information.

[0113] The matrix acquisition module 451 is used to determine the weight matrix, wherein the elements of the weight matrix are used to characterize the semantic influence between the row dimension elements and the column dimension elements in the vector matrix.

[0114] The decomposition module 452 is used to decompose the weight matrix of the acquisition matrix module 451 based on the number of dimensions to be decomposed, and obtain the decomposed low-rank matrix. The number of dimensions to be decomposed represents the number of lexical units to be analyzed for semantic influence.

[0115] The transformation module 453 is used to calculate the change in the weight matrix based on the low-rank matrix decomposed by the decomposition module 452.

[0116] The update module 454 is used to determine the final weight matrix based on the change in the weight matrix and the weight matrix, and update the corpus analysis model according to the final weight matrix.

[0117] Furthermore, such as Figure 5 As shown, when the weight matrix is ​​a decomposed low-rank matrix, and the decomposed low-rank matrix includes a first low-rank matrix and a second low-rank matrix, the decomposition module 452 further includes: calculating a first semantic influence degree based on the first low-rank matrix, wherein the first semantic influence degree is the influence degree of row dimension elements on column dimension elements; calculating a second semantic influence degree based on the second low-rank matrix, wherein the second semantic influence degree is the influence degree of column dimension elements on row dimension elements; if the first semantic influence degree and / or the second semantic influence degree do not conform to the corresponding preset range, then based on the number of dimensions to be decomposed, the first low-rank matrix and / or the second low-rank matrix that do not conform to the corresponding preset range are decomposed to obtain a re-decomposed low-rank matrix.

[0118] Furthermore, such as Figure 5 As shown, the output result unit 45 includes:

[0119] When the target corpus information is input into the corpus analysis model, the accuracy of the corpus analysis result output by any calculation model is calculated;

[0120] If the accuracy of the corpus analysis result is greater than or equal to a preset accuracy threshold, the output of the calculation model corresponding to the accuracy of the corpus analysis result is determined as the corpus analysis result corresponding to the corpus information.

[0121] Furthermore, such as Figure 5 As shown, the device further includes a result acquisition unit 46, which includes:

[0122] The satisfaction acquisition module 461 is used to acquire the user's satisfaction with the corpus analysis results corresponding to the corpus information;

[0123] The segmentation module 462 is used to, when the satisfaction level of the satisfaction acquisition module 461 is less than or equal to the preset satisfaction level, after obtaining the cloud model access instruction, divide the target corpus information into first security level information and second security level information using the preset information security level.

[0124] Encryption module 463 is used to encrypt the first security level information of partitioning module 462 using an asymmetric algorithm to obtain first encrypted information;

[0125] The encryption module 463 is used to encrypt the second security level information of the partitioning module 462 using a symmetric algorithm to obtain the second encrypted information.

[0126] The analysis result module 464 is used to send the first encrypted information and the second encrypted information of the encryption module 463 to the private cloud in a progressive sending manner to obtain the first corpus analysis result.

[0127] Furthermore, such as Figure 5 As shown, the result acquisition unit 46 includes:

[0128] The Result Satisfaction Module 465 is used to obtain user satisfaction with the analysis results of the first corpus.

[0129] The judgment module 466 is used to determine whether the target corpus information is first security level information when the satisfaction of the obtained result satisfaction module 465 is less than or equal to the preset satisfaction level, using the preset information security level.

[0130] The judgment module 466 is used to determine that the target corpus information is information of the first security level, and then send an instruction based on the received user information to input the target corpus information into the generative large model to obtain the second corpus analysis result.

[0131] The judgment module 466 is used to determine that the target corpus information is second security level information, and then input the target corpus information into the generative large model to obtain the second corpus analysis result.

[0132] Furthermore, such as Figure 5 As shown, the matching unit 42 includes:

[0133] Extraction module 421 is used to extract keyword information from the semantic information;

[0134] The keyword matching module 422 is used to match the keyword information of the extraction module 421 with preset keywords corresponding to different application scenarios, so as to obtain the keyword information matching degree under different application scenarios.

[0135] The scenario determination module 423 is used to determine the matching degree of different scenarios based on the keyword information matching degree of different application scenarios of the keyword matching module 422, and to determine the application scenario corresponding to the semantic information by the application scenario corresponding to the maximum value of the matching degree of different scenarios.

[0136] Furthermore, such as Figure 5 As shown, the assembly unit 44 includes:

[0137] The module 441 is used to call a predefined scenario template bound to the application scenario. The predefined scenario template defines the field mapping relationship and combination order between the corpus information and the supplementary information.

[0138] The target corpus information acquisition module 442 is used to fill the corpus information and the supplementary information into the corresponding fields of the predefined scene template based on the field mapping relationship and combination order of the calling module 441, so as to acquire the target corpus information.

[0139] Furthermore, such as Figure 5As shown, the output result unit 45 includes: training the corpus analysis model. The specific training process is as follows: acquiring target training corpus information, which is determined based on initial training corpus information and corresponding training supplementary information, and the training supplementary information is obtained based on the application scenario matched by the initial training corpus information; vectorizing the target training corpus information to obtain a training vector matrix, where the elements of the training vector matrix represent the position information of each word in the target training corpus information; acquiring a training weight matrix, where the elements of the training weight matrix represent the semantic influence between row dimension elements and column dimension elements in the training vector matrix; decomposing the training weight matrix based on the number of dimensions to be decomposed to obtain a decomposed low-rank matrix, where the number of dimensions to be decomposed represents the number of words to be analyzed for semantic influence; obtaining the weight matrix change based on the decomposed low-rank matrix; and determining the final training weight matrix based on the weight matrix change and the training weight matrix, and determining the corpus analysis model based on the final training weight matrix.

[0140] When the weight matrix is ​​a decomposed low-rank matrix, and the decomposed low-rank matrix includes a first low-rank matrix and a second low-rank matrix, the output result unit 45 includes: obtaining the training first low-rank matrix and the training second low-rank matrix; calculating the training first semantic influence degree based on the training first low-rank matrix, wherein the training first semantic influence degree is the influence degree of row dimension elements on column dimension elements; calculating the training second semantic influence degree based on the training second low-rank matrix, wherein the training second semantic influence degree is the influence degree of column dimension elements on row dimension elements; if the training first semantic influence degree and / or the training second semantic influence degree do not conform to the corresponding training preset range, then based on the number of dimensions to be decomposed, the training first low-rank matrix and / or the training second low-rank matrix that do not conform to the corresponding training preset range are decomposed to obtain the decomposed training low-rank matrix.

[0141] Furthermore, such as Figure 6 As shown, this application also provides a corpus analysis model training device, the device comprising:

[0142] The information acquisition unit 51 is used to acquire target training corpus information, which is determined based on initial training corpus information and training supplementary information corresponding to the initial training corpus information. The training supplementary information is obtained based on the application scenario matched by the initial training corpus information.

[0143] The information acquisition unit 51 is used to vectorize the target training corpus information to obtain a training vector matrix of the target training corpus information. The elements of the training vector matrix are used to represent the position information of each word in the target training corpus information.

[0144] The information acquisition unit 51 is used to acquire a training weight matrix, wherein the elements of the training weight matrix are used to characterize the semantic influence between the row dimension elements and the column dimension elements in the training vector matrix.

[0145] The matrix unit 52 is used to decompose the training weight matrix in the information acquisition unit based on the number of dimensions to be decomposed, and to obtain the decomposed low-rank matrix. The number of dimensions to be decomposed represents the number of lexical units to be analyzed for semantic influence.

[0146] The matrix acquisition unit 52 is used to obtain the change in the weight matrix based on the decomposed low-rank matrix.

[0147] Model determination unit 53 is used to determine the final training weight matrix based on the weight matrix change in the acquisition matrix unit 512 and the training weight matrix, and to determine the corpus analysis model based on the final training weight matrix.

[0148] Furthermore, when the weight matrix is ​​a decomposed low-rank matrix, and the decomposed low-rank matrix includes a first low-rank matrix and a second low-rank matrix, the matrix acquisition unit 52 includes:

[0149] The training matrix acquisition module 521 is used to acquire the first low-rank training matrix and the second low-rank training matrix;

[0150] The training matrix calculation module 522 is used to calculate the first semantic influence degree of training based on the first low-rank training matrix in the training matrix acquisition module 521, wherein the first semantic influence degree of training is the influence degree of row dimension elements on column dimension elements.

[0151] The training matrix calculation module 522 is used to calculate the second semantic influence degree of training based on the second low-rank matrix of training obtained in the training matrix acquisition module 521. The second semantic influence degree of training is the influence degree of column dimension elements on row dimension elements.

[0152] If the training first semantic influence degree and / or the training second semantic influence degree do not conform to the corresponding training preset range, the training matrix calculation module 522 decomposes the training first low-rank matrix and / or the training second low-rank matrix that do not conform to the corresponding training preset range based on the number of dimensions to be decomposed, and obtains the decomposed training low-rank matrix.

[0153] Furthermore, embodiments of the present invention also provide a computing device, the computing device comprising: at least one processor, and a memory, wherein the memory stores instructions executable by the processor, the instructions being executed by the processor, thereby enabling the processor to perform the above-described operations. Figure 1-2The corpus analysis method or the corpus analysis model training method described herein.

[0154] Furthermore, embodiments of the present invention also provide a readable storage medium for storing a computer program, wherein the computer program, when running, controls the device where the storage medium is located to perform the above-described actions. Figure 1-2 The corpus analysis method or the corpus analysis model training method described herein.

[0155] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0156] It is understood that the relevant features in the above methods and apparatus can be referenced interchangeably. Furthermore, the terms "first," "second," etc., in the above embodiments are used to distinguish between embodiments and do not represent the superiority or inferiority of any particular embodiment.

[0157] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0158] The algorithms and displays provided herein are not inherently related to any particular computer, virtual system, or other device. Various general-purpose systems can also be used in conjunction with the teachings herein. The required structure for constructing such systems is apparent from the above description. Furthermore, this invention is not directed to any particular programming language. It should be understood that the contents of the invention described herein can be implemented using various programming languages, and the above description of specific languages ​​is for the purpose of disclosing the best mode of implementation of the invention.

[0159] In addition, the memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0160] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0161] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0162] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0163] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0164] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0165] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0166] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0167] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0168] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, variations or substitutions readily conceived by those skilled in the art within the scope of the technology disclosed in this application, embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code, should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A corpus analysis method, characterized in that, The method includes: The input corpus information is analyzed to obtain the corresponding semantic information; The semantic information is matched with a preset application scenario, wherein different application scenarios correspond to different supplementary information. The supplementary information is auxiliary data retrieved from the local knowledge base that is adapted to the application scenario and can enhance the understanding of the scenario. Collect corresponding supplementary information based on the application scenario; The corpus information and the supplementary information are combined to form target corpus information that conforms to the application scenario. The target corpus information is information that is structured and integrated with the corpus information and supplementary information according to a predefined scenario template. The target corpus information is input into a locally deployed corpus analysis model, and the corpus analysis results corresponding to the corpus information are output. The method further includes: When the corpus analysis results do not meet the expected analysis results, the target corpus information is vectorized to obtain a vector matrix of the target corpus information. The elements of the vector matrix are used to represent the position information of each word in the target corpus information. Determine a weight matrix, the elements of which are used to characterize the semantic influence between the row-dimensional elements and the column-dimensional elements in the vector matrix; Based on the number of dimensions to be decomposed, the weight matrix is ​​decomposed to obtain a low-rank matrix after decomposition. The number of dimensions to be decomposed represents the number of lexical units to be analyzed for semantic influence. The change in the weight matrix is ​​calculated based on the decomposed low-rank matrix. Based on the changes in the weight matrix and the weight matrix itself, the final weight matrix is ​​determined and the corpus analysis model is updated according to the final weight matrix. When the weight matrix is ​​a decomposed low-rank matrix, and the decomposed low-rank matrix includes a first low-rank matrix and a second low-rank matrix, the step of decomposing the weight matrix based on the number of dimensions to be decomposed to obtain the decomposed low-rank matrix includes: Based on the first low-rank matrix, the first semantic influence degree is calculated, which is the influence degree of the row dimension element on the column dimension element; Based on the second low-rank matrix, the second semantic influence degree is calculated, which is the influence degree of column dimension elements on row dimension elements; If the first semantic influence degree and / or the second semantic influence degree do not conform to the corresponding preset range, then based on the number of dimensions to be decomposed, the first low-rank matrix and / or the second low-rank matrix that do not conform to the corresponding preset range are decomposed to obtain the decomposed low-rank matrix.

2. The method according to claim 1, characterized in that, The locally deployed corpus analysis model includes multiple computational models, and the method includes: When the target corpus information is input into the corpus analysis model, the accuracy of the corpus analysis result output by any calculation model is calculated; If the accuracy of the corpus analysis result is greater than or equal to a preset accuracy threshold, the output of the calculation model corresponding to the accuracy of the corpus analysis result is determined as the corpus analysis result corresponding to the corpus information.

3. The method according to claim 1, characterized in that, The method further includes: Obtain user satisfaction with the corpus analysis results corresponding to the corpus information; When the satisfaction level is less than or equal to the preset satisfaction level, after obtaining the instruction to access the cloud model, the target corpus information is divided into first security level information and second security level information based on the preset information security level. The first security level information is encrypted using an asymmetric algorithm to obtain the first encrypted information; The second security level information is encrypted using a symmetric algorithm to obtain the second encrypted information; The first and second encrypted information are sent to the private cloud to obtain the first corpus analysis results.

4. The method according to claim 1, characterized in that, The method further includes: Obtain user satisfaction with the analysis results of the first corpus; If the satisfaction level is less than or equal to a preset satisfaction level, then based on a preset information security level, when the target corpus information is determined to be information of the first security level, upon receiving a user information sending instruction, the target corpus information is input into a generative large model to obtain the second corpus analysis result; or If the satisfaction level is less than or equal to the preset satisfaction level, then based on the preset information security level, when the second security level information of the target corpus information is determined, the target corpus information is input into the generative large model to obtain the second corpus analysis result.

5. The method according to claim 1, characterized in that, The step of matching a preset application scenario based on the semantic information includes: Extract keyword information from the semantic information; The keyword information is matched with preset keywords corresponding to different application scenarios to obtain the keyword information matching degree under different application scenarios; The matching degree of different scenarios is determined based on the keyword information matching degree under different application scenarios, and the application scenario corresponding to the maximum matching degree of the different scenarios is determined as the application scenario corresponding to the semantic information.

6. The method according to claim 1, characterized in that, The step of combining the corpus information with the supplementary information to form target corpus information that conforms to the application scenario includes: Call a predefined scenario template bound to the application scenario, wherein the predefined scenario template defines the field mapping relationship and combination order between corpus information and supplementary information; Based on the field mapping relationship and combination order, the corpus information and the supplementary information are filled into the corresponding fields in the predefined scene template to obtain the target corpus information.

7. A method for training a corpus analysis model, characterized in that, include: Obtain target training corpus information, which is determined based on initial training corpus information and training supplementary information corresponding to the initial training corpus information, and the training supplementary information is obtained based on the application scenario matched by the initial training corpus information; The target training corpus information is vectorized to obtain a training vector matrix of the target training corpus information. The elements of the training vector matrix are used to represent the position information of each word in the target training corpus information. Obtain a training weight matrix, the elements of which are used to characterize the semantic influence between row-dimensional elements and column-dimensional elements in the training vector matrix; Based on the number of dimensions to be decomposed, the training weight matrix is ​​decomposed to obtain a low-rank matrix after decomposition. The number of dimensions to be decomposed represents the number of lexical units to be analyzed for semantic influence. Based on the decomposed low-rank matrix, the change in the weight matrix is ​​obtained; Based on the changes in the weight matrix and the training weight matrix, the final training weight matrix is ​​determined, and the corpus analysis model is determined according to the final training weight matrix. When the weight matrix is ​​a decomposed low-rank matrix, and the decomposed low-rank matrix includes a first low-rank matrix and a second low-rank matrix, the step of decomposing the weight matrix based on the number of dimensions to be decomposed to obtain the decomposed low-rank matrix includes: Obtain the first low-rank training matrix and the second low-rank training matrix; Based on the first low-rank training matrix, the first semantic influence degree of training is calculated, which is the influence degree of row dimension elements on column dimension elements. Based on the trained second low-rank matrix, the trained second semantic influence degree is calculated, which is the influence degree of column dimension elements on row dimension elements; If the first semantic influence of training and / or the second semantic influence of training do not conform to the corresponding preset training range, then based on the number of dimensions to be decomposed, the first low-rank training matrix and / or the second low-rank training matrix that do not conform to the corresponding preset training range are decomposed to obtain the decomposed low-rank training matrix.

8. A corpus analysis device, characterized in that, The device includes: The unit is used to analyze the input corpus information and obtain the corresponding semantic information; A matching unit is used to match a preset application scenario based on the obtained unit semantic information. Different application scenarios correspond to different supplementary information. The supplementary information is auxiliary data retrieved from the local knowledge base that is adapted to the application scenario and can enhance the understanding of the scenario. The acquisition unit is used to acquire corresponding supplementary information according to the application scenario of the matching unit; The combination unit is used to combine the corpus information with the supplementary information of the acquisition unit into target corpus information that conforms to the application scenario. The target corpus information is information after the corpus information and supplementary information are structured and integrated according to a predefined scenario template. The output result unit is used to input the target corpus information of the combined unit into the locally deployed corpus analysis model and output the corpus analysis result corresponding to the corpus information; The corpus analysis model contains multiple computational models, and the output result unit includes... The matrix acquisition module is used to vectorize the target corpus information and obtain a vector matrix of the target corpus information when the corpus analysis result does not meet the expected analysis result. The elements of the vector matrix are used to represent the position information of each word in the target corpus information. The matrix acquisition module is used to determine the weight matrix, wherein the elements of the weight matrix are used to characterize the degree of semantic influence between the row dimension elements and the column dimension elements in the vector matrix; The decomposition module is used to decompose the weight matrix of the matrix acquisition module based on the number of dimensions to be decomposed, and obtain the decomposed low-rank matrix. The number of dimensions to be decomposed represents the number of lexical units to be analyzed for semantic influence. The transformation module is used to calculate the change in the weight matrix based on the low-rank matrix decomposed by the decomposition module. The update module is used to determine the final weight matrix based on the change in the weight matrix and the weight matrix, and update the corpus analysis model according to the final weight matrix; When the weight matrix is ​​a decomposed low-rank matrix, and the decomposed low-rank matrix includes a first low-rank matrix and a second low-rank matrix, the decomposition module further includes: calculating a first semantic influence degree based on the first low-rank matrix, wherein the first semantic influence degree is the influence degree of row dimension elements on column dimension elements; calculating a second semantic influence degree based on the second low-rank matrix, wherein the second semantic influence degree is the influence degree of column dimension elements on row dimension elements; if the first semantic influence degree and / or the second semantic influence degree do not conform to the corresponding preset range, then based on the number of dimensions to be decomposed, the first low-rank matrix and / or the second low-rank matrix that do not conform to the corresponding preset range are decomposed to obtain a re-decomposed low-rank matrix.

9. A corpus analysis model training device, characterized in that, The device includes: An information acquisition unit is used to acquire target training corpus information, wherein the target training corpus information is determined based on initial training corpus information and training supplementary information corresponding to the initial training corpus information, and the training supplementary information is obtained based on the application scenario matched by the initial training corpus information; The information acquisition unit is used to vectorize the target training corpus information to obtain a training vector matrix of the target training corpus information, wherein the elements of the training vector matrix are used to characterize the position information of each word in the target training corpus information. The information acquisition unit is used to acquire a training weight matrix, wherein the elements of the training weight matrix are used to characterize the semantic influence between the row dimension elements and the column dimension elements in the training vector matrix. The matrix unit is used to decompose the training weight matrix in the information acquisition unit based on the number of dimensions to be decomposed, and to obtain the decomposed low-rank matrix. The number of dimensions to be decomposed represents the number of lexical units to be analyzed for semantic influence. The matrix acquisition unit is used to obtain the change in the weight matrix based on the decomposed low-rank matrix. A model unit is defined to determine the final training weight matrix based on the change in the weight matrix in the acquisition matrix unit and the training weight matrix in the acquisition matrix unit, and to determine the corpus analysis model based on the final training weight matrix. When the weight matrix is ​​a decomposed low-rank matrix, and the decomposed low-rank matrix includes a first low-rank matrix and a second low-rank matrix, the matrix acquisition unit includes: The module for obtaining the training matrix is ​​used to obtain the first low-rank training matrix and the second low-rank training matrix; The training matrix calculation module is used to calculate the first semantic influence of training based on the first low-rank training matrix in the training matrix acquisition module, wherein the first semantic influence of training is the influence of row dimension elements on column dimension elements. The training matrix calculation module is used to calculate the second semantic influence degree of training based on the second low-rank matrix of training obtained in the training matrix acquisition module. The second semantic influence degree of training is the influence degree of column dimension elements on row dimension elements. If the training first semantic influence degree and / or the training second semantic influence degree do not conform to the corresponding training preset range, the training matrix calculation module decomposes the training first low-rank matrix and / or the training second low-rank matrix that do not conform to the corresponding training preset range based on the number of dimensions to be decomposed, and obtains the decomposed training low-rank matrix.

10. A computing device, characterized in that, The computing device includes: at least one processor, and a memory, wherein the memory stores instructions executable by the processor, the instructions being executed by the processor so that the processor can perform the corpus analysis method as described in any one of claims 1-6 or the corpus analysis model training method as described in claim 7.

11. A readable storage medium, characterized in that, The readable storage medium is used to store a computer program, wherein the computer program, when running, controls the device where the storage medium is located to execute the corpus analysis method as described in any one of claims 1-6 or the corpus analysis model training method as described in claim 7.

Images