Term data intelligent management system based on machine learning

Abstract

The invention discloses an intelligent term data management system based on machine learning. The intelligent term data management system comprises the steps that multi-source data are collected through a data collection module, the data are preprocessed, the preprocessed data are subjected to related storage, and the data are orderly and effectively stored through a table structure in the storage process; the method comprises the following steps: receiving query requests from different users through a query request module, analyzing the query requests into query tasks, intelligently scheduling and optimizing the query tasks by using a query optimization algorithm through an intelligent query engine to obtain query results, and returning the query results to the users through a result feedback module. And receiving feedback data of the user, acquiring the feedback data and historical query data of the user through the learning module, and continuously training and updating the query model according to the feedback data and the historical query data of the user.

Description

Technical Field

[0001] The present invention belongs to the technical field of data intelligent query, and in particular relates to a terminology data intelligent management system based on machine learning. Background Art

[0002] With the development of big data technology and the popularization of artificial intelligence, query systems have been widely used in various fields. However, traditional query systems often rely on static data structures and predefined query rules, which are difficult to adapt to dynamically changing needs and diverse query conditions. Existing query systems have low query efficiency when facing large-scale data and cannot provide accurate query results in a timely manner. Therefore, there is an urgent need for an intelligent query system that can adaptively adjust query strategies and improve query efficiency and accuracy.

[0003] 1. Low data retrieval efficiency: When processing large-scale data, the existing query system has a slow retrieval speed and a long response time, which makes it difficult to meet users' needs for efficient query.

[0004] 2. Insufficient accuracy of query results: The results of traditional query systems are often not accurate enough and cannot fully meet the user's query intent, resulting in users having to conduct multiple queries to find the required information.

[0005] 3. Poor user experience: Complex query interfaces and operation steps increase the difficulty of user use and affect user experience

[0006] 4. Difficulty in processing multi-source heterogeneous data: Traditional systems are complex to process data from different data sources and in different formats, and it is difficult to achieve unified query and analysis of data. Summary of the invention

[0007] In order to solve the above technical problems, the present invention proposes a terminology data intelligent management system based on machine learning to solve the problems existing in the above-mentioned prior art.

[0008] To achieve the above object, the present invention provides a terminology data intelligent management system based on machine learning, comprising:

[0009] The query request module, the intelligent query engine and the result feedback module are sequentially connected; the intelligent query engine is also connected to the learning module and the data collection module;

[0010] The multi-source heterogeneous data are collected through the data collection module, and the multi-source heterogeneous data are preprocessed and stored. The query request module receives query requests from different users and parses the query requests into query tasks. The query tasks are intelligently scheduled and optimized using the query optimization algorithm through the intelligent query engine. According to the optimization result, the stored data is extracted to obtain the query result, wherein the query result is the standard data and the corresponding associated data. The query result is returned to the user through the result feedback module, and the user's feedback data is received. The feedback data and the user's historical query data are obtained through the learning module. According to the feedback data and the user's historical query data, the query optimization algorithm is continuously trained and updated through the machine learning model.

[0011] Optionally, the preprocessed data is stored in a database, wherein the database uses a table structure to store the data, wherein different data types are set in the table structure, wherein the data types include national standard numbers, term numbers, Chinese terms, English terms, other preferred Chinese terms, other preferred English terms, permitted Chinese terms, Chinese definitions, abbreviations, symbols, units, formulas, sources, figures, tables, notes, references, examples, deprecated terms in Chinese, deprecated terms in English, rejected terms in Chinese, and rejected terms in English, and associated data are stored in the rows of the table structure, and the data in each row is stored in corresponding columns according to different data types.

[0012] Optionally, in the data acquisition module, the preprocessed data is stored in a distributed database, and the stored data is indexed and optimized. Index optimization includes: selecting appropriate columns to create indexes, strategies when using joint indexes, operations to avoid index invalidation, considering covering indexes, index maintenance, and using auto-incrementing primary keys.

[0013] Optionally, a user interface is provided in the query request module, wherein a custom query port and a result filtering port are provided in the user interface for inputting query requests and visualizing and selecting query results, respectively, wherein the query request is any one or more data types of national standard number, term number, Chinese term, English term, other preferred Chinese term, other preferred English term, and permitted Chinese term.

[0014] Optionally, in the query request module, the process of parsing the query request into a query task includes:

[0015] The query request is converted into the same language, and keywords are extracted from the converted query request through natural language processing technology. The keyword extraction results are converted, and query tasks are generated based on the converted keywords.

[0016] Optionally, in a query optimization algorithm, the process of scheduling and optimizing query tasks through an optimization algorithm includes: obtaining an initial query task, constructing and initializing a population based on the initial query task, wherein the population includes a number of individuals, wherein the individuals include different query tasks; setting an objective function, fitness, and population update method, wherein the objective function is to minimize the sum of the CPU cost and IO cost corresponding to the query task; and generating the optimal query task by searching for the best solution through an optimization algorithm based on the objective function, fitness, and population update method.

[0017] Optional population update methods include:

[0018] Mi=Di·e bθ ·cos(2πθ)+ω·Fj

[0019] Among them, b is a constant used to define the shape of the logarithmic spiral; θ is a random number between [-1,1]; Mi represents the i-th moth, that is, the individual; Fj represents the j-th flame; Di represents the distance between the i-th moth and the j-th flame, and the formula is: Di = |Fj-Mi|, and ω is the inertia weight.

[0020] Optionally, the inertia weight is:

[0021]

[0022] Where t is the current iteration number; T is the maximum iteration number; a and d are constants respectively.

[0023] Optionally, in the learning module, the machine learning model includes an input layer, an attention mechanism, a convolutional layer and an output layer; wherein the input data of the machine learning model includes the query results and corresponding query paths under historical query data, and the output is the path corresponding to the future query results and relevant configurable parameters of the relevant optimization scheduling algorithm.

[0024] Optionally, in the learning module, the training process of the machine learning model includes: taking satisfactory query results and corresponding query paths in the feedback data as positive samples, taking unsatisfactory query results and corresponding query paths as negative samples, taking the corresponding future user feedback satisfactory query results and corresponding query paths and relevant configurable parameters of the current relevant optimization scheduling algorithm as labels, and training the machine learning model.

[0025] Compared with the prior art, the present invention has the following advantages and technical effects:

[0026] 1. Significantly improve query efficiency: Through data-driven intelligent algorithms and optimized data structure design, combined with improved optimization algorithms, the query speed is greatly improved, the response time is shortened, and the overall performance of the system is improved.

[0027] 2. Improve the accuracy of query results: Use advanced machine learning and natural language processing technologies to improve the accuracy of query results and ensure that users can quickly obtain the information they need.

[0028] 3. Optimize user experience: design a simple and easy-to-use user interface, simplify the query operation steps, and enable users to use the system for query more conveniently.

[0029] 4. Realize unified query of multi-source heterogeneous data: Through data integration and standardized processing, realize unified query, analysis and storage of data from different sources and in different formats, and through table structures with data association, be able to associate different data and provide comprehensive data support. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The drawings constituting a part of the present application are used to provide a further understanding of the present application. The illustrative embodiments and descriptions of the present application are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0031] Figure 1 A schematic diagram of the system structure of an embodiment of the present invention;

[0032] Figure 2 A schematic diagram of the conceptual table structure of an embodiment of the present invention;

[0033] Figure 3 A schematic diagram of the term library table structure of an embodiment of the present invention;

[0034] Figure 4 A schematic diagram of a foreign language vocabulary structure corresponding to terms in an embodiment of the present invention;

[0035] Figure 5 It is a schematic diagram of the structure of the conceptual other content table of an embodiment of the present invention. DETAILED DESCRIPTION

[0036] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0037] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0038] like Figure 1As shown, in this embodiment, a terminology data intelligent management system based on machine learning is provided, comprising: a query request module, an intelligent query engine and a result feedback module in sequence; the intelligent query engine is also connected to a learning module and a data collection module;

[0039] The multi-source data is collected through the data collection module, and the data is preprocessed, and the preprocessed data is stored in a related manner; the query request module receives query requests from different users, wherein the query request is any one or more data types of national standard number, term number, Chinese term, English term, other preferred Chinese term, other preferred English term, and permitted Chinese term, and the query request is parsed into an executable query task, and the query task is intelligently scheduled and optimized by the intelligent query engine using the query optimization algorithm to obtain the query result, wherein the query result is standard data and corresponding associated data, and the query result is returned to the user through the result feedback module, and the user's feedback data is received, and the feedback data and the user's historical query data are obtained through the learning module, and the query model is continuously trained and updated according to the feedback data and the user's historical query data.

[0040] As some embodiments, in the data acquisition module, the relevant contents of preprocessing multi-source heterogeneous data include data cleaning and data standardization, etc.; in the relevant storage of the preprocessed data, a distributed storage method is used in the storage process, and the high-performance operation of the system is improved by storing different data in a distributed database, and a corresponding index is constructed for the stored data, and the index is used as the basis for searching. When constructing the index, the index is optimized by a relevant index optimization method, and the index optimization method includes: selecting a suitable column to create an index, using a strategy when using a joint index, avoiding operations that cause index failure, considering covering indexes, index maintenance, using an auto-increment primary key or other index optimization methods to construct and optimize the index, and ensure the validity of the index. This ensures effective data query and efficient system operation.

[0041] As some embodiments, in the data acquisition module, multi-source heterogeneous data are subjected to relevant standardized processing through the access port before storage, including format conversion, data integration and related operations such as the above-mentioned preprocessing, and are stored in a database that reflects their relationship based on their relevance.

[0042] In the database, this embodiment stores data from different sources or with correlation through relevant table structures. During the storage process, the table structure in the database is first constructed. In the table structure, the data attributes or data types corresponding to each column are national standard number, term number, Chinese term, English term, other preferred Chinese term, other preferred English term, permitted Chinese term, Chinese definition, abbreviation, symbol, unit, formula, source, figure, table, note, see, example, deprecated term in Chinese, deprecated term in English, rejected term in Chinese, rejected term in English.

[0043] In the storage of related data, in the table structure, each row represents data with related connections. For example, for the related data with "Chinese term" "A", the above data is supplemented and stored in the corresponding row and in the corresponding column. For example, the national label "B" is related to "A", and "B" is stored in the column corresponding to the national label in the corresponding row of "A", so as to improve the relevance of the stored related data. At the same time, when searching or extracting certain data in the database, such as providing the term number "C" in the same row of "A", based on "C", "A", "B" and other data in the same row are provided or provided, that is, the related data of different columns or data attributes in the corresponding row are also provided or extracted, so as to provide users with more comprehensive data and information extraction. In storage, the data is stored in the corresponding column to distinguish the data attributes of different data. The relevance of the data is improved through the storage relationship in the same row, so that the data can be effectively and orderly stored, and the relevance and orderliness of data extraction can be improved.

[0044] In another embodiment, the present invention stores the above data through multiple table structures, wherein during the storage process, a concept table, a term table, a term corresponding to a foreign word table, and a concept other content table are used;

[0045] For redundant information that may exist in a table, such as a term or multiple related information, such as multiple sources or forms, when using a table structure for storage, the current term needs to be stored multiple times (represented in the table structure as using multiple rows for the same information or storing it by splitting it with a delimiter) to store all its information. This will cause some redundant content. In order to reduce the redundancy, the above information is stored through multiple table structures through the association between its information and related information.

[0046] like Figure 2 As shown in the table, in the concept table, the data attributes or data types corresponding to each column are serial number, national standard number, term number, Chinese definition, source standard number, source term number, whether it is modified, modification description, and footnote; Figure 3As shown in the table of terms, the data attributes or data types corresponding to each column are serial number, concept table-serial number, term, adoption level, language code, order, form, and footnote; Figure 4 As shown in the table of foreign terms corresponding to terms, the data attributes or data types corresponding to each column are serial number, term list-serial number, foreign word, language code, adoption level (preferred, allowed, rejected), and order; Figure 5 As shown, in the concept other content table, the data attributes or data types corresponding to each column are serial number, concept table-serial number, content, type (example, note, refer to), order, first example, second example...), footnote.

[0047] The above-mentioned concept table, term list, term corresponding foreign word list and concept other content table are associated through the serial number in the table structure and the serial number of other table structures in the table structure. Specifically, in the term list, the concept table-serial number corresponds to the serial number in the concept table, so that the related row data in the concept table and the term list are associated; in the term corresponding foreign word list, the term list-serial number is associated with the serial number in the term list, specifically, the term corresponding foreign word list and the related row data in the term list are associated; in the concept other content table, the concept table-serial number corresponds to the serial number in the concept table, so that the related row data in the concept other content table are associated; and then the above-mentioned data are stored through different table structures, so that different table structures can store different data, and then more storage methods are provided for different storage methods, providing a structural basis for parallel, serial or distributed storage, and providing a basis for improving the storage performance of the database.

[0048] In the storage of related data, in the above different table structures, each row represents data with related connections, such as Figure 2 As shown, in the concept table, exemplarily, for the types of the above different columns, for the associated data in the row with the serial number "1", the corresponding national standard number, term number, term number, Chinese definition, source standard number, source term number, and whether there is any modification are stored as "1", "GB / T 2001.1-2024", "3.3.1", "Define the reference (3.4.1.1.1) and definition (3.4.2.1) of concepts (3.2.1) used in specific fields (3.3.1) or professions (3.3.2).", "GB / T 20000.1-2014", "7.2", "Modified";

[0049] like Figure 3As shown, in the term list, exemplarily, for the types of the above-mentioned different columns, for the concept table-serial number, term, adoption level, language code in the associated data in the row with serial number "1", they are stored as: "1", "Term Standard", "Preferred", "CH" respectively; the subsequent sequence, form, and footnote are filled in according to the specific content, which are not displayed here; for the concept table-serial number, term, adoption level, language code, sequence, form, and footnote in the associated data in the row with serial number "2", they are stored as: "2", "Term Entry", "Preferred", "CH" respectively; the subsequent sequence, form, and footnote are filled in according to the specific content, which are not displayed here; in the term list, the concept table is associated with the concept table according to the concept table-serial number, wherein the corresponding associated data of the above-mentioned exemplary serial numbers 1 and 2 are associated with the corresponding associated data of serial numbers 1 and 2 in the concept table;

[0050] like Figure 4 As shown, in the terminological corresponding foreign-language vocabulary, exemplarily, in the types of the above-mentioned different columns, for the serial number "5", the corresponding terminological table-serial number, foreign word, language code, adoption level (preferred, permitted, rejected), and order are stored as "5", "terminological data collection", "ENG", "permitted", and "1" respectively; wherein the terminological corresponding foreign-language vocabulary is associated with the terminological table according to the terminological table-serial number, wherein the corresponding associated data of the above-mentioned exemplary serial number 6 is also associated with the corresponding associated data of serial number 5 in the terminological table, and based on the above-mentioned content, in the exemplary table structure content, for the terminological table serial number 5, the two rows of data 5 and 6 in the table structure are used for explanation and corresponding association, and the redundancy of a single table structure is reduced through multiple table structures;

[0051] like Figure 5 As shown, in the table of other contents of concepts, exemplarily, for the types of the above-mentioned different columns, for the serial number "1", the corresponding concept table-serial number, content, type (example, note, see), and order are stored as "1", "Terminology standards include terms and their definitions, sometimes with figures, notes, examples, formulas, etc.", "Note", "1"; wherein the table of other contents of concepts is associated with the concept table according to the concept table-serial number, wherein the corresponding associated data of the above-mentioned exemplary serial number 2 is associated with the corresponding associated data of the serial number 1 in the concept table, and at the same time, based on the above-mentioned content, in the exemplary table structure content, for the content of the concept table serial number 1, it is explained through the two row data of serial numbers 1 and 2, and the redundancy of a single table structure is reduced through multiple table structures.

[0052] When outputting the subsequent results, in addition to outputting the associated row data of the same table for the found standard terms, the data in different tables are output through the associated structure of the related serial numbers of different table structures, so that the queried data is more comprehensive.

[0053] At the same time, as it is feasible and clearly explained, the above-mentioned multiple table structures can also be used as the database corresponding to the query system of student elective course information. For the query system of student elective course information, when using a single table structure as mentioned above, it is necessary to set the data attributes or data types corresponding to each column separately. The data types involved include student ID, student name, class, course name, course number, professor, etc. If all the above data types can be put into one table, then this table must be very redundant. If a student chooses multiple courses, all information must be stored multiple times (such as storing different courses for the same student through multiple rows of data). If, in order to avoid multiple storage, multiple courses are put into the course name and separated by some separators, it will definitely not be as convenient as putting them separately when querying, and it will definitely not be convenient to modify after the query.

[0054] As an improvement, the above content is stored through multiple table structures. Specifically, if the information student number, student name, class, course name, course number, and professor are divided into three tables, the first table structure stores student information, student number, student name, and class; the second table structure stores subject information, course name, course number, and professor; the third table structure stores the connection between students and courses, and only the student number and subject number are written. In this way, if a student chooses multiple courses, the contents of the first and second tables are fixed, and only the connection between students and courses (the third table) needs to be written. There is no need to repeatedly store the contents of the first and second tables, and there is no need to store them multiple times. In addition, it is very convenient to query and modify. As a key technical concept, this part of the implementation content can be used as an implementation example, and at the same time, a simple example is used to explain the technical concept of the above multiple table structures to relevant personnel.

[0055] As some embodiments, the query request module belongs to a query port for different users. It can be user-oriented by designing a relevant user interface and performing relevant custom query and result filtering interfaces in the user interface to improve the user's experience and satisfaction. After the user enters a relevant query request in the query interface, the query request is processed by a natural language processing model to generate data suitable for database query. The query request is a relevant query statement or a phrase with relevant search terms input by the user. After the query port obtains the user's query request, the query request is first judged for its language category, and its language category is judged by a language category recognition model. At the same time, after the judgment, it is converted into a unified language category, such as Chinese, through a language conversion model to support complex queries and multilingual queries, and improve the relevance of query results. The language category recognition model and the language conversion model adopt relevant deep learning models, which can be used through existing models and will not be described here.

[0056] As some embodiments, after the query request is converted into a standard query statement through the above two models, the above standard query statement is subjected to natural language keyword extraction through natural language processing technology or natural language processing model. Before the natural language processing model is used, the stop words in the query statement are removed through a stop word list, and then the keyword extraction is performed through the natural language processing model, and the extracted keywords are converted into a format suitable for data query, such as related characters or vectors. The Llama model or Bert model adopted by the natural language processing model is not described here. It should be noted that if a phrase of a search word is used, it can be judged by whether it uses a phrase of a search word, such as setting an option for keyword combination retrieval in the query port to indicate that it uses a phrase of a search word, or judging by setting a threshold value of its character position or the number of times it appears in the search, and after judging, the search words in the phrase with related search words are removed, and the keywords or phrases are used as the basis of the query task. Apply natural language processing technology to understand the user's query intention, support complex queries and multilingual queries, and improve the relevance of query results.

[0057] As some embodiments, in an intelligent query engine, query tasks are intelligently scheduled and optimized through a data-driven query optimization algorithm to generate an optimal query task, i.e., an optimal query plan, and the data in the database is queried through the optimal query task to obtain query results. The operation content of the specific settings of the query optimization algorithm and the query task belongs to conventional content and will not be repeated here. In the present invention, the optimization algorithm conventionally used in the scheduling and optimization process of the query task in the query optimization algorithm is improved to improve the scheduling and optimization efficiency of its query task. Conventional optimization algorithms generally use greedy algorithms or genetic algorithms to schedule and optimize query tasks, but first of all, greedy algorithms and genetic algorithms are prone to fall into local optimal solutions and cannot provide the optimal query task under the global situation. At the same time, the scheduling optimization efficiency of their query tasks is not high. In the present invention, an improved new optimization method is introduced to adjust and optimize the query task to provide the optimal query task.

[0058] Specifically, the improved new optimization method adopts the improved moth algorithm, and its specific goal is to minimize the sum of the CPU cost and IO cost corresponding to the query task. The specific steps of the optimization algorithm are:

[0059] 1) Initialize the population, where the population contains several different query tasks, where the positions are selected based on manual experience or randomly as initial individuals, and set the relevant objective function, constraints and corresponding fitness. The constraints are the same as those of the conventional optimization algorithm, and the fitness function is the inverse of the sum of the CPU cost and the IO cost. The population is composed of multiple individuals, and the individual is the basic relationship in each query task or the temporary relationship generated in the middle in the multi-table connection. The specific content of the temporary relationship is represented by a string.

[0060] 2) In the improved moth-to-fire algorithm, the moth needs to update and iterate the flames around it. The flames represent the local optimal solution, and the number of flames is determined. At the same time, during the iteration process, the number of flames needs to be adaptively reduced. The adaptive reduction mechanism formula is as follows:

[0061] FN = round(Nt*(N-1) / T)

[0062] Where: N is the initial number of flames; t is the current number of iterations; T is the maximum number of iterations, and FN represents the number of flames during the tth iteration.

[0063] 3) Calculate the fitness of different individuals. The larger the fitness, the better the quality of the individual. During the calculation process, the sum of the corresponding CPU cost and IO cost is calculated according to the query tasks in scheduling and optimization to obtain the corresponding fitness. The sum of the CPU cost and IO cost corresponding to the above query task calculation is publicly provided in the conventional query optimization process and will not be repeated here.

[0064] The fitness is sorted from large to small. After sorting, the corresponding number of flames is selected according to the sorting result, and the individuals in the fitness sorting are used as flames. The number of flames is calculated according to the above fitness reduction mechanism formula, and the current best flame is stored. The best flame is the individual with the highest number of flames in the fitness sorting.

[0065] 4) According to the above individual update method, or the moth position update method, the update is performed. In the update process, the initial position update method is first used for updating. In the initial position update method, the above moth position update method is adaptively updated through the inertia weight:

[0066] Mi=Di·e bθ ·cos(2πθ)+ω·Fj

[0067] Among them, b is a constant used to define the shape of the logarithmic spiral; θ is a random number between [-1,1]; Mi represents the i-th moth, that is, the individual; Fj represents the j-th flame; Di represents the distance between the i-th moth and the j-th flame, and the formula is: Di=|Fj-Mi|, ω is the inertia weight, which refers to the inertia weight idea of ​​the particle swarm algorithm. The larger the ω, the greater the search intensity in the global range of the algorithm, and the smaller the ω, the higher the search accuracy in the local range of the algorithm.

[0068]

[0069] Where t is the current iteration number; T is the maximum iteration number; a and d are constants respectively.

[0070] 5) After the update, update the moth position again:

[0071]

[0072] Among them, MI represents the label of the moth individual that is updated again, rand() represents a random number, and Levy represents the corresponding Levy random step.

[0073]

[0074] Among them, β is a constant, which has been set to 1.5, and the distribution number μ~N(0,σ 2 ), σ 2 represents the upper limit; random distribution number v~N(0,1);

[0075]

[0076] Where Г() represents the standard gamma function.

[0077] 6) Determine whether the flame data satisfies the requirement of only one flame remaining. If only one flame remains, directly output the final result;

[0078] 8) Determine whether the maximum number of iterations has been reached. If not, repeat the above steps 2-6. If reached, output the optimal moth individual, that is, the optimal query task, that is, the optimal query plan.

[0079] The relevant query path is obtained through the optimal query task, and the query-related data in the database is extracted as the query result through the relevant index.

[0080] As some embodiments, after obtaining the query results, a result feedback module is used to provide the user with a visualized query result in a user interface and obtain user feedback, wherein the feedback data includes the user's use of specific data in the query results and a score of satisfaction.

[0081] As some embodiments, in the learning module, the relevant parameters and query paths (query plans) in the above-mentioned query optimization algorithm are optimized by machine learning methods through the relevant feedback data and user historical query data obtained. Improve retrieval efficiency and accuracy. Specifically, the relevant feedback data and user historical query data belong to the content of the query path and the relevant configurable parameters of the relevant optimization scheduling algorithm. The above data are used to predict the possible relevant query paths and optimal relevant configurable parameters of future users, and based on this, the query tasks are scheduled and optimized through the above-mentioned optimization scheduling algorithm to improve its query efficiency and performance.

[0082] In the machine learning method, a machine learning model is used to process the above data as input data to output the corresponding results, wherein the machine learning model includes an input layer, an attention mechanism, a convolution layer and an output layer. The input data includes the query results and corresponding paths under the relevant historical query data, and the output is the path corresponding to the future query results and the relevant configurable parameters of the relevant optimization scheduling algorithm, such as CPU scheduling strategy and adjustable parameters, process priority, I / O waiting time, and scheduling class settings. During the training process, the query results and corresponding query paths that the user feedback is satisfactory are used as positive samples, and the unsatisfactory ones are used as negative samples. The corresponding future user feedback satisfactory query results and corresponding query paths and the relevant configurable parameters of the current relevant optimization scheduling algorithm are used as labels to train the above machine learning model. The training process is not described in detail here.

[0083] In the machine learning model, the above input data is input through the input layer and the input data is adjusted to the same length or size.

[0084] The attention mechanism is used to assign weights to the input data at different times, where the attention mechanism is used to dynamically assign weights to each input data V i Assign weight w i,j , i and j represent the labels of input data at different times, and N represents the total number of corresponding input data, where score(V i ) is used to measure the relative importance of each input data. After calculating the weighted average, the comprehensive feature data V is obtained t :

[0085]

[0086] The use of the attention mechanism not only effectively retains the correlation information of the upper and lower time order of the data, but also avoids the information loss that may be caused by the limitation of data length.

[0087] In the multiple convolutional layers connected in sequence, the comprehensive feature vector V in the sample is sequentially calculated in the convolutional layer. t Perform convolution operations to extract rich local features C k The operation of the convolution kernel k in each convolution layer can be expressed as:

[0088] C k =ReLU(W K *V t +b k )

[0089] Among them, W K is the weight of filter k, b k is the bias term, * represents the convolution operation, which enables each convolution kernel to slide on the comprehensive feature vector and capture specific local patterns, and ReLU represents the activation function.

[0090] In the pooling layer, the maximum pooling operation is applied to each query result and the corresponding local features in the corresponding query path data to extract the most significant features and reduce the feature dimension:

[0091] P k =MaxPool(C k )

[0092] The pooled feature P k Fusion and processing are performed to form a long convolution feature vector F dense :

[0093] F concat =[P1; P2; ..; P N ]

[0094] F dense =ReLU(W dense *Fconcat +b dense )

[0095] Among them, W dense and b dense are the weights and biases of the fully connected layer respectively.

[0096] In the output layer, different output functions are used to output the paths corresponding to future query results and the relevant configurable parameters of the optimization scheduling algorithm.

[0097] The outputted parameters of the optimization scheduling algorithm are used to set the parameters in the optimization scheduling algorithm, and the paths corresponding to the future query results predicted in the future are used as the initial population or individual in the above optimization algorithm to improve the operation efficiency of the optimization scheduling algorithm and the optimization algorithm, and ensure the accuracy of its query task scheduling and optimization. Figure 2 The output result of the table structure shown is, among which, the national standards of relevant Chinese terms "magnetic field", "basic magnetic deflection polarity" and "magnetic dipole" are used as keywords of the query request respectively, and relevant queries are performed. Through the above-mentioned relevant query optimization system, the corresponding source or the corresponding index of the corresponding data is output, and then the standard terms and corresponding related data are extracted through the relevant keyword extraction method, including the relevant definitions, national standard numbers, term numbers, and English terms are extracted from the data corresponding to the index. At the same time, in order to facilitate the relevant display, important information is extracted through the relevant keyword extraction method, and the relevant related data is continued to be output. The keyword extraction method adopts big data deep learning or other existing related technologies. This is the content of query optimization post-processing and will not be elaborated here. For the above-mentioned query content, the query time is less than 1s, at the millisecond level. At the same time, the source or index position of the query provided as the query after the query request has a strong connection with the above-mentioned query request, which meets the relevant query expectations.

[0098] Compared with the prior art, the present invention has the following advantages and technical effects:

[0099] 1. Significantly improve query efficiency: Through data-driven intelligent algorithms and optimized data structure design, the query speed is greatly improved, the response time is shortened, and the overall performance of the system is improved.

[0100] 2. Improve the accuracy of query results: Use advanced machine learning and natural language processing technologies to improve the accuracy of query results and ensure that users can quickly obtain the information they need.

[0101] 3. Optimize user experience: design a simple and easy-to-use user interface, simplify the query operation steps, and enable users to use the system for query more conveniently.

[0102] 4. Realize unified query of multi-source heterogeneous data: Through data integration and standardized processing, realize unified query and analysis of data from different sources and in different formats, and provide comprehensive data support.

[0103] The above are only preferred specific implementations of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims

1. A terminology data intelligent management system based on machine learning, characterized in that: include: Query request module, intelligent query engine and result feedback module in sequence; The intelligent query engine is also connected to a learning module and a data collection module; The multi-source heterogeneous data is collected through the data collection module, and the multi-source heterogeneous data is preprocessed, and the preprocessed data is stored; Query requests from different users are received through the query request module, and the query requests are parsed into query tasks. The query tasks are intelligently scheduled and optimized using the query optimization algorithm through the intelligent query engine. According to the optimization result, the stored data is extracted to obtain the query result, wherein the query result is the standard data and the corresponding associated data. The query result is returned to the user through the result feedback module, and the user's feedback data is received. The feedback data and the user's historical query data are obtained through the learning module. According to the feedback data and the user's historical query data, the query optimization algorithm is continuously trained and updated through the machine learning model.

2. The system according to claim 1, characterized in that The preprocessed data is stored in a database, wherein a table structure is used to store the data in the database, wherein different data types are arranged in the table structure, wherein the data types include national standard number, term number, Chinese term, English term, other preferred Chinese term, other preferred English term, permitted Chinese term, Chinese definition, abbreviation, symbol, unit, formula, source, figure, table, note, reference, example, Chinese name of deprecated term, English name of deprecated term, Chinese name of rejected term, English name of rejected term, associated data is stored in the rows of the table structure, and the data in each row is stored in corresponding columns according to different data types.

3. The system according to claim 1, characterized in that In the data acquisition module, the preprocessed data is stored through a distributed database, and the stored data is indexed and optimized. Index optimization includes: selecting appropriate columns to create indexes, strategies when using joint indexes, operations to avoid index invalidation, considering covering indexes, index maintenance, and using auto-incrementing primary keys.

4. The system according to claim 1, characterized in that A user interface is provided in the query request module, wherein a custom query port and a result filtering port are provided in the user interface for inputting query requests and visualizing and selecting query results, respectively, wherein the query request is a data type related to any one or several of national standard numbers, term numbers, Chinese terms, English terms, other preferred Chinese terms, other preferred English terms, and permitted Chinese terms.

5. The system according to claim 1, characterized in that In the query request module, the process of parsing a query request into a query task includes: The query request is converted into the same language, and keywords are extracted from the converted query request through natural language processing technology. The keyword extraction results are converted, and query tasks are generated based on the converted keywords.

6. The system according to claim 1, characterized in that In the query optimization algorithm, the process of scheduling and optimizing query tasks through the optimization algorithm includes: obtaining an initial query task, constructing and initializing a population based on the initial query task, wherein the population includes a number of individuals, wherein the individuals include different query tasks; setting an objective function, fitness and population update method, wherein the objective function is to minimize the sum of the CPU cost and the IO cost corresponding to the query task; and searching through the optimization algorithm based on the objective function, fitness and population update method to generate the optimal query task.

7. The system according to claim 1, characterized in that Population update methods include: Mi=Di·e bθ ·cos(2πθ)+ω·Fj Among them, b is a constant used to define the shape of the logarithmic spiral; θ is a random number between [-1,1]; Mi represents the i-th moth, that is, the individual; Fj represents the j-th flame; Di represents the distance between the i-th moth and the j-th flame, and the formula is: Di = |Fj-Mi|, and ω is the inertia weight.

8. The system according to claim 1, characterized in that The inertia weight is: Where t is the current iteration number; T is the maximum iteration number; a and d are constants respectively.

9. The system according to claim 1, characterized in that In the learning module, the machine learning model includes an input layer, an attention mechanism, a convolutional layer and an output layer; among them, the input data of the machine learning model includes the query results and the corresponding query paths under the historical query data, and the output is the path corresponding to the future query results and the relevant configurable parameters of the relevant optimization scheduling algorithm.

10. The system according to claim 9, characterized in that In the learning module, the training process of the machine learning model includes: taking the satisfactory query results and the corresponding query paths in the feedback data as positive samples, taking the unsatisfactory query results and the corresponding query paths as negative samples, taking the corresponding future user feedback satisfactory query results and the corresponding query paths and the relevant configurable parameters of the current relevant optimization scheduling algorithm as labels, and training the machine learning model.

Images