A named entity recognition method and system based on knowledge-enhanced pre-training model

Abstract

The present invention discloses a method and system for named entity recognition based on a knowledge-enhanced pre-training model, relates to the technical field of named entity recognition, and is intended to solve the problem that the existing training model cannot perform named entity recognition more accurately. The present invention introduces a knowledge-enhanced pre-training model, so that the model can learn more prior knowledge and semantic information, thereby improving the generalization ability of unseen text data. Through a detailed performance evaluation and optimization process, the interpretability of the model is enhanced, which helps to understand the working mechanism of the model, thereby better adjusting and optimizing the model. The optimized model also improves in generalization ability and can better adapt to different application scenarios and data sets. The generation of monitoring logs makes performance analysis and problem tracking more convenient, which helps to continuously optimize model performance, and can improve the standard knowledge-enhanced training model according to user feedback on the target application terminal.

Description

Technical Field

[0001] The present invention relates to the technical field of named entity recognition, and in particular to a method and system for named entity recognition based on a knowledge-enhanced pre-training model. Background Art

[0002] Named entity recognition is a task in the field of natural language processing. Its purpose is to identify entities with specific meaning or strong referentiality from text, such as names of people, places, organizations, time expressions, amounts, etc.

[0003] The Chinese patent application document with publication number CN116451764A discloses a query set-based entity recognition model training method, device, equipment and storage medium. The method mainly uses the entity query set as a fixed learnable vector set to treat named entity recognition as a sequence-to-set task, and then performs self-attention weight calculation on each entity query vector representation in the entity query set to obtain a first enhanced vector expression set, which can capture the dependency relationship between entities and avoid problems such as duplication and redundancy of the final output entity. Then, the entity query set and the text vector are cross-attention weighted to obtain a second enhanced vector expression set, which allows the entity query set and the text vector to interact and optimize the entity query set, so that the entity recognition model can obtain richer data information and more accurate entity recognition results. Although the above patent application document solves the problem of model training, the following problems still exist in actual operation:

[0004] 1. There is no targeted graph construction and model selection based on text data, which leads to reduced accuracy when training models based on text data.

[0005] 2. The constructed model and text data are not effectively embedded, which results in the model being unable to perform better named entity recognition based on the text content.

[0006] 3. The completed training model is not monitored in real time on the application terminal, which makes it impossible to further understand the actual application of the training model. Summary of the invention

[0007] The purpose of the present invention is to provide a method and system for named entity recognition based on a knowledge-enhanced pre-training model. By introducing the knowledge-enhanced pre-training model, the model can learn more prior knowledge and semantic information, thereby improving the generalization ability of unseen text data. Through a detailed performance evaluation and optimization process, the interpretability of the model is enhanced, which helps to understand the working mechanism of the model, thereby better adjusting and optimizing the model. The optimized model also has improved generalization ability and can better adapt to different application scenarios and data sets. The generation of monitoring logs makes performance analysis and problem tracking more convenient, which helps to continuously optimize model performance. The standard knowledge-enhanced training model can be improved according to user feedback on the target application terminal, and the problems in the prior art can be solved.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A method for named entity recognition based on a knowledge-enhanced pre-training model comprises the following steps:

[0010] S1: Text data collection and processing: collect text data from the database, process the text data after collection, and obtain the target text data after data processing;

[0011] S2: Knowledge base construction: Confirm the knowledge information source from the knowledge base, construct the knowledge graph with the confirmed knowledge information source and the target text data, and obtain the target knowledge graph data after the knowledge graph is constructed;

[0012] S3: Training model selection and processing: selecting a training model according to the named entity recognition rule, and adjusting the selected training model to obtain a recognition training model after adjustment;

[0013] S4: Knowledge enhancement model training: embed the target knowledge graph data into the recognition training model, and perform model training on the recognition training module embedded with the target knowledge graph data. After model training, the knowledge enhancement model is obtained.

[0014] S5: Named entity recognition: input the text to be recognized into the knowledge enhancement model for named entity simulation recognition, and perform recognition result processing on the named entity simulation recognition result to obtain named entity data after the recognition result processing;

[0015] S6: Enhanced model evaluation and optimization: Perform performance evaluation on named entity data and optimize the knowledge enhancement model based on the performance evaluation results.

[0016] Preferably, in S1, the text data is collected from the database, and data processing is performed after the text data is collected, including:

[0017] Collect text data from the database using the data access interface;

[0018] Clean the collected text data and standardize the data after cleaning;

[0019] The data related to the target domain in the text data is screened according to the named entity recognition rules, and the named entity recognition rules are retrieved from the database;

[0020] And mark the screened data related to the target field;

[0021] After data labeling is completed, the target text data is obtained.

[0022] Preferably, the knowledge information source is confirmed from the knowledge base in S2, and a knowledge graph is constructed using the confirmed knowledge information source and the target text data, including:

[0023] Confirm the knowledge information source from the knowledge base, and extract the key entities, relationships and attribute data in the knowledge information source using natural language processing technology;

[0024] Use named entity recognition technology to identify entities and concepts in target text data;

[0025] Align the entities and concepts identified in the target text data with the key entities, relations, and attribute data extracted from the knowledge information source;

[0026] After data alignment, the mapping relationship between entities is established;

[0027] Use knowledge graph construction technology to construct knowledge graphs of entities and mapping relationships;

[0028] Conduct a quality assessment on the constructed knowledge graph and optimize it based on the assessment results. The optimization includes adding missing entities and relationships, correcting erroneous data, and removing redundant information.

[0029] The constructed knowledge graph is optimized to obtain the target knowledge graph data.

[0030] Preferably, setting the refresh frequency of the target knowledge graph data includes:

[0031] Extract the number of entities in the target knowledge graph data;

[0032] Extract the number of relationships corresponding to each entity;

[0033] Extract the correlation coefficient corresponding to each relationship of each entity;

[0034] The relationship strength coefficient corresponding to each entity is obtained by using the number of relationships corresponding to each entity and the association coefficient corresponding to each relationship of each entity;

[0035] The relationship strength coefficient is obtained by the following formula:

[0036]

[0037] Among them, S represents the relationship strength coefficient; n represents the number of relationships corresponding to each entity; G i represents the correlation coefficient of the i-th relationship corresponding to each entity; N represents the total number of relationships contained in the target knowledge graph data; G c Indicates the preset reference value of the correlation coefficient; G max Indicates the maximum value of the correlation coefficient corresponding to each entity; G zmax Represents the maximum value of the correlation coefficient contained in the target knowledge graph data; G z Indicates the middle value of the correlation coefficient contained in the target knowledge graph data;

[0038] comparing the relationship strength coefficient with a preset strength coefficient threshold;

[0039] When the number of entities whose relationship strength coefficient exceeds a preset strength coefficient threshold exceeds a preset number threshold, the refresh frequency of the target knowledge graph data is set, and the target knowledge graph data is refreshed according to the set refresh frequency;

[0040] When the number of entities whose relationship strength coefficient exceeds the preset strength coefficient threshold does not exceed the preset number threshold, the target knowledge graph data is refreshed using the preset initial refresh frequency.

[0041] Preferably, when the number of entities whose relationship strength coefficient exceeds a preset strength coefficient threshold exceeds a preset number threshold, the refresh frequency of the target knowledge graph data is set, including:

[0042] Entities exceeding a preset strength coefficient threshold are regarded as target entities;

[0043] Entities that do not exceed a preset intensity coefficient threshold are regarded as observed entities;

[0044] Extracting the relationship strength coefficient corresponding to the target entity;

[0045] Extracting the relationship strength coefficient corresponding to the observed entity;

[0046] Obtaining a frequency adjustment coefficient using a relationship strength coefficient corresponding to the target entity and a relationship strength coefficient corresponding to the observed entity;

[0047] The frequency adjustment coefficient is obtained by the following formula:

[0048]

[0049] Where, J represents the frequency adjustment coefficient; S b Represents the standard deviation of the relationship strength coefficients corresponding to all entities in the target knowledge graph data; S mb represents the standard deviation of the relationship strength coefficient corresponding to the target entity; S kb represents the standard deviation of the relationship strength coefficient corresponding to the observed entity; f represents the adjustment factor, and the adjustment factor is obtained by the following formula:

[0050]

[0051] Where f represents the adjustment factor; m represents the number of target entities; k represents the number of observed entities; S gi represents the intensity coefficient corresponding to the i-th observed entity; S mi represents the intensity coefficient corresponding to the i-th target entity; S y Indicates the preset intensity coefficient threshold; S b Represents the standard deviation of the relationship strength coefficients corresponding to all entities in the target knowledge graph data; S mb represents the standard deviation of the relationship strength coefficient corresponding to the target entity; S kb represents the standard deviation of the relationship strength coefficient corresponding to the observed entity; e represents the number of all entities contained in the target knowledge graph data;

[0052] The refresh frequency of the target knowledge graph data is set by using the frequency adjustment coefficient in combination with the preset initial refresh frequency;

[0053] The refresh frequency after setting is obtained by the following formula:

[0054]

[0055] Among them, F t represents the refresh frequency; F0 represents the preset initial refresh frequency; J represents the frequency adjustment coefficient.

[0056] Preferably, selecting a training model according to the named entity recognition rule in S3 and adjusting the selected training model include:

[0057] Retrieving named entity recognition rules from a database and confirming recognition requirements in the named entity recognition rules, the recognition requirements including entity type definition, recognition priority, and context dependency;

[0058] The standards for training models are formulated according to the recognition requirements. The standards for training models include the model architecture, learning ability, generalization ability, and the ability to integrate domain knowledge.

[0059] Select a training model based on the established training model standards, including BERT, RoBERTa, or ALBERT;

[0060] Perform model adjustment on the selected training model according to the established training model standards;

[0061] Model adjustment includes architecture adjustment, parameter adjustment and recognition layer adjustment;

[0062] After the model adjustment is completed, the recognition training model is obtained.

[0063] Preferably, embedding the target knowledge graph data into the recognition training model in S4, and performing model training on the recognition training module embedded with the target knowledge graph data, includes:

[0064] Calculate entity vectors of entities and relationships of target knowledge graph data;

[0065] After the entity vector is calculated, the vector representation of the target knowledge graph data is obtained;

[0066] Add a knowledge embedding layer to the recognition training model, where the knowledge embedding layer includes an input layer or an intermediate layer;

[0067] Integrate the vector representation of the target knowledge graph data into the knowledge embedding layer added by the recognition training model;

[0068] Adjust the model parameters of the recognition training model after the vector representation integration, including the learning rate, batch size and regularization term;

[0069] After the model parameters are adjusted, the recognition training model integrated with the vector representation is used for model training, and the change of the loss function is monitored during the model training process;

[0070] Adjust the structure of the model's knowledge embedding layer based on the model training results;

[0071] Finally, a knowledge-enhanced model is obtained.

[0072] Preferably, in S5, inputting the text to be recognized into the knowledge enhancement model for named entity simulation recognition, and performing recognition result processing on the named entity simulation recognition result, includes:

[0073] Retrieve the text to be recognized from the database, where the text to be recognized is simulated recognition text data for named entity recognition;

[0074] Use the word segmenter to segment and tokenize the text to be recognized;

[0075] Use entity linking technology to match entities in the text to be recognized after word segmentation and tokenization with entities in the target knowledge graph data, and perform format conversion after matching;

[0076] The text to be recognized after format conversion is input into the knowledge enhancement model, and the knowledge enhancement model uses the knowledge embedding layer to infer the input text to be recognized;

[0077] The knowledge enhancement model generates an output tensor for the inferred text to be recognized. The generated output tensor contains the classification score of each input token.

[0078] Based on the classification scores in the output tensor, an activation function is used to predict the entity category of each word in the output tensor;

[0079] Extract named entities in the text to be recognized based on the prediction results, identify the starting and ending positions of the extracted named entities, and correspond the identified positions to the entity types;

[0080] Organize the corresponding completed text to be recognized into a structured data format, including the text, category, location and confidence of the named entity;

[0081] Annotate structured data formats as named entity data.

[0082] Preferably, the performance evaluation of the named entity data in S6 is performed, and the knowledge enhancement model is optimized and adjusted according to the performance evaluation result, including:

[0083] Generate a standard evaluation data set of text to be recognized;

[0084] Match the named entity data with the generated standard evaluation dataset and compare the matched data;

[0085] Perform performance evaluation based on the comparison results, including accuracy evaluation, recall evaluation, and F1 score evaluation;

[0086] Optimize the knowledge enhancement model according to the evaluation results;

[0087] The optimized knowledge enhancement model is validated using the cross-validation method;

[0088] After model verification is completed, the standard knowledge-enhanced training model for named entity recognition is obtained.

[0089] A named entity recognition system based on a knowledge-enhanced pre-training model, comprising:

[0090] Model deployment and application unit, used to:

[0091] Export the standard knowledge-enhanced training model for named entity recognition to a standard model format;

[0092] Integrate the exported standard knowledge-enhanced training model into a target application terminal, wherein the target application terminal includes an interpreter, a deep learning framework, a CUDA library, and a driver;

[0093] After the standard knowledge-enhanced training model is integrated into the target application terminal, load testing is performed, and key indicators are monitored based on the load test results;

[0094] And generate monitoring logs based on key indicator monitoring data;

[0095] Improvements to the standard knowledge enhancement training model are made based on the generated monitoring logs and user feedback on the target application terminal. The improvement projects include retraining the model and adjusting the recognition rules.

[0096] Compared with the prior art, the present invention has the following beneficial effects:

[0097] 1. The present invention provides a method and system for named entity recognition based on a knowledge-enhanced pre-training model. By constructing a knowledge graph, knowledge resources can be managed and utilized more effectively, and the efficiency of data analysis can be improved. By formulating standards for training models, especially emphasizing generalization ability, the selected training model can show stronger adaptability when processing different fields or different data sets, which can reduce the risk of overfitting the model on a specific data set and improve its recognition performance on other data sets.

[0098] 2. The present invention provides a method and system for named entity recognition based on a knowledge-enhanced pre-training model. By introducing the knowledge-enhanced pre-training model, the model can learn more prior knowledge and semantic information, thereby improving the generalization ability of unseen text data. Through a detailed performance evaluation and optimization process, the interpretability of the model is enhanced, which helps to understand the working mechanism of the model, thereby better adjusting and optimizing the model. The optimized model also has improved generalization ability and can better adapt to different application scenarios and data sets.

[0099] 3. The present invention provides a method and system for named entity recognition based on a knowledge-enhanced pre-training model. The generation of monitoring logs makes performance analysis and problem tracking more convenient, helps to continuously optimize model performance, and can improve the standard knowledge-enhanced training model based on user feedback on the target application terminal. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1A schematic diagram of the steps of named entity recognition of the knowledge enhancement pre-training model of the present invention;

[0101] Figure 2 A schematic diagram of the named entity recognition process of the knowledge enhancement pre-training model of the present invention. DETAILED DESCRIPTION

[0102] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0103] In order to solve the problem in the prior art that there is no targeted graph construction and model selection based on text data, which leads to reduced accuracy when training models with text data, please refer to Figure 1 and Figure 2 , this embodiment provides the following technical solutions:

[0104] A method for named entity recognition based on a knowledge-enhanced pre-training model comprises the following steps:

[0105] S1: Text data collection and processing: collect text data from the database, process the text data after collection, and obtain the target text data after data processing;

[0106] Among them, by retrieving named entity recognition rules from the database, customized recognition can be performed for the target field to improve the accuracy and pertinence of recognition;

[0107] S2: Knowledge base construction: Confirm the knowledge information source from the knowledge base, construct the knowledge graph with the confirmed knowledge information source and the target text data, and obtain the target knowledge graph data after the knowledge graph is constructed;

[0108] Among them, the construction of the knowledge graph further enhances the model's reasoning ability in complex scenarios, enabling the model to better understand and interpret the implicit information in text data;

[0109] S3: Training model selection and processing: selecting a training model according to the named entity recognition rule, and adjusting the selected training model to obtain a recognition training model after adjustment;

[0110] Among them, by carefully selecting and adjusting the training model, it can be ensured that the model has higher accuracy, recall rate and F1 value when recognizing named entities;

[0111] S4: Knowledge enhancement model training: embed the target knowledge graph data into the recognition training model, and perform model training on the recognition training module embedded with the target knowledge graph data. After model training, the knowledge enhancement model is obtained.

[0112] Among them, with the help of semantic information and feature representation of knowledge graph, the model can more accurately identify named entities in text;

[0113] S5: Named entity recognition: input the text to be recognized into the knowledge enhancement model for named entity simulation recognition, and perform recognition result processing on the named entity simulation recognition result to obtain named entity data after the recognition result processing;

[0114] Among them, the effect of named entity recognition can be further improved by adjusting the content and structure of the knowledge graph and optimizing the parameters of the knowledge enhancement model;

[0115] S6: Enhanced model evaluation and optimization: Perform performance evaluation on named entity data and optimize the knowledge enhancement model based on the performance evaluation results;

[0116] Among them, through a detailed performance evaluation and optimization process, the interpretability of the model is enhanced, which helps to understand the working mechanism of the model and thus better adjust and optimize the model.

[0117] In S1, text data is collected from the database and data processing is performed after the text data is collected, including:

[0118] Collect text data from the database using the data access interface;

[0119] Clean the collected text data and standardize the data after cleaning;

[0120] The data related to the target domain in the text data is screened according to the named entity recognition rules, and the named entity recognition rules are retrieved from the database;

[0121] And mark the screened data related to the target field;

[0122] After data labeling is completed, the target text data is obtained.

[0123] Specifically, collecting text data from a database through a data access interface can greatly improve the efficiency of data collection and reduce the time and cost of manual collection. Databases usually have powerful data storage and retrieval capabilities, making the data collection process faster and more accurate. Data cleaning can remove noise and redundant information in the original data and improve data quality. Data standardization ensures the consistency and comparability of the data, providing a reliable foundation for subsequent data analysis and named entity recognition. Using a pre-trained named entity recognition model, named entities in the text, such as names of people, places, and organizations, can be automatically identified. By retrieving named entity recognition rules from the database, customized recognition can be performed for the target field to improve the accuracy and pertinence of recognition. By screening text data according to named entity recognition rules, data related to the target field can be quickly located.

[0124] Confirm the knowledge information source from the knowledge base in S2, and construct a knowledge graph with the confirmed knowledge information source and the target text data, including:

[0125] Confirm the knowledge information source from the knowledge base, and extract the key entities, relationships and attribute data in the knowledge information source using natural language processing technology;

[0126] Use named entity recognition technology to identify entities and concepts in target text data;

[0127] Align the entities and concepts identified in the target text data with the key entities, relations, and attribute data extracted from the knowledge information source;

[0128] After data alignment, the mapping relationship between entities is established;

[0129] Use knowledge graph construction technology to construct knowledge graphs of entities and mapping relationships;

[0130] Conduct a quality assessment on the constructed knowledge graph and optimize it based on the assessment results. The optimization includes adding missing entities and relationships, correcting erroneous data, and removing redundant information.

[0131] The constructed knowledge graph is optimized to obtain the target knowledge graph data.

[0132] Specifically, through the knowledge-enhanced pre-training model, the model can learn richer semantic information and contextual relationships, so as to more accurately identify named entities in the target text data. The entities and relationships in the knowledge graph provide the model with additional contextual information, which helps the model make more accurate judgments when identifying entities. Natural language processing technology is used to extract key entities, relationships, and attribute data from knowledge information sources. It has a high degree of automation, reduces manual intervention, and improves construction efficiency. Through data alignment and the establishment of mapping relationships, the accuracy of entities and relationships in the knowledge graph is ensured, and the reliability of the knowledge graph is enhanced. The knowledge-enhanced pre-training model can handle more complex The model can perform semantic understanding and reasoning tasks, and thus excels in processing target text data containing complex semantic relationships. The construction of the knowledge graph further enhances the model's reasoning ability in complex scenarios, enabling the model to better understand and interpret implicit information in text data. After the knowledge graph was constructed, a quality assessment was conducted, and the graph was optimized based on the evaluation results, including adding missing entities and relationships, correcting erroneous data, and removing redundant information. This helps to ensure the accuracy and completeness of the knowledge graph, and provides a reliable foundation for subsequent knowledge application and analysis. By constructing a knowledge graph, knowledge resources can be managed and utilized more effectively, and the efficiency of data analysis can be improved.

[0133] Specifically, set the refresh frequency of the target knowledge graph data, including:

[0134] Extract the number of entities in the target knowledge graph data;

[0135] Extract the number of relationships corresponding to each entity;

[0136] Extract the correlation coefficient corresponding to each relationship of each entity;

[0137] The relationship strength coefficient corresponding to each entity is obtained by using the number of relationships corresponding to each entity and the association coefficient corresponding to each relationship of each entity;

[0138] The relationship strength coefficient is obtained by the following formula:

[0139]

[0140] Among them, S represents the relationship strength coefficient; n represents the number of relationships corresponding to each entity; G i represents the correlation coefficient of the i-th relationship corresponding to each entity; N represents the total number of relationships contained in the target knowledge graph data; G c Indicates the preset reference value of the correlation coefficient; G max Indicates the maximum value of the correlation coefficient corresponding to each entity; G zmax Represents the maximum value of the correlation coefficient contained in the target knowledge graph data; G zIndicates the middle value of the correlation coefficient contained in the target knowledge graph data;

[0141] comparing the relationship strength coefficient with a preset strength coefficient threshold;

[0142] When the number of entities whose relationship strength coefficient exceeds a preset strength coefficient threshold exceeds a preset number threshold, the refresh frequency of the target knowledge graph data is set, and the target knowledge graph data is refreshed according to the set refresh frequency;

[0143] When the number of entities whose relationship strength coefficient exceeds the preset strength coefficient threshold does not exceed the preset number threshold, the target knowledge graph data is refreshed using the preset initial refresh frequency.

[0144] The technical effect of the above technical solution is: by extracting the number of entities of the target knowledge graph data, the number of relationships of each entity, and the correlation coefficient of each relationship, the solution can dynamically adjust the refresh frequency of the knowledge graph based on the actual complexity and activity of the data. This dynamic adjustment ensures the timeliness and accuracy of the knowledge graph, which is particularly important when the data changes rapidly or a specific field requires high-frequency updates. By calculating the relationship strength coefficient and comparing it with the preset strength coefficient threshold, the solution can identify which entities and their relationships have high importance or activity in the knowledge graph. When the number of these important entities exceeds a certain threshold, the refresh frequency is increased, thereby avoiding unnecessary frequent updates and effectively saving computing resources and storage resources. The solution calculates the relationship strength coefficient by comprehensively considering the number of relationships and the correlation coefficient, which not only considers the direct association of the entity, but also reflects the strength and importance of the relationship through the difference in the correlation coefficient. This refined processing helps to more accurately evaluate the timeliness and reliability of information in the knowledge graph, thereby providing users with higher quality knowledge services. The formulas and thresholds in the solution can be adjusted and optimized according to the actual application scenario. For example, parameters such as the association coefficient reference value, strength coefficient threshold, and number threshold can be adjusted according to the characteristics of knowledge graphs in different fields to adapt to changes in demand in different scenarios. This flexibility and scalability enable the solution to be widely used in various knowledge graph management and maintenance scenarios. By regularly refreshing and updating knowledge graph data, the solution helps maintain the timeliness and accuracy of the knowledge graph, thereby promoting its continuous optimization and development. This is of great significance for improving the application effect of knowledge graphs in the fields of intelligent question answering, recommendation systems, semantic search, etc.

[0145] In summary, this technical solution improves resource utilization efficiency, enhances the reliability and accuracy of the knowledge graph by dynamically adjusting the refresh frequency of the knowledge graph, and demonstrates good flexibility and scalability, providing strong support for the continuous optimization and application of the knowledge graph.

[0146] Specifically, when the number of entities whose relationship strength coefficient exceeds a preset strength coefficient threshold exceeds a preset number threshold, the refresh frequency of the target knowledge graph data is set, including:

[0147] Entities exceeding a preset strength coefficient threshold are regarded as target entities;

[0148] Entities that do not exceed a preset intensity coefficient threshold are regarded as observed entities;

[0149] Extracting the relationship strength coefficient corresponding to the target entity;

[0150] Extracting the relationship strength coefficient corresponding to the observed entity;

[0151] Obtaining a frequency adjustment coefficient using a relationship strength coefficient corresponding to the target entity and a relationship strength coefficient corresponding to the observed entity;

[0152] The frequency adjustment coefficient is obtained by the following formula:

[0153]

[0154] Where, J represents the frequency adjustment coefficient; S b Represents the standard deviation of the relationship strength coefficients corresponding to all entities in the target knowledge graph data; S mb represents the standard deviation of the relationship strength coefficient corresponding to the target entity; S kb represents the standard deviation of the relationship strength coefficient corresponding to the observed entity; f represents the adjustment factor, and the adjustment factor is obtained by the following formula:

[0155]

[0156] Where f represents the adjustment factor; m represents the number of target entities; k represents the number of observed entities; S gi represents the intensity coefficient corresponding to the i-th observed entity; S mi represents the intensity coefficient corresponding to the i-th target entity; S y Indicates the preset intensity coefficient threshold; S b Represents the standard deviation of the relationship strength coefficients corresponding to all entities in the target knowledge graph data; S mb represents the standard deviation of the relationship strength coefficient corresponding to the target entity; S kb represents the standard deviation of the relationship strength coefficient corresponding to the observed entity; e represents the number of all entities contained in the target knowledge graph data;

[0157] The refresh frequency of the target knowledge graph data is set by using the frequency adjustment coefficient in combination with the preset initial refresh frequency;

[0158] The refresh frequency after setting is obtained by the following formula:

[0159]

[0160] Among them, F t represents the refresh frequency; F0 represents the preset initial refresh frequency; J represents the frequency adjustment coefficient.

[0161] The technical effect of the above technical solution is: by extracting the number of entities of the target knowledge graph data, the number of relationships of each entity, and the correlation coefficient of each relationship, the solution can dynamically adjust the refresh frequency of the knowledge graph based on the actual complexity and activity of the data. This dynamic adjustment ensures the timeliness and accuracy of the knowledge graph, which is particularly important when the data changes rapidly or a specific field requires high-frequency updates. By calculating the relationship strength coefficient and comparing it with the preset strength coefficient threshold, the solution can identify which entities and their relationships have high importance or activity in the knowledge graph. When the number of these important entities exceeds a certain threshold, the refresh frequency is increased, thereby avoiding unnecessary frequent updates and effectively saving computing resources and storage resources. The solution calculates the relationship strength coefficient by comprehensively considering the number of relationships and the correlation coefficient, which not only considers the direct association of the entity, but also reflects the strength and importance of the relationship through the difference in the correlation coefficient. This refined processing helps to more accurately evaluate the timeliness and reliability of information in the knowledge graph, thereby providing users with higher quality knowledge services. The formulas and thresholds in the solution can be adjusted and optimized according to the actual application scenario. For example, parameters such as the association coefficient reference value, strength coefficient threshold, and number threshold can be adjusted according to the characteristics of knowledge graphs in different fields to adapt to changes in demand in different scenarios. This flexibility and scalability enable the solution to be widely used in various knowledge graph management and maintenance scenarios. By regularly refreshing and updating knowledge graph data, the solution helps maintain the timeliness and accuracy of the knowledge graph, thereby promoting its continuous optimization and development. This is of great significance for improving the application effect of knowledge graphs in the fields of intelligent question answering, recommendation systems, semantic search, etc.

[0162] In summary, this technical solution improves resource utilization efficiency, enhances the reliability and accuracy of the knowledge graph by dynamically adjusting the refresh frequency of the knowledge graph, and demonstrates good flexibility and scalability, providing strong support for the continuous optimization and application of the knowledge graph.

[0163] Select a training model based on the named entity recognition rules in S3 and adjust the selected training model, including:

[0164] Retrieving named entity recognition rules from a database and confirming recognition requirements in the named entity recognition rules, the recognition requirements including entity type definition, recognition priority, and context dependency;

[0165] The standards for training models are formulated according to the recognition requirements. The standards for training models include the model architecture, learning ability, generalization ability, and the ability to integrate domain knowledge.

[0166] Select a training model based on the established training model standards, including BERT, RoBERTa, or ALBERT;

[0167] Perform model adjustment on the selected training model according to the established training model standards;

[0168] Model adjustment includes architecture adjustment, parameter adjustment and recognition layer adjustment;

[0169] After the model adjustment is completed, the recognition training model is obtained.

[0170] Specifically, by retrieving the named entity recognition rules from the database and confirming the recognition requirements (such as entity type definition, recognition priority and context dependency), it is possible to ensure that these rules are fully considered and utilized when training the model, thereby improving the accuracy and efficiency of recognition. According to the recognition requirements, the standards for training models are formulated, including the model's architecture, learning ability, generalization ability and the ability to integrate domain knowledge, which helps to ensure that the selected training model (such as BERT, RoBERTa or ALBERT) can meet specific recognition requirements. The mentioned model adjustments (architecture adjustment, parameter adjustment and recognition layer adjustment) provide great flexibility and scalability for the training model. This can fine-tune or optimize the model according to the actual application scenarios and requirements to achieve better recognition results. By formulating the standards for training models, especially emphasizing generalization ability, it helps the selected training model to show stronger adaptability when dealing with different fields or different data sets. This can reduce the risk of overfitting the model on a specific data set and improve its recognition performance on other data sets. By carefully selecting and adjusting the training model, it can be ensured that the model has higher accuracy, recall rate and F1 value when recognizing named entities. This is of great significance to improving the overall effect of named entity recognition.

[0171] In order to solve the problem in the existing technology that the constructed model and text data are not effectively embedded, resulting in the model being unable to perform better named entity recognition based on the text content, please refer to Figure 1 and Figure 2 , this embodiment provides the following technical solutions:

[0172] In S4, the target knowledge graph data is embedded into the recognition training model, and the recognition training module embedded with the target knowledge graph data is trained, including:

[0173] Calculate entity vectors of entities and relationships of target knowledge graph data;

[0174] After the entity vector is calculated, the vector representation of the target knowledge graph data is obtained;

[0175] Add a knowledge embedding layer to the recognition training model, where the knowledge embedding layer includes an input layer or an intermediate layer;

[0176] Integrate the vector representation of the target knowledge graph data into the knowledge embedding layer added by the recognition training model;

[0177] Adjust the model parameters of the recognition training model after the vector representation integration, including the learning rate, batch size and regularization term;

[0178] After the model parameters are adjusted, the recognition training model integrated with the vector representation is used for model training, and the change of the loss function is monitored during the model training process;

[0179] Adjust the structure of the model's knowledge embedding layer based on the model training results;

[0180] Finally, a knowledge-enhanced model is obtained.

[0181] Specifically, by calculating the entity vectors of entities and relationships in the target knowledge graph data, the model can more effectively understand and utilize the semantic information in the graph. The entities and relationships in the knowledge graph provide the model with rich background knowledge and context, which helps the model to more accurately identify named entities in the text. After embedding the knowledge graph data, the model can learn more features about entities and relationships, thereby enhancing its ability to process unseen text data. By adjusting model parameters (such as learning rate, batch size, and regularization terms), the generalization performance of the model can be further optimized, making its performance more stable in different fields and scenarios. By adding a knowledge embedding layer to the recognition training model, the position of the knowledge embedding layer (input layer or intermediate layer) can be flexibly adjusted according to actual task requirements, thereby optimizing the model structure. By monitoring the changes in the loss function and adjusting the structure of the knowledge embedding layer according to the model training results, the model performance can be gradually optimized to make it more adaptable to the task requirements of named entity recognition. With the help of the semantic information and feature representation of the knowledge graph, the model can more accurately identify named entities in the text, including names of people, places, and organizations.

[0182] In S5, the text to be recognized is input into the knowledge enhancement model for named entity simulation recognition, and the named entity simulation recognition result is processed, including:

[0183] Retrieve the text to be recognized from the database, where the text to be recognized is simulated recognition text data for named entity recognition;

[0184] Use the word segmenter to segment and tokenize the text to be recognized;

[0185] Use entity linking technology to match entities in the text to be recognized after word segmentation and tokenization with entities in the target knowledge graph data, and perform format conversion after matching;

[0186] The text to be recognized after format conversion is input into the knowledge enhancement model, and the knowledge enhancement model uses the knowledge embedding layer to infer the input text to be recognized;

[0187] The knowledge enhancement model generates an output tensor for the inferred text to be recognized. The generated output tensor contains the classification score of each input token.

[0188] Based on the classification scores in the output tensor, an activation function is used to predict the entity category of each word in the output tensor;

[0189] Extract named entities in the text to be recognized based on the prediction results, identify the starting and ending positions of the extracted named entities, and correspond the identified positions to the entity types;

[0190] Organize the corresponding completed text to be recognized into a structured data format, including the text, category, location and confidence of the named entity;

[0191] Annotate structured data formats as named entity data.

[0192] Specifically, through entity linking technology, entities in the text to be identified are matched with entities in the target knowledge graph, which can make full use of the rich information in the knowledge graph and improve the accuracy and depth of named entity recognition. The knowledge enhancement model uses the knowledge embedding layer for reasoning, which helps the model understand the contextual information of the text and more accurately identify named entities. The knowledge embedding layer can introduce external knowledge and improve the model's ability to understand complex semantic relationships. The word segmentation and tokenization processing through the word segmenter, as well as the entity category prediction using the classification score of the output tensor, can refine the processing steps and improve the recognition accuracy. At the same time, the structured data format can clearly display the text, category, location and confidence of the entity, which is convenient for subsequent analysis and processing. It can handle different types of text to be identified and adapt to different knowledge graphs. By adjusting the content and structure of the knowledge graph and optimizing the parameters of the knowledge enhancement model, the effect of named entity recognition can be further improved, and the process of named entity recognition can be decomposed into multiple steps, each of which has a clear function and output. This modular design makes the solution easy to implement and maintain, and also facilitates troubleshooting and performance optimization. By introducing the knowledge-enhanced pre-training model, the model can learn more prior knowledge and semantic information, thereby improving the generalization ability of unseen text data. This helps the model maintain a high recognition accuracy when processing new fields or new types of text.

[0193] The performance of named entity data in S6 is evaluated, and the knowledge enhancement model is optimized and adjusted according to the performance evaluation results, including:

[0194] Generate a standard evaluation data set of text to be recognized;

[0195] Match the named entity data with the generated standard evaluation dataset and compare the matched data;

[0196] Perform performance evaluation based on the comparison results, including accuracy evaluation, recall evaluation, and F1 score evaluation;

[0197] Optimize the knowledge enhancement model according to the evaluation results;

[0198] The optimized knowledge enhancement model is validated using the cross-validation method;

[0199] After model verification is completed, the standard knowledge-enhanced training model for named entity recognition is obtained.

[0200] Specifically, by generating a standard evaluation dataset, the consistency and fairness of the evaluation are ensured. This helps to accurately compare the performance of the model at different times and under different conditions, thereby avoiding subjectivity and uncertainty in the evaluation process. The data matching process between the named entity data and the standard evaluation dataset can accurately identify the performance of the model in identifying specific named entities. By comparing the matched data, we can intuitively see in which aspects the model is deficient, providing a clear direction for subsequent optimization. The use of multiple indicators such as accuracy evaluation, recall evaluation and F1 score evaluation can fully reflect the performance of the model in the named entity recognition task. These indicators complement each other and together constitute a complete evaluation of the model performance, which helps to discover the advantages and disadvantages of the model in different aspects. According to the performance evaluation results, the knowledge enhancement model can be optimized and adjusted in a targeted manner. This optimization is targeted and directional, and can significantly improve the performance of the model. At the same time, since the optimization is based on actual evaluation data, the optimized model is more adaptable to actual application scenarios. Using the cross-validation method to verify the optimized model can further ensure the stability and reliability of the model. The cross-validation method can fully evaluate the performance of the model on different data sets through multiple training and testing, thereby avoiding the problem of overfitting or underfitting. Through the detailed performance evaluation and optimization process, the interpretability of the model is enhanced, which helps to understand the working mechanism of the model and thus better adjust and optimize the model. At the same time, the optimized model also improves its generalization ability and can better adapt to different application scenarios and data sets.

[0201] In order to solve the problem in the prior art that the completed training model is not monitored in real time on the application terminal, which makes it impossible to further understand the actual application of the training model, please refer to Figure 1 and Figure 2 , this embodiment provides the following technical solutions:

[0202] A named entity recognition system based on a knowledge-enhanced pre-training model, comprising:

[0203] Model deployment and application unit, used to:

[0204] Export the standard knowledge-enhanced training model for named entity recognition to a standard model format;

[0205] Integrate the exported standard knowledge-enhanced training model into a target application terminal, wherein the target application terminal includes an interpreter, a deep learning framework, a CUDA library, and a driver;

[0206] After the standard knowledge-enhanced training model is integrated into the target application terminal, load testing is performed, and key indicators are monitored based on the load test results;

[0207] And generate monitoring logs based on key indicator monitoring data;

[0208] Improvements to the standard knowledge enhancement training model are made based on the generated monitoring logs and user feedback on the target application terminal. The improvement projects include retraining the model and adjusting the recognition rules.

[0209] Specifically, the standard knowledge-enhanced training model for named entity recognition is first exported to a standard model format, which is conducive to the versatility and portability of the model. The exported standard model can be easily integrated into various target application terminals, including those equipped with interpreters, deep learning frameworks, CUDA libraries, and drivers. This integration method simplifies the deployment process and lowers the technical threshold. After the model is integrated into the target application terminal, load testing is performed, which helps to evaluate the performance of the model in the actual application environment. Through key indicator monitoring, the system can promptly discover performance bottlenecks or potential problems and provide data support for subsequent model improvements. The generation of monitoring logs makes performance analysis and problem tracking more convenient, which helps to continuously optimize model performance and improve the standard knowledge-enhanced training model based on user feedback on the target application terminal. This user-centric approach ensures that the model can better adapt to actual application needs. Improvement projects include retraining the model and adjusting the recognition rules, which help to improve the accuracy and generalization of the model. The knowledge-enhanced pre-training model enhances the semantic understanding ability of the model by introducing an external knowledge base, thereby improving the accuracy of named entity recognition. The pre-training process enables the model to learn common features on a large amount of unlabeled data, and then fine-tune it on specific tasks, thereby improving the adaptability and efficiency of the model.

[0210] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0211] While the embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that many changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the invention.

Claims

1. A method for named entity recognition based on a knowledge-enhanced pre-training model, characterized in that: The steps include: S1: Text data collection and processing: collect text data from the database, process the text data after collection, and obtain the target text data after data processing; S2: Knowledge base construction: Confirm the knowledge information source from the knowledge base, construct the knowledge graph with the confirmed knowledge information source and the target text data, and obtain the target knowledge graph data after the knowledge graph is constructed; Setting the refresh frequency of the target knowledge graph data includes: obtaining the relationship strength coefficient corresponding to each entity using the number of relationships corresponding to each entity in the target knowledge graph data and the association coefficient corresponding to each relationship of each entity: Among them, S represents the relationship strength coefficient; n represents the number of relationships corresponding to each entity; G i represents the correlation coefficient of the i-th relationship corresponding to each entity; N represents the total number of relationships contained in the target knowledge graph data; G c Indicates the preset reference value of the correlation coefficient; G max Indicates the maximum value of the correlation coefficient corresponding to each entity; G zmax Represents the maximum value of the correlation coefficient contained in the target knowledge graph data; G z Indicates the middle value of the correlation coefficient contained in the target knowledge graph data; comparing the relationship strength coefficient with a preset strength coefficient threshold to determine a refresh frequency; S3: Training model selection and processing: selecting a training model according to the named entity recognition rule, and adjusting the selected training model to obtain a recognition training model after adjustment; S4: Knowledge enhancement model training: embed the target knowledge graph data into the recognition training model, and perform model training on the recognition training model embedded with the target knowledge graph data. After model training, the knowledge enhancement model is obtained. S5: Named entity recognition: input the text to be recognized into the knowledge enhancement model for named entity simulation recognition, and perform recognition result processing on the named entity simulation recognition result to obtain named entity data after the recognition result processing; S6: Enhanced model evaluation and optimization: Perform performance evaluation on named entity data and optimize the knowledge enhancement model based on the performance evaluation results.

2. According to claim 1, a method for named entity recognition based on a knowledge-enhanced pre-training model is characterized in that: In S1, text data is collected from the database and data processing is performed after the text data is collected, including: Collect text data from the database using the data access interface; Clean the collected text data and standardize the data after cleaning; The data related to the target domain in the text data is screened according to the named entity recognition rules, and the named entity recognition rules are retrieved from the database; And mark the screened data related to the target field; After data labeling is completed, the target text data is obtained.

3. According to claim 2, a method for named entity recognition based on a knowledge-enhanced pre-training model is characterized in that: Confirm the knowledge information source from the knowledge base in S2, and construct a knowledge graph with the confirmed knowledge information source and the target text data, including: Confirm the knowledge information source from the knowledge base, and extract the key entities, relationships and attribute data in the knowledge information source using natural language processing technology; Use named entity recognition technology to identify entities and concepts in target text data; Align the entities and concepts identified in the target text data with the key entities, relations, and attribute data extracted from the knowledge information source; After data alignment, the mapping relationship between entities is established; Use knowledge graph construction technology to construct knowledge graphs of entities and mapping relationships; Conduct a quality assessment on the constructed knowledge graph and optimize it based on the assessment results. The optimization includes adding missing entities and relationships, correcting erroneous data, and removing redundant information. The constructed knowledge graph is optimized to obtain the target knowledge graph data.

4. A method for named entity recognition based on a knowledge-enhanced pre-training model according to claim 3, characterized in that: Set the refresh frequency of the target knowledge graph data, including: When the number of entities whose relationship strength coefficient exceeds a preset strength coefficient threshold exceeds a preset number threshold, the refresh frequency of the target knowledge graph data is set, and the target knowledge graph data is refreshed according to the set refresh frequency; When the number of entities whose relationship strength coefficient exceeds the preset strength coefficient threshold does not exceed the preset number threshold, the target knowledge graph data is refreshed using the preset initial refresh frequency.

5. A method for named entity recognition based on a knowledge-enhanced pre-training model according to claim 4, characterized in that: When the number of entities whose relationship strength coefficient exceeds the preset strength coefficient threshold exceeds the preset number threshold, the refresh frequency of the target knowledge graph data is set, including: Entities exceeding a preset strength coefficient threshold are regarded as target entities; Entities that do not exceed a preset intensity coefficient threshold are regarded as observed entities; Extracting the relationship strength coefficient corresponding to the target entity; Extracting the relationship strength coefficient corresponding to the observed entity; Obtaining a frequency adjustment coefficient using a relationship strength coefficient corresponding to the target entity and a relationship strength coefficient corresponding to the observed entity; The frequency adjustment coefficient is obtained by the following formula: Where, J represents the frequency adjustment coefficient; S b Represents the standard deviation of the relationship strength coefficients corresponding to all entities in the target knowledge graph data; S mb represents the standard deviation of the relationship strength coefficient corresponding to the target entity; S kb represents the standard deviation of the relationship strength coefficient corresponding to the observed entity; f represents the adjustment factor, and the adjustment factor is obtained by the following formula: Where f represents the adjustment factor; m represents the number of target entities; k represents the number of observed entities; S gi represents the intensity coefficient corresponding to the i-th observed entity; S mi represents the intensity coefficient corresponding to the i-th target entity; S y Indicates the preset intensity coefficient threshold; S b Represents the standard deviation of the relationship strength coefficients corresponding to all entities in the target knowledge graph data; S mb represents the standard deviation of the relationship strength coefficient corresponding to the target entity; S kb represents the standard deviation of the relationship strength coefficient corresponding to the observed entity; e represents the number of all entities contained in the target knowledge graph data; The refresh frequency of the target knowledge graph data is set by using the frequency adjustment coefficient in combination with the preset initial refresh frequency; The refresh frequency after setting is obtained by the following formula: Among them, F t represents the refresh frequency; F0 represents the preset initial refresh frequency; J represents the frequency adjustment coefficient.

6. The method for named entity recognition based on a knowledge-enhanced pre-training model according to claim 3, characterized in that: Select a training model based on the named entity recognition rules in S3 and adjust the selected training model, including: Retrieving named entity recognition rules from a database and confirming recognition requirements in the named entity recognition rules, the recognition requirements including entity type definition, recognition priority, and context dependency; The standards for training models are formulated according to the recognition requirements. The standards for training models include the model architecture, learning ability, generalization ability, and the ability to integrate domain knowledge. Select a training model based on the established training model standards, including BERT, RoBERTa, or ALBERT; Perform model adjustment on the selected training model according to the established training model standards; Model adjustment includes architecture adjustment, parameter adjustment and recognition layer adjustment; After the model adjustment is completed, the recognition training model is obtained.

7. A method for named entity recognition based on a knowledge-enhanced pre-training model according to claim 6, characterized in that: In S4, the target knowledge graph data is embedded into the recognition training model, and the recognition training model embedded with the target knowledge graph data is trained, including: Calculate entity vectors of entities and relationships of target knowledge graph data; After the entity vector is calculated, the vector representation of the target knowledge graph data is obtained; Add a knowledge embedding layer to the recognition training model, where the knowledge embedding layer includes an input layer or an intermediate layer; Integrate the vector representation of the target knowledge graph data into the knowledge embedding layer added by the recognition training model; Adjust the model parameters of the recognition training model after the vector representation integration, including the learning rate, batch size and regularization term; After the model parameters are adjusted, the recognition training model integrated with the vector representation is used for model training, and the change of the loss function is monitored during the model training process; Adjust the structure of the model's knowledge embedding layer based on the model training results; Finally, a knowledge-enhanced model is obtained.

8. A method for named entity recognition based on a knowledge-enhanced pre-training model according to claim 7, characterized in that: In S5, the text to be recognized is input into the knowledge enhancement model for named entity simulation recognition, and the named entity simulation recognition result is processed, including: Retrieve the text to be recognized from the database, where the text to be recognized is simulated recognition text data for named entity recognition; Use the word segmenter to segment and tokenize the text to be recognized; Use entity linking technology to match entities in the text to be recognized after word segmentation and tokenization with entities in the target knowledge graph data, and perform format conversion after matching; The text to be recognized after format conversion is input into the knowledge enhancement model, and the knowledge enhancement model uses the knowledge embedding layer to infer the input text to be recognized; The knowledge enhancement model generates an output tensor for the inferred text to be recognized. The generated output tensor contains the classification score of each input token. Based on the classification scores in the output tensor, an activation function is used to predict the entity category of each word in the output tensor; Extract named entities in the text to be recognized based on the prediction results, identify the starting and ending positions of the extracted named entities, and correspond the identified positions to the entity types; Organize the corresponding completed text to be recognized into a structured data format, including the text, category, location and confidence of the named entity; Annotate structured data formats as named entity data.

9. A method for named entity recognition based on a knowledge-enhanced pre-training model according to claim 8, characterized in that: The performance of named entity data in S6 is evaluated, and the knowledge enhancement model is optimized and adjusted according to the performance evaluation results, including: Generate a standard evaluation data set of text to be recognized; Match the named entity data with the generated standard evaluation dataset and compare the matched data; Perform performance evaluation based on the comparison results, including accuracy evaluation, recall evaluation, and F1 score evaluation; Optimize the knowledge enhancement model according to the evaluation results; The optimized knowledge enhancement model is validated using the cross-validation method; After model verification is completed, the standard knowledge-enhanced training model for named entity recognition is obtained.

10. A named entity recognition system based on a knowledge-enhanced pre-trained model, applied in the method for named entity recognition based on a knowledge-enhanced pre-trained model as claimed in claim 9, characterized in that: include: Model deployment and application unit, used to: Export the standard knowledge-enhanced training model for named entity recognition to a standard model format; Integrate the exported standard knowledge-enhanced training model into a target application terminal, wherein the target application terminal includes an interpreter, a deep learning framework, a CUDA library, and a driver; After the standard knowledge-enhanced training model is integrated into the target application terminal, load testing is performed, and key indicators are monitored based on the load test results; And generate monitoring logs based on key indicator monitoring data; Improvements to the standard knowledge enhancement training model are made based on the generated monitoring logs and user feedback on the target application terminal. The improvement projects include retraining the model and adjusting the recognition rules.

Images