A method for constructing a knowledge extraction platform based on active learning

Abstract

The application discloses a kind of based on active learning's construction method of knowledge extraction platform, and knowledge extraction platform includes label system construction layer, intelligent service driving layer, service capability evaluation layer and knowledge capability evolution layer;The process of construction method is: to original data or annotation data lead-in, data lead-in is in knowledge extraction platform;According to the task of knowledge extraction and the feature selection or newly created editing label system of data source file;The machine labeling of man-in-the-loop and model training release are carried out, and according to original data set and label, annotation data set construction and service release are carried out;Service encapsulation is carried out based on the model trained in step (3), service is exported service installation package, and according to knowledge service demand, service is screened and called.The present application based on active learning carries out secondary annotation to error example sample, improves the accuracy of extraction, solves the problem that the knowledge required when constructing various applications at present is difficult to prepare, the accuracy is low and the utilization is low.

Description

Technical Field

[0001] This invention relates to knowledge extraction platforms in the field of machine learning technology, and more particularly to a method for constructing a knowledge extraction platform based on active learning. Background Technology

[0002] The knowledge required to generate the capabilities of intelligent systems needs to be continuously learned and evolved during application. Current system capabilities are custom-built based on foreseeable needs; once the system is finalized, its capabilities become fixed. Furthermore, the knowledge required for building various applications is currently difficult to prepare, has low accuracy, and is poorly utilized.

[0003] To adapt to the complex and ever-changing nature of future intelligent scenarios, the system needs to have dynamic evolution capabilities. It needs to continuously accumulate new data and learn new knowledge during the application process, gradually optimize various models, and achieve iterative enhancement of core capabilities.

[0004] The delivery model for various intelligent devices in the future will undergo significant changes, transforming from pre-defined and fixed products with pre-set functions and capabilities to dynamic, evolving systems. This will shift the intelligent system's capabilities from "fixed functions, one-time delivery" to "use once, gain intelligence." Therefore, a systematic learning platform needs to be built to achieve the co-construction, sharing, and optimization of intelligent capabilities, providing data, knowledge, algorithms, models, development frameworks, and training environments to incubate intelligent AI and address the low-level duplication of existing decentralized intelligent R&D models. Summary of the Invention

[0005] Purpose of the Invention: To address the problems of insufficient knowledge preparation, low accuracy, and low utilization in building various applications, as well as the difficulty in accurately and quickly generating the heterogeneous knowledge required for system capabilities, this invention proposes a method for constructing a knowledge extraction platform based on active learning. By collecting feedback information from user business operations to determine the level of performance degradation, the method initiates model retraining in a timely manner, selects a training algorithm, and automatically tunes parameters based on the Auto-PyTorch framework. This lowers the technical threshold for developers while enabling the knowledge extraction model to actively learn and adapt to new data, promoting rapid and high-quality learning and growth of intelligent capabilities. This solves the problems of insufficient knowledge preparation, low accuracy, and low utilization in building current knowledge extraction platforms.

[0006] Technical solution: The knowledge extraction platform in the method for constructing a knowledge extraction platform based on active learning of the present invention includes a tag system construction layer, an intelligent service driving layer, a service capability evaluation layer, and a knowledge capability evolution layer;

[0007] The label system construction layer constructs a label system by uniformly accessing and preprocessing the original dataset, selecting the corresponding representation modes of entity knowledge, event knowledge, and rule knowledge based on the characteristics of the original data;

[0008] The intelligent service-driven layer uses the pre-built model in the intelligent pre-labeling to pre-label the original dataset according to the label system built by the label system construction layer. Through active learning, it provides hints for pre-labeled examples with low accuracy and performs difficult example labeling. After the results of the difficult example labeling are fed back, the model is retrained and the intelligent service is encapsulated to iteratively generate intelligent labeling capabilities.

[0009] The service capability assessment layer uses the knowledge service invocation engine to call the services encapsulated in the knowledge-building intelligent service driver layer on demand, extract knowledge, and perform real-time calculation of indicators to comprehensively assess the capabilities of intelligent services.

[0010] The knowledge capability evolution layer provides unified management of knowledge building services and iteratively evolves knowledge service capabilities based on user feedback.

[0011] The construction method of a knowledge extraction platform includes the following steps:

[0012] Step (1) involves importing the original data or labeled data into the knowledge extraction platform by uploading or configuring a database connection, and then converting and saving the data as a text file in UTF-8 format.

[0013] Step (2): Select or create a new editing tag system based on the knowledge extraction task and the characteristics of the data source file;

[0014] Step (3) involves machine annotation of people in the loop and model training and deployment, constructing an annotated dataset and deploying the service based on the original dataset and labels;

[0015] Step (4) Service encapsulation, invocation and iterative evolution: based on the model trained in step (3), service encapsulation is performed, the service is exported as a service installation package, and services are selected and invoked according to knowledge service requirements.

[0016] In step (2), based on the knowledge extraction task, feature analysis of the text content is selected; the tag system with the highest matching degree is selected as the tag system to be labeled. If the matching degree does not reach the threshold, the tag system is manually created and edited.

[0017] Step (3) includes the following steps:

[0018] Step (3.1) involves constructing the labeled dataset using three pre-annotation methods: first, implicit annotation is performed using logical rules; then, entity relationships are extracted using remote supervision based on existing knowledge graphs; and finally, pre-annotation is performed using existing intelligent pre-annotation services to jointly generate the labeled dataset.

[0019] Step (3.2) involves training the model on the existing labeled data. The binary files generated by the training are used to create and publish intelligent services via container cloud. The published services are added to the pre-labeled service list. The difficult example data called by the services are added to the active learning module for manual labeling. New labeled data and models are generated iteratively.

[0020] In step (4), during the call process, the indicator function is collected and monitored in real time. If the curvature or absolute value of the indicator function's decreasing curve reaches the threshold, the model is retrained and the service is published.

[0021] The intelligent service-driven layer actively learns to provide hints for pre-labeled examples with low accuracy and performs hard example labeling. After the hard example labeling results are fed back, the model is retrained and the intelligent service is encapsulated.

[0022] Model training can be divided into two methods: model initialization training based on full data and incremental learning.

[0023] The intelligent service-driven layer includes an intelligent pre-annotation module, an active learning module, a hard example annotation module, a model training module, and a service encapsulation module;

[0024] The intelligent pre-annotation module performs machine intelligent annotation on existing intelligent models before manual annotation;

[0025] The active learning module filters the data to be labeled;

[0026] After analyzing and identifying difficult cases that are judged incorrectly, the difficult case annotation module prioritizes manual annotation of difficult cases with high error frequency.

[0027] The model training module trains deep learning models on labeled datasets by scheduling hardware cluster resources through the container cloud platform and dynamically allocating resources to start containers for model training.

[0028] The service encapsulation module encapsulates the trained model files into services for external systems to call.

[0029] The service encapsulation module first creates a Django project, which then creates a REST service using Python code. Next, it adds the inference function from the model inference script to the project startup code. Finally, it performs inference calculations and returns the results by calling the model inference function.

[0030] The service capability assessment layer includes a knowledge service invocation engine, an assessment indicator system, and an indicator real-time calculation module;

[0031] The knowledge service call engine automatically matches the required knowledge representation pattern and selects intelligent extraction services under different knowledge types and tag systems.

[0032] The evaluation indicator system constructs different domain indicator libraries according to business types, and displays different types of indicators in order;

[0033] The real-time indicator calculation module performs real-time calculations of indicators based on different indicator systems and the frequency of data collection, and displays the indicator scores according to the keyness ranking of the evaluation indicators.

[0034] The knowledge capability evolution layer includes the knowledge service operation support environment, the evolution trigger engine, and the knowledge service integration framework;

[0035] The knowledge service runtime support environment provides support for the operation of knowledge extraction services in heterogeneous systems.

[0036] The evolution trigger engine re-labels the training dataset based on changes in real-time calculated metric parameter values, and then incrementally trains and publishes it as a service for use.

[0037] The knowledge service integration framework matches and selects the installation package to be deployed and provides the deployment script.

[0038] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0039] (1) This invention is designed to adapt to the complex and ever-changing characteristics of future intelligent scenarios, which require intelligent systems to have dynamic evolution capabilities. It uses active learning to assist in the construction of a knowledge extraction model, extracting and generating various types of knowledge required in the process of generating system capabilities, such as rules, entities, and event knowledge information. For knowledge that is extracted incorrectly, it uses active learning to perform secondary annotation on the incorrect sample, continuously iterates and evolves the model capabilities, and greatly improves the accuracy of extraction.

[0040] (2) This invention extracts knowledge from heterogeneous data, integrates and models three different types of knowledge (rules, entities, and events) to construct three types of knowledge extraction models, builds a knowledge construction loop through human loop, establishes a feedback loop of knowledge usage frequency and accuracy, constructs knowledge based on the continuous evolution system of active learning, stores various types of knowledge through knowledge engines such as rule engine and knowledge graph, and provides knowledge services for various external systems.

[0041] (3) This invention continuously integrates intelligent systems into application systems, continuously iterates and evolves the system to build knowledge extraction capabilities, and improves knowledge extraction efficiency and utilization. Attached Figure Description

[0042] Figure 1 This is a diagram illustrating the overall architecture of the knowledge extraction platform based on active learning according to the present invention.

[0043] Figure 2 This is a flowchart illustrating the process of modeling different knowledge representations during the tag system construction phase of this invention.

[0044] Figure 3 This is a flowchart illustrating the active learning and annotation process for data in the intelligent service-driven layer of this invention.

[0045] Figure 4 This is a flowchart illustrating the process of calling intelligent services and performing real-time indicator calculations in the generation of system capabilities according to the present invention. Detailed Implementation

[0046] The method for constructing a knowledge extraction platform based on active learning provided in this embodiment of the invention is as follows: Figure 1 As shown, the knowledge extraction platform includes a tag system construction layer, an intelligent service driving layer, a service capability evaluation layer, and a knowledge capability evolution layer.

[0047] The tagging system construction layer uniformly accesses and preprocesses the original dataset, selecting corresponding representation modes for entity-based, event-based, and rule-based knowledge based on data characteristics to construct the tagging system to be labeled. This layer uniformly accesses structured and unstructured data, cleans and replaces it, and then classifies it by knowledge type. It categorizes data types to determine whether they belong to entity-based, event-based, or rule-based knowledge, and identifies the corresponding tagging system to which each data type belongs.

[0048] The entity tagging system includes the type of entity and the types of relationships between entities; the event tagging system includes the initiator, recipient, event type, location, or other elements of the event; the rule tagging system includes rule execution elements, such as parallelism, OR, execution conditions, and execution steps.

[0049] In one implementation of this embodiment, entity-based, event-based, and rule-based knowledge representation modeling is performed after accessing and analyzing the original data. Entity-based knowledge representation describes the characteristics of people, organizations, locations, countries, or other objectively existing objects, and assists in the construction of entity-based knowledge through named entity recognition and entity relationship extraction techniques. Event-based knowledge representation assists in the construction of event-based knowledge by extracting event elements from the text, such as initiator, recipient, time, location, type, and relationships between events, and then extracting event relationships. Tactical rule-based knowledge representation describes the basic elements of entities involved in rule logic and action logic in a standardized manner, and represents them in the form of production rules, state machines, and activity diagrams. The extraction of conditions, actions, and logical controls assists in the construction of rule-based knowledge.

[0050] The intelligent service-driven layer uses the pre-built model in the intelligent pre-annotation to annotate according to the label system. For results with low prediction probability, it actively learns and provides feedback to human annotators to annotate difficult examples. Finally, all the annotated data is unified, trained on the model, and released as a service for use.

[0051] Among them, active learning feedback for manual annotation means that in the case of intelligent pre-annotation, low-probability annotation prompts are manually processed, while machine-predicted labels and corresponding probabilities are provided, making it easier for annotators to select from a smaller candidate set and improving annotation accuracy.

[0052] Model training can be divided into two methods: one is to perform model initialization training based on the full amount of data, completing the model construction work without human participation in annotation; the other is incremental learning, which is to incrementally learn and build the model after adding labeled samples to the hard examples.

[0053] In one implementation of this embodiment, the intelligent service-driven layer includes an intelligent pre-annotation module, an active learning module, a difficult example annotation module, a model training module, and a service encapsulation module.

[0054] The intelligent pre-annotation module is used to assist machine intelligent annotation of existing intelligent models before manual annotation, thereby reducing the workload of manual data annotation. When manual annotation is performed, the machine-assisted annotation is used as a reference for modification and correction. The methods used in intelligent pre-annotation include data augmentation, implicit annotation based on business logic, and intelligent model service-assisted annotation.

[0055] The active learning module filters the labeled data, prioritizing the most informative or valuable data for human annotation. From the model's perspective, the earlier the most needed examples are manually labeled, the better the model's performance will improve after training, and the higher the accuracy of automatic annotation is likely to be.

[0056] The difficult example annotation module analyzes the difficult examples identified as erroneous by the knowledge extraction model based on user operations, sorts them according to the frequency of errors, and prioritizes providing the difficult examples with high error frequencies to human annotators for annotation.

[0057] The model training module uses a pre-built knowledge extraction algorithm to train a deep learning model on the labeled dataset. It schedules hardware cluster resources through a container cloud platform and dynamically allocates resources to start containers for model training.

[0058] The service encapsulation module packages the trained model files into services for external systems to call. First, a standard Django project is created, which then establishes a REST service using Python code. Next, the inference function from the intelligent model inference script is added to the project's startup code. Upon project startup, due to the dependency on the inference script, the code block in the inference script that loads the model is directly executed, thus preloading the intelligent model file. Finally, the model inference function is called to perform inference calculations and return the results.

[0059] In one implementation of this embodiment, the service capability evaluation layer includes a knowledge service invocation engine, an evaluation index system, and an index real-time calculation module.

[0060] The knowledge service retrieval engine automatically matches the required knowledge representation pattern based on different knowledge extraction needs and selects intelligent extraction services under different knowledge types and tag systems.

[0061] The evaluation indicator system first establishes a common indicator evaluation system, which includes accuracy, recall, rogue, bleu, and throughput; then, it constructs different domain-specific indicator libraries according to different business types; finally, it comprehensively sorts and displays the different types of indicators in a key order.

[0062] The real-time indicator calculation module automatically calculates indicators in real time based on different indicator systems and the frequency of data collection, and displays indicator scores according to the keyness ranking of the evaluation indicators.

[0063] In one implementation of this embodiment, the knowledge capability evolution layer includes a knowledge service runtime support environment, an evolution triggering engine, and a knowledge service integration framework.

[0064] The knowledge service runtime support environment provides support for the knowledge extraction service to run in heterogeneous systems. Specifically, it refers to the Docker container image for service deployment, the program for running the service, and the startup scripts that are packaged and exported.

[0065] The evolution trigger engine, based on the changing trend of real-time calculated indicator parameter values, triggers model retraining iteration if the decline reaches a certain threshold. Through the active learning of the intelligent service-driven layer, the training dataset is re-labeled, incrementally trained, and published as a service for use.

[0066] The knowledge service integration framework automatically matches and selects the installation package to be deployed based on the knowledge extraction service requirements and provides deployment scripts.

[0067] The service encapsulation in this invention is not done by manually writing code, but by automatically packaging the corresponding algorithm code using an intelligent service-driven framework and publishing the corresponding REST service based on the Django framework.

[0068] The intelligent service-driven layer pre-labels the original dataset according to the label system through intelligent pre-labeling. Through active learning, it provides prompts for pre-labeled examples with low accuracy and feeds them back to the labelers for difficult example labeling. After the results of difficult example labeling are fed back, the model is trained and the intelligent service is encapsulated, and intelligent labeling capabilities are generated iteratively.

[0069] The service capability assessment layer uses the knowledge service invocation engine to call the services encapsulated in the knowledge-building intelligent service-driven layer on demand for knowledge extraction, and performs real-time calculation of the evaluation index system, namely, calculating the accuracy, recall, and throughput of entity recognition and entity relationship extraction, and comprehensively evaluating the capabilities of intelligent services.

[0070] The service capability assessment layer constructs an assessment index system that includes accuracy, recall, and throughput. During model training, it records and visualizes the trend of index changes in real time. After the service is released, the knowledge service call engine dynamically scales up and down the number of service replicas. Based on the characteristics of business calls and the current service load, it selects the most suitable knowledge service for invocation and records the input and output, response time, and call frequency of the service calls.

[0071] The knowledge capability evolution layer unifies the management of knowledge building services and iteratively evolves knowledge service capabilities based on user feedback.

[0072] In this embodiment, the tag system construction layer uses entity-based knowledge as an example for overall description. Unstructured text data is introduced, and heterogeneous data is stored in files. Based on different knowledge extraction tasks, features of the text content are selected for analysis. Entity-based knowledge involves matching typical entity names in the tag system. The article with the highest matching degree for a certain tag system is selected as the tag system if it reaches a matching degree threshold. If the matching degree does not reach the threshold, the user manually creates and edits the tag system.

[0073] In this embodiment, the intelligent service driving layer marks the original text data according to the label system established by the label system construction layer through intelligent pre-annotation. Specifically, it actively learns to provide prompts for pre-annotated examples with low accuracy and provides feedback to the annotators for difficult example annotation. After the relevant annotation results are fed back, the model is retrained and the intelligent service is encapsulated to iteratively generate intelligent annotation capabilities.

[0074] In this embodiment, the knowledge capability evolution layer provides unified management of knowledge construction services. First, it builds an underlying platform for entity knowledge extraction services based on the knowledge service runtime support environment. Then, it establishes the capability to deploy and build knowledge extraction services through a knowledge service integration framework, enabling one-click deployment and publishing to application systems. Finally, it receives feedback from application systems regarding knowledge capabilities through a knowledge evolution trigger engine. When service capability indicators drop to a certain level, it triggers a knowledge service retraining mechanism to evolve and iterate the model.

[0075] The knowledge capability evolution layer is based on the Kubernetes cloud platform and uses an underlying container cloud to support the operation of knowledge services, scheduling underlying hardware and software resources to build the service environment. All knowledge services are displayed in a unified directory through the knowledge service integration framework, based on the type of business. The knowledge service integration framework automatically creates, starts, and stops knowledge services according to the request volume, achieving integration with user systems.

[0076] The evolution trigger engine of the knowledge capability evolution layer collects the call status of knowledge services and calculates the current usage status of knowledge services according to the set evolution trigger index calculation formula. For example, in terms of throughput and response time of REST services, it evaluates the container service capabilities in real time; in terms of service accuracy, it calculates and collects feedback from users' systems using intelligent service calls to calculate accuracy, recall, and LOSS and other indicators. When the relevant indicators drop to a certain threshold, it triggers retraining or incremental training, thereby achieving the effect of continuous knowledge evolution.

[0077] Combined with appendix Figure 1 The overall process of this invention will be described below. The steps of the method for constructing a knowledge extraction platform based on active learning according to this invention are as follows:

[0078] Step (1) involves importing the original or labeled data into the knowledge extraction platform by uploading or configuring a database connection, and then converting and saving the data as a UTF-8 format text file.

[0079] Step (2): Select or create a new editing tag system based on the knowledge extraction task and the characteristics of the data source file, such as... Figure 2As shown, the specific steps are as follows:

[0080] Step (2.1) involves analyzing the features of the text content based on different knowledge extraction tasks. Entity-based knowledge involves matching typical entity names in the tag system; event-based knowledge involves determining the type of events in the tag system; and rule-based knowledge is mapped to a specific domain and operated using atomic interfaces within that domain. The tag system with the highest matching degree for a given tag system is selected as the labeled tag system. If the matching degree reaches a threshold, it is used as the tag system; otherwise, the user manually creates and edits the tag system.

[0081] Step (3) involves machine annotation of the human-in-the-loop model and its training and deployment. This includes constructing an annotated dataset and deploying the service based on the original dataset and labels. Figure 3 As shown, the specific steps are as follows:

[0082] Step (3.1) involves constructing the labeled dataset using three pre-annotation methods: first, implicit annotation based on user operations during system usage using business logic rules; second, entity relationships are extracted using existing knowledge graphs with the aid of remote supervision; and finally, pre-annotation is performed using existing intelligent pre-annotation services to jointly generate the labeled dataset.

[0083] Step (3.2) involves training the model on the existing labeled data using the existing model training module tools. The binary files generated by the training are used to create and publish intelligent services via container cloud. The published services are added to the pre-labeled service list for selection. The difficult example data called by the service enters the active learning module. New labeled data is generated iteratively through manual annotation by users, and the model is continuously generated iteratively.

[0084] Step (4), service encapsulation, invocation, and iterative evolution, such as... Figure 4 As shown, service encapsulation is mainly based on the model trained in step (3) to encapsulate services, support multi-copy deployment of services and high-load calls, and thus quickly export the service into the service installation package.

[0085] The service installation package includes the Docker image of the service, the service startup script, the service API description, and other information. Appropriate services are selected and invoked based on different knowledge service needs. During the invocation process, accuracy, efficiency, response time, and other metrics are collected and monitored in real time. If the curvature trend or absolute value of the declining curve reaches a threshold, model retraining is triggered, and the service is published.

Claims

1. A method for constructing a knowledge extraction platform based on active learning, characterized in that: The knowledge extraction platform includes a tag system construction layer, an intelligent service driving layer, a service capability evaluation layer, and a knowledge capability evolution layer. The label system construction layer constructs a label system to be labeled by uniformly accessing and preprocessing the original dataset, selecting the corresponding representation modes of entity knowledge, event knowledge, and rule knowledge based on the characteristics of the original data. The intelligent service driving layer uses the pre-built model in the intelligent pre-labeling to pre-label the original dataset according to the label system established by the label system construction layer, train the model, and encapsulate the intelligent services. The service capability assessment layer extracts knowledge and performs real-time calculation of metrics by calling the services encapsulated in the knowledge building intelligent service driving layer through the knowledge service invocation engine. The knowledge capability evolution layer provides unified management of knowledge construction services. The method for constructing the knowledge extraction platform includes the following steps: Step (1) involves importing the original data or labeled data into the knowledge extraction platform by uploading or configuring a database connection, and then converting and saving the data as a text file in UTF-8 format. Step (2): Select or create a new editing tag system based on the knowledge extraction task and the characteristics of the data source file; Step (3) involves machine annotation of people in the loop and model training and deployment, constructing an annotated dataset and deploying the service based on the original dataset and labels; Step (4) Service encapsulation, invocation and iterative evolution: based on the model trained in step (3), service encapsulation is performed, the service is exported as a service installation package, and services are selected and invoked according to knowledge service requirements.

2. The method for constructing a knowledge extraction platform based on active learning according to claim 1, characterized in that: In step (2), based on the knowledge extraction task, feature analysis of the text content is selected; the tag system with the highest matching degree is selected as the tag system to be labeled. If the matching degree does not reach the threshold, the tag system is manually created and edited.

3. The method for constructing a knowledge extraction platform based on active learning according to claim 1, characterized in that: Step (3) includes the following steps: Step (3.1) involves constructing the labeled dataset using three pre-annotation methods: first, implicit annotation is performed using logical rules; then, entity relationships are extracted using remote supervision based on existing knowledge graphs; and finally, pre-annotation is performed using existing intelligent pre-annotation services to jointly generate the labeled dataset. Step (3.2) involves training the model on the existing labeled data. The binary files generated by the training are used to create and publish intelligent services via container cloud. The published services are added to the pre-labeled service list. The difficult example data called by the services are added to the active learning module for manual labeling. New labeled data and models are generated iteratively.

4. The method for constructing a knowledge extraction platform based on active learning according to claim 1, characterized in that: The intelligent service driving layer actively learns to provide hints for pre-labeled examples with low accuracy and performs difficult example labeling. After the results of the difficult example labeling are fed back, the model is retrained and the intelligent service is encapsulated.

5. The method for constructing a knowledge extraction platform based on active learning according to claim 1, characterized in that: The model training is divided into two methods: model initialization training based on full data and incremental learning.

6. The method for constructing a knowledge extraction platform based on active learning according to claim 1, characterized in that: The intelligent service-driven layer includes an intelligent pre-labeling module, an active learning module, a hard example labeling module, a model training module, and a service encapsulation module; The intelligent pre-annotation module performs machine intelligent annotation on existing intelligent models before manual annotation. The active learning module filters the data to be labeled; After analyzing and identifying difficult cases that are judged incorrectly, the difficult case labeling module prioritizes manually labeling difficult cases with high error frequency. The model training module trains deep learning models on labeled datasets by scheduling hardware cluster resources through a container cloud platform and dynamically allocating resources to start containers for model training. The service encapsulation module encapsulates the trained model files into services for external systems to call.

7. The method for constructing a knowledge extraction platform based on active learning according to claim 6, characterized in that: The service encapsulation module first creates a Django project, which then creates a REST service using Python code. Next, it adds the inference function from the model inference script to the project startup code. Finally, it performs inference calculations and returns the results by calling the model inference function.

8. The method for constructing a knowledge extraction platform based on active learning according to claim 1, characterized in that: The service capability assessment layer includes a knowledge service invocation engine, an assessment indicator system, and an indicator real-time calculation module. The knowledge service calling engine automatically matches the required knowledge representation pattern and selects intelligent extraction services under different knowledge types and tag systems. The evaluation indicator system constructs different domain indicator libraries according to business types, and displays different types of indicators in sorted order. The real-time indicator calculation module performs real-time calculations of indicators based on different indicator systems and the frequency of data collection, and displays the indicator scores according to the keyness ranking of the evaluation indicators.

9. The method for constructing a knowledge extraction platform based on active learning according to claim 1, characterized in that: The knowledge capability evolution layer includes a knowledge service operation support environment, an evolution triggering engine, and a knowledge service integration framework. The knowledge service runtime support environment provides support for the knowledge extraction service to run in heterogeneous systems. The evolution trigger engine re-labels the training dataset based on changes in real-time calculated index parameter values, and then incrementally trains and publishes it as a service for use. The knowledge service integration framework matches and selects the installation package to be deployed and provides the deployment script.

Images