A data set management platform

Abstract

The application relates to the technical field of data management, and particularly relates to a data set management platform; the data set management platform comprises an input interface, a database, a scene analysis module, a data labeling module, a data analysis module and an output interface; the input interface receives user requirements and new data; the database stores all historical user requirements and data sets output according to the requirements; the scene analysis module maps the new data to the most suitable historical scene, and preferentially attempts to reuse a target data set and data elements thereof; if reuse is successful, the target data set elements are directly borrowed to enter the labeling module and are pre-labeled by using related model weights; if reuse fails, the new data is sent to an element modification module; the element modification module creates new data elements based on domain knowledge to feed back to the data labeling module for labeling; when receiving a new requirement, the data analysis module responds to the new requirement of the user in real time, and optimizes and updates data elements and labeling tools by driving natural language understanding.

Description

Technical Field

[0001] This invention relates to the field of data management technology, specifically a dataset management platform. Background Technology

[0002] With the development of artificial intelligence technology, high-quality, large-scale, and well-structured labeled datasets have become crucial for model training. Common data annotation content includes rectangles, polygons, key points, and semantic segmentation in images, as well as entity annotation, relation annotation, and sentiment annotation in text data.

[0003] With the widespread application of deep learning and large models, model performance is highly dependent on a large amount of high-quality labeled data. Traditional data labeling and dataset management usually treats labeling tools, storage systems, and version management as isolated modules and processes them separately. This lacks the ability to link domain knowledge, automated parsing, and interactive requirements, resulting in poor dataset reusability, high iterative maintenance costs, and insufficient matching with business needs. Summary of the Invention

[0004] The purpose of this invention is to provide a dataset management platform to solve the problems mentioned in the background section.

[0005] The objective of this invention can be achieved through the following technical solution: a dataset management platform, comprising: an input interface, a database, a scene parsing module, a data annotation module, a feature modification module, a data parsing module, and an output interface; The input interface is used to receive new user requests and new data, and then send them to the data parsing module and the scenario parsing module respectively. The database is used to store the needs of all historical users and their corresponding datasets, data element templates for all scenarios, and domain knowledge tables; The scene analysis module performs similar scene retrieval and matching based on the received newly added data to filter out the target dataset corresponding to the newly added data, and extracts the data elements of the target dataset as target data elements and sends them to the data annotation module; if the target dataset does not exist, the newly added data is sent to the element modification module. The feature modification module defines the representation method of the newly added data based on domain knowledge, and determines the target data features accordingly, and sends them to the data annotation module. The data parsing module decomposes new user requirements and identifies changed data elements based on domain knowledge. Then, it adjusts the annotation tools based on the changed data elements to output the target data elements and sends them to the data annotation module. The data annotation module annotates the new data based on the target data features to output a dataset and sends it to the output interface; at the same time, it updates the database according to the output location of the target data features; the output interface is used to convey the dataset of new data or new requirements to the user.

[0006] Preferably, the process of performing similar scene retrieval and matching based on the received newly added data is as follows: 2-1. Extract the newly added data and perform scene matching analysis on it to filter out candidate scenes that match the newly added data from all scenes in the database; 2-2, Filter out the target scenarios from the candidate scenarios that precisely match the newly added data; 2-3. Extract all datasets in the target scene and perform similarity analysis between them and the currently added data to output the overall similarity. When the overall similarity is greater than or equal to the preset similarity, this comparison dataset is used as the relevant dataset of the currently added data. Count the number of relevant datasets. If the number of relevant datasets is ≥1, select the relevant dataset with the highest overall similarity from the relevant datasets as the target dataset of the currently added data, and extract the data elements of the target dataset as the target data elements of the currently added data. If the number of relevant datasets is 0, send the currently added data to the element modification module.

[0007] Preferably, candidate scenarios that match the newly added data are selected: The newly added data is divided into text data and image data according to its content. Pre-trained visual models and natural language processing models are used to identify and extract the text data and image data respectively to obtain the keywords involved in the text data and image data. These keywords are then concatenated to obtain the keywords representing the current newly added data scenario. The keywords describing the scenario representing the newly added data are compared with the keywords of all scenarios in the database. If the keywords of a certain scenario in the database contain all the keywords representing the scenario of the newly added data, then this scenario is directly used as a candidate scenario for the newly added data. If the keywords of all scenarios in the database do not contain all the keywords of the scenario of the newly added data, then the scenario that overlaps with the keywords representing the scenario of the newly added data is selected as a candidate scenario.

[0008] Preferably, the target scenario that precisely matches the newly added data is selected from the candidate scenarios: The keywords of the candidate scenarios are compared with the keywords of the newly added data to obtain overlapping keywords. The number of overlapping keywords and the number of keywords in the newly added data are counted, and the number of overlapping keywords is divided by the number of keywords in the newly added data to obtain the overlap ratio. The scenario with the largest overlap ratio is selected as the primary scenario. Then the number of keywords in the primary scenario is calculated, and the number of keywords in the primary scenario is divided by the number of overlapping keywords to obtain the blank ratio. The scenario with the smallest blank ratio is selected as the target scenario of the newly added data.

[0009] Preferably, the similarity analysis outputs an overall similarity score: 5-1. Extract the dataset corresponding to the scene that matches the target scene from the database and use it as the comparison dataset for the newly added data; extract the original data from the comparison dataset and divide the original data into text data and image data according to the data types it includes; similarly, divide the newly added data into text data and image data. 5-2. Calculate the text similarity between the text data of the newly added data and the text data of the comparison dataset. Specifically, the text is converted into numerical vectors using a pre-trained natural language processing model. This yields the numerical vectors of the text data of the newly added data and the text data of the comparison dataset. The cosine similarity is then used to calculate the similarity between the numerical vectors of the two text data to obtain the text similarity. 5-3. Calculate the image similarity between the image data of the newly added data and the image data of the comparison dataset. Specifically, use a pre-trained visual model to extract the feature vector of the image. This will give you the feature vectors of the image data of the newly added data and the image data of the comparison dataset. Use cosine similarity to measure the similarity between the feature vectors of the two image data to obtain the image similarity. 5-4. Assign a weight to the text and image similarity, and the sum of the weights of the text similarity and image similarity is 1. Linearly weight the text and image similarities according to the weights assigned to the text and image similarities to obtain the overall similarity. Thus, the overall similarity between the newly added data and all comparison datasets can be obtained.

[0010] Preferably, the representation method of the newly added data is defined based on domain knowledge, and the target data elements are determined accordingly: Extract the newly added data and its corresponding scenario, and extract the domain knowledge table and data element templates in the same scenario as the newly added data in the database. Call the historical task objectives and the definition of each task objective of the data element template and domain knowledge for the scenario in the domain knowledge table to determine the types of objects that need to be identified, labeled or calculated in the newly added data. Further analyze the representation of each type of object in the data, and then associate recommended labeling tool types with each representation method to generate all representation methods of the newly added data and the recommended labeling tools associated with each representation method. Each representation method is refined into several basic annotation items, i.e., data elements. From this, we can obtain all representation methods of the newly added data, as well as the data elements and annotation tools corresponding to each representation method. These are used as the target data elements of the newly added data and fed back to the data annotation module.

[0011] Preferably, the new user requirements are broken down and the changed data elements are clearly defined based on domain knowledge: The system extracts new user requirements, performs text cleaning to remove irrelevant characters, and segments the text to extract core verbs and noun phrases to output a standardized word list. An integrated large language model is used to perform deep semantic analysis on the word list to identify the user's change type, target, and expected change effect. The change type, target, and expected change effect are merged into a new requirement intent, which is then matched with a domain knowledge base to obtain the change data elements of the new requirement.

[0012] Preferably, the annotation tool is then adjusted based on the changed data features to output the target data features: Step 1: Extract all annotation tools stored in the database and match them with the changed data features to identify the supported and unsupported features; Step two: Check whether the annotation tool corresponding to the supporting element meets the expected change effect of the supporting element. If it does, record the supporting element as an effect-satisfied element; otherwise, record the supporting element as an effect-unsatisfied element. Count the number of unsupported elements, the number of effect-satisfied elements, and the number of effect-unsatisfied elements. Step 3: If the number of elements is not supported to be 0, and the effect does not meet the requirement of the number of elements being 0, then the changed data element is directly sent to the data annotation module as the target data element of the new requirement. Step 4: If the number of elements is not supported = 0, and the effect does not meet the requirement of the number of elements ≥ 1, then generate a work order for adjusting the accuracy of the annotation tool and send it to the corresponding human to adjust the accuracy so that the annotation tool can meet the requirements of the new requirements. When the instruction that the accuracy adjustment of the annotation tool is completed is received, return to step 1 until the requirements of step 3 are met, or the maximum number of loops is reached, then request human intervention. Step 5: If the conditions other than those in Steps 3 and 4 are met, load the annotation tool corresponding to the unsupported feature. When the loading completion instruction is received, return to Step 1 until the requirements of Step 3 are met, or the maximum number of loops is reached, then request manual intervention.

[0013] Preferably, the database is updated according to the output location of the target data elements: In the data annotation module, annotation tasks are created based on data elements. This includes setting annotation tools, importing labels, assigning tasks, and executing the annotation process to output initial annotation results. Human annotators then adjust the initial annotation results to finally obtain the dataset.

[0014] Preferably, the database is updated according to the output location of the target data elements: Identify the output location of the target data element. If the output location is the scene parsing module, there is no need to update the data element template and domain knowledge table of the scene to which the target data element belongs. Instead, compare the target data element in the dataset generated by the data annotation module with all data elements in the data element template of its scene, update the target data element in the data element template, and supplement the domain knowledge table with the representation method in the dataset generated by the newly added data. If the output location is the data parsing module, update the target data element in the data element template, and update the annotation tool of the dataset in the domain knowledge table.

[0015] The beneficial effects of this invention are: 1. Through multi-level intelligent matching algorithms, it can accurately locate the historical scenarios most relevant to the new data, automatically recommend the optimal data element solution, enable data assets to be reused efficiently, and effectively improve the efficiency and accuracy of data reuse. 2. By automatically translating target data features into executable annotation tasks and combining model pre-annotation with manual correction workflows, annotation efficiency and consistency are significantly improved. It can automatically configure appropriate annotation tools and prioritize the use of model weights closest to the target dataset for batch pre-annotation to reduce redundancy. When connecting to the feature modification module and data parsing module, it can also write back newly discovered data features or adjusted annotation tools to the data feature template and domain knowledge table in the scenario, forming a closed-loop learning and reuse, thereby reducing annotation costs, shortening delivery cycles, and improving data quality. 3. When there are no reusable scenarios, the element modification module decomposes the business objectives of the newly added data into observable and labelable representations and data elements. By structuring domain expert knowledge into machine-executable labeling specifications, i.e. data elements, it ensures a high degree of consistency between the dataset and the business objectives of the newly added data, provides a standardized solution for handling novel and complex scenarios, and lowers the threshold for using domain knowledge. 4. By using natural language processing technology to understand new user needs and drive the platform to adaptively adjust the annotation tools, the platform has the ability to continuously learn and respond quickly to business changes. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the system module connections of the present invention; Figure 2 This is a schematic diagram illustrating the process of modifying and updating the database using the MCP of this invention. Detailed Implementation

[0017] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0018] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention. Example 1

[0019] Please see Figure 1 As shown, the present invention provides a dataset management platform, which includes: an input interface, a database, a scene parsing module, a data annotation module, a feature modification module, a data parsing module, and an output interface; The input interface receives user requests and new data, and sends them to the database for storage. The database stores all new requirements from historical users, as well as datasets output based on those requirements. These datasets include raw data, labeled data, and data elements. It's important to note that the datasets are labeled according to user requirements and have undergone quality control. The database also stores data element templates and domain knowledge tables for all scenarios. Each scenario is described using several keywords. Data element templates are created by summarizing the data elements from all datasets within the same scenario and then removing duplicate elements. The scene parsing module performs similar scene retrieval and matching based on the received new data volume to determine whether a similar dataset exists. If it does, it outputs the target data elements based on the similar dataset and sends them to the data annotation module; if it does not, it sends the currently added data to the element modification module. The process of performing similar scene retrieval and matching based on newly received data to determine whether a similar dataset exists is as follows: The newly added data is extracted and divided into text and image data according to their content. Pre-trained visual and natural language processing models are used to identify and extract keywords from the text and image data respectively. These keywords are then concatenated to obtain keywords representing the scenario of the newly added data. These scenario keywords are compared with keywords from all scenarios in the database. If a scenario in the database contains all the keywords representing the scenario of the newly added data, it indicates an exact match, and this scenario is selected as a candidate scenario for the newly added data. If no scenario in the database contains all the keywords representing the scenario of the newly added data, then scenarios that overlap with the keywords representing the scenario of the newly added data are selected as candidate scenarios. Fuzzy matching is used here. To avoid overlooking other scenarios related to the new data, the keywords of candidate scenarios are compared with the keywords of the current new data to obtain overlapping keywords. The number of overlapping keywords and the number of keywords in the current new data are counted, and the overlap ratio is obtained by dividing the number of overlapping keywords by the number of keywords in the current new data. The scenario with the largest overlap ratio is selected as the primary scenario. Then, the number of keywords in the primary scenario is calculated, and the blank ratio is obtained by dividing the number of keywords in the primary scenario by the number of overlapping keywords. The scenario with the smallest blank ratio is selected as the target scenario for the current new data. By calculating the overlap ratio first and then the blank ratio, the keywords of the new data are first ensured to be covered to the greatest extent, and then the matching accuracy of the scenarios is optimized. The overlap ratio and blank ratio ensure that the selected target scenarios are highly consistent with the new data in terms of content and scope, thus providing a reliable foundation for subsequent data element extraction and annotation generation. Extract the dataset corresponding to the scene consistent with the target scene from the database and use it as the comparison dataset for the newly added data. Extract the original data from the comparison dataset and divide it into text data and image data according to the data types it includes. Similarly, divide the newly added data into text data and image data. It should be noted that the current model training results usually support text data and image data. First, calculate the text similarity between the text data of the newly added data and the text data of the comparison dataset. Specifically, use a pre-trained natural language processing model (such as BERT, Word2Vec, or TF-IDF) to convert the text into numerical vectors (embedding vectors). For short text, directly use the sentence vectors generated by the model as numerical vectors. For long text (such as paragraphs or documents), first segment it into sentences or use a document-level model (such as Doc). 2Vec) generates an overall vector as a numerical vector; from this, we can obtain the numerical vectors of the text data of the newly added data and the text data of the comparison dataset. For the newly added text data, we also extract its numerical vector, and use cosine similarity to calculate the similarity between the numerical vectors of the two text data to obtain the text similarity. The text similarity formula is: Text Similarity = (A·B) / (||A||×||B||), where A and B are the numerical vectors of the text data of the newly added data and the text data of the comparison dataset, respectively. If the comparison dataset contains multiple text data samples, we can first calculate the similarity between the text data of the newly added data and each text data sample in the comparison dataset, and then take the average value as the final text similarity. It should be noted that the output text similarity ranges from 0 to 1. When the text similarity is equal to 1, it means that the two text data are completely similar. Then, the image similarity between the newly added image data and the image data in the comparison dataset is calculated. Specifically, a pre-trained visual model (such as ResNet, VGG, or CLIP) is used to extract the feature vectors of the images. Specifically, through the forward propagation of the model, the output of the last convolutional layer or fully connected layer is obtained as the image feature vector. Thus, the feature vectors of the newly added image data and the image data in the comparison dataset are obtained. Cosine similarity is used to measure the similarity between the feature vectors of the two image data to obtain the image similarity. The image similarity calculation formula is: Image Similarity = (C·D) / (||C||×||D||), where C and D are the feature vectors of the newly added image data and the image data in the comparison dataset, respectively. If the comparison dataset contains multiple image data samples, the similarity between the newly added image data and each image data sample in the comparison dataset can be calculated first, and then the average value can be taken as the final image similarity. It should be noted that the output image similarity range is within 0-1. When the image similarity is equal to 1, it means that the two image data are completely similar. Finally, text and image similarities are combined into a single overall similarity score to comprehensively reflect the similarity between the current data and the comparison dataset. Specifically, a weight is assigned to text and image similarities based on the business scenario and data importance, with the sum of the weights for text and image similarities being 1. For example, in scenarios where text is dominant, text has a higher weight (e.g., 0.7) and images a lower weight (e.g., 0.3); in scenarios where images are dominant (e.g., image classification), images have a higher weight (e.g., 0.7) and text a lower weight (e.g., 0.3); if text and images are equally important (e.g., multimodal data), each has a weight of 0.5. The text and image similarities are then linearly weighted according to their assigned weights to obtain the overall similarity score. This yields the overall similarity between the newly added data and all comparison datasets. A preset similarity threshold is set, designed based on the confidence interval theory in statistics and the precision-recall balance principle in machine learning. This threshold maximizes the similarity between the current data and all comparison datasets while ensuring dataset quality. The efficiency of historical data reuse; those skilled in the art set the similarity threshold to 0.75 because they have found through data statistics that 0.75 achieves the best balance between precision and recall, effectively filtering out truly similar datasets while avoiding missing suitable candidate sets due to overly strict standards; when the overall similarity is greater than or equal to the preset similarity, it indicates that the currently added data and this comparison dataset are highly similar, and this comparison dataset is then used as the relevant dataset for the currently added data; the number of relevant datasets is counted, and if the number of relevant datasets is ≥1, the relevant dataset with the highest overall similarity is selected as the target dataset for the currently added data, and the data elements of the target dataset are extracted as the target data elements for the currently added data; if the number of relevant datasets is 0, it indicates that there is no highly similar dataset in the database with the same scenario as the currently added data, and the currently added data is sent to the element modification module; Through multi-level intelligent matching algorithms, the most relevant historical scenarios to new data can be accurately located, and the optimal data element solutions can be automatically recommended, enabling data assets to be reused efficiently and effectively improving the efficiency and accuracy of data reuse.

[0020] The data annotation module annotates newly added data based on target data elements to output a dataset, which is then sent to the interface output for delivery to the user. It identifies the output location of the target data elements. If the output location is the scene parsing module, there is no need to update the data element template and domain knowledge table for the scene to which the target data elements belong, avoiding redundant operations. If the output location is the element modification module, the target data elements in the dataset generated by the data annotation module are compared with all data elements in the data element template of its scene, and the target data elements are updated to the data element template (i.e., data elements that only exist in the target data elements are added to the data element template). Simultaneously, the representation methods in the dataset generated from the newly added data are added to the domain knowledge table. If the output location is the data parsing module, the target data elements are updated to the data element template (i.e., data elements that only exist in the target data elements are added to the data element template). It should be noted that changes in user requirements, i.e., new requirements, may lead to changes in data elements. Simultaneously, the annotation tools for the dataset are updated to the domain knowledge table. These annotation tools include the annotation tools with adjusted precision and the newly loaded annotation tools. The specific process of annotating newly added data based on target data elements to output a dataset is as follows: The data annotation module creates annotation tasks based on data features. This includes setting annotation tools, importing labels, assigning tasks, and executing the annotation process to output initial annotation results. Human annotators then adjust these initial results to obtain the final dataset. Specifically, the annotation tool settings are configured according to the annotation type (e.g., bounding boxes, polygons). Label import involves loading the target data features. Task assignment automates the annotation process by assigning tasks sequentially based on the annotators' skills. The annotation process uses a pre-trained model (e.g., YOLO for object detection) to pre-annotate the new data in batches, generating initial annotation results (e.g., automatically drawing bounding boxes). It should be noted that the pre-annotated model prioritizes model weights related to the closest dataset (i.e., the target dataset) to improve accuracy. By automatically translating target data features into executable annotation tasks and combining model pre-annotation with manual correction workflows, annotation efficiency and consistency are significantly improved. It can automatically configure appropriate annotation tools and prioritize the use of model weights closest to the target dataset for batch pre-annotation to reduce redundancy. When connecting to the feature modification module and data parsing module, it can also write back newly discovered data features or adjusted annotation tools to the data feature template and domain knowledge table in the scenario, forming a closed-loop learning and reuse, thereby reducing annotation costs, shortening delivery cycles, and improving data quality.

[0021] The feature modification module, based on domain knowledge, clarifies the representation method of the newly added data, determines the target data features of the newly added data according to the representation method, and sends them to the data annotation module. The specific process for determining the target data features of the newly added data is as follows: Extract the newly added data and its corresponding scenario, and extract the domain knowledge table and data element templates (including historical task objectives and definitions of each task objective, industry ontology knowledge, expert experience rules, representation methods, and related recommendation annotation tools, etc.) from the database for the same scenario as the newly added data. Call the data element templates and historical task objectives and definitions of each task objective (such as detection, classification, segmentation, key point recognition, etc.) in the domain knowledge table for this scenario to determine the types of objects in the newly added data that need to be identified, labeled, or calculated. Further analyze the representation of each object in the data. Specifically: if the object is a visible boundary area, such as a crack or repair area, then determine... Polygon segmentation is used for representation; if the object is an independent entity, such as a component or vehicle, bounding box representation is used; if the object is a key point that needs to be described in terms of spatial location or morphological features, such as a weld point or joint, key point representation is used; if the object involves state, property, or hierarchical information, such as severity, type, pass / fail, classification labeling is used; if the object features need to be quantified, such as area, length, grayscale mean, or brightness change rate, numerical representation is used and a calculation formula is set in the feature definition; then, recommended labeling tool types (such as box selection, point selection, mask, polygon drawing) are associated with each representation method to generate all representation methods for the currently added data and recommended labeling tools associated with each representation method; Each representation method is broken down into several basic annotation items, i.e., data elements. For example, the representation method of disease area is broken down into elements such as disease boundary coordinates, disease type, and disease area; the representation method of metal weld joint is broken down into elements such as weld joint coordinates, weld joint diameter, and weld joint defect level. Thus, all representation methods of the newly added data and the data elements and annotation tools corresponding to each representation method can be obtained. These are used as the target data elements of the newly added data and fed back to the data annotation module. When there are no reusable scenarios, the feature modification module breaks down the business objectives of the newly added data into observable and labelable representations and data elements. By structuring domain expert knowledge into machine-executable annotation specifications, i.e. data elements, it ensures a high degree of consistency between the dataset and the business objectives of the newly added data, provides a standardized solution for handling novel and complex scenarios, and lowers the threshold for using domain knowledge.

[0022] If a user raises a new requirement, the data parsing module breaks down the new requirement and identifies the changed data elements based on domain knowledge. It then determines whether the platform's annotation tools meet the annotation requirements for the changed data elements. If they do, the module directly outputs the changed data elements as target data elements and sends them to the data annotation module. If not, it modifies or loads the annotation tools until the user's new requirement is met. The process of breaking down the new requirement and identifying the changed data elements based on domain knowledge is as follows: Extract the user's new requirements (e.g., the area bounding box for block repair is too large, I need more precision), and perform text cleaning on the new requirements to remove irrelevant characters and perform word segmentation to extract core verbs and noun phrases to output a standardized word list. Use an integrated large language model to perform deep semantic analysis on the word list to identify the user's change type (e.g., area bounding box), operation object (e.g., block repair), and expected change effect (e.g., more precision). Merge the change type, operation object, and expected change effect into a new requirement intent, and then match the parsed new requirement intent with the domain knowledge base to obtain the change data elements of the new requirement. Specifically, query the standard definition related to the operation object in the domain knowledge table (e.g., the standard definition of block repair in the field of road defects) and retrieve existing data element specifications (e.g., the current bounding box annotation method) to match and obtain the change data elements of the new requirement. This includes determining whether the annotation tools on this platform meet the annotation requirements for changed data elements. The process for determining whether a changed data element meets the requirements for newly added data is as follows: Step 1: Extract all annotation tools stored in the database and match them with the changed data features to identify supported and unsupported features. For example, if a new feature requires polygon splitting annotation, check if there is an annotation tool that supports polygon splitting annotation. If it exists, the feature is a supported feature; otherwise, it is an unsupported feature. Step two: Check whether the annotation tool corresponding to the supporting element meets the expected change effect of the supporting element. If it does, record the supporting element as an effect-satisfied element; otherwise, record the supporting element as an effect-unsatisfied element. Count the number of unsupported elements, the number of effect-satisfied elements, and the number of effect-unsatisfied elements. Step 3: If the number of elements is not supported = 0, and the effect does not meet the requirement of the number of elements = 0, it means that the annotation tool of this platform meets the new requirements. Then, the changed data elements are directly sent to the data annotation module as the target data elements of the new requirements. Step 4: If the number of unsupported elements is 0, and the effect does not meet the requirement of ≥1 element count, it means that the expected effect of the annotation tool on this platform does not meet the new requirements. In this case, an annotation tool accuracy adjustment work order is generated and sent to the corresponding human to adjust the accuracy so that the annotation tool can meet the requirements of the new requirements. When the instruction that the annotation tool accuracy adjustment is completed is received, return to step 1 until the requirements of step 3 are met, that is, the number of unsupported elements is 0, and the effect does not meet the requirement of 0 element count; or the maximum number of loops is reached (those skilled in the art take the maximum number of loops as 5 times), then request human intervention. Step 5: If any of the conditions other than those in Step 3 and Step 4 are met, it means that the annotation tools of this platform cannot meet the new requirements. Then, the annotation tools corresponding to the unsupported features will be loaded. When the loading completion instruction is received, the process will return to Step 1 until the requirements of Step 3 are met, that is, the number of unsupported features is 0 and the effect does not meet the requirement of the number of features being 0; or the maximum number of loops is reached (those skilled in the art will take the maximum number of loops as 5 times), then manual intervention will be requested. By using natural language processing technology to understand new user needs and drive the platform to adaptively adjust annotation tools, the platform has the ability to continuously learn and respond quickly to business changes.

[0023] With the widespread application of deep learning and large-scale models, model performance highly depends on a large amount of high-quality labeled data. Traditional data labeling and dataset management typically treats labeling tools, storage systems, and version management as isolated modules, lacking the ability to integrate domain knowledge, automated parsing, and interactive requirements. This results in poor dataset reusability, high iterative maintenance costs, and insufficient alignment with business needs. To address this technical problem, this paper proposes a dataset management platform. Its operation begins with the input of new user requirements and new data. Through intelligent scene parsing, it matches historical experience or creates new specifications, generating high-quality datasets through labeling, ultimately forming a continuously optimized knowledge loop. Specifically, when new data enters the platform, the scene parsing module... Keyword extraction and multi-dimensional similarity analysis are used to find the best-matching historical dataset. If a match is found, the data elements are directly reused. If no match is found, the element modification module creates new data elements based on domain knowledge, and the data annotation module executes the data annotation process based on the determined data elements to output the corresponding dataset. At the same time, the data parsing module responds to new user needs in real time, driving the optimization and updating of data elements and annotation tools through natural language understanding. This achieves the organic integration of domain knowledge, automated tools, and user feedback, and builds a data ecosystem that can continuously evolve with business needs through three mechanisms: intelligent matching, dynamic optimization, and continuous learning. The platform not only solves the pain points in current data annotation and management, but more importantly, it forms a virtuous cycle of knowledge accumulation and reuse. Example 2

[0024] like Figure 2 As shown in Example 1, MCP (Model Context Protocol) can also be used to call other functional modules to modify and update the data element templates and domain knowledge tables in the database. The specific process is as follows: The new user requirements (i.e., user-generated verbal instructions) received from the input interface are parsed by the data parsing module to obtain the new requirement intent and the changed data elements. Then, the database is modified and updated by calling other functional modules through MCP (Model Context Protocol). Among them, LLM parsing is a large-scale language model intelligent agent, namely the element modification and data parsing modules described in Example 1. This is because the dataset generated by data annotation using the target data elements output by these two modules needs to be updated with the data element templates and domain knowledge tables in the database.

[0025] The above description is merely an example and illustration of the structure of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the structure of the invention or exceed the scope defined in the claims, all of which should fall within the protection scope of the present invention.

Claims

1. A dataset management platform, characterized in that, This platform includes: input interface, database, scene parsing module, data annotation module, feature modification module, data parsing module, and output interface; The input interface is used to receive new user requests and new data, and then send them to the data parsing module and the scenario parsing module respectively. The database is used to store the needs of all historical users and their corresponding datasets, data element templates for all scenarios, and domain knowledge tables; The scene parsing module performs similar scene retrieval and matching based on the received newly added data to filter out the target dataset corresponding to the newly added data, and extracts the data elements of the target dataset as target data elements and sends them to the data annotation module; if the target dataset does not exist, the newly added data is sent to the element modification module. The feature modification module defines the representation method of the newly added data based on domain knowledge, and determines the target data features accordingly, and sends them to the data annotation module. The data parsing module is used to break down new user requirements and identify the changed data elements based on domain knowledge. Then, based on the changed data elements, the annotation tool is adjusted to output the target data elements and send them to the data annotation module. The data annotation module annotates the new data based on the target data features to output a dataset and sends it to the output interface; at the same time, it updates the database according to the output location of the target data features; the output interface is used to convey the dataset of new data or new requirements to the user.

2. The dataset management platform according to claim 1, characterized in that, The process of performing similar scene retrieval and matching based on the newly received data is as follows: 2-1. Extract the newly added data and perform scene matching analysis on it to filter out candidate scenes that match the newly added data from all scenes in the database; 2-2, Filter out the target scenarios from the candidate scenarios that precisely match the newly added data; 2-3. Extract all datasets in the target scene and perform similarity analysis between them and the newly added data to output the overall similarity. When the overall similarity is greater than or equal to the preset similarity, this comparison dataset is used as the relevant dataset for the newly added data. The number of related datasets is counted. If the number of related datasets is ≥1, the related dataset with the highest overall similarity is selected as the target dataset for the new data. The data elements of the target dataset are then extracted as the target data elements for the new data. If the number of related datasets is 0, the newly added data will be sent to the feature modification module.

3. The dataset management platform according to claim 2, characterized in that, Filter out candidate scenarios that match the newly added data: The newly added data is divided into text data and image data according to its content. Pre-trained visual models and natural language processing models are used to identify and extract the text data and image data respectively to obtain the keywords involved in the text data and image data. These keywords are then concatenated to obtain the keywords representing the current newly added data scenario. The keywords describing the scenario representing the newly added data are compared with the keywords of all scenarios in the database. If the keywords of a certain scenario in the database contain all the keywords representing the scenario of the newly added data, then this scenario is directly used as a candidate scenario for the newly added data. If the keywords of all scenarios in the database do not contain all the keywords of the scenario of the newly added data, then the scenario that overlaps with the keywords representing the scenario of the newly added data is selected as a candidate scenario.

4. A dataset management platform according to claim 3, characterized in that, Filter out the target scenarios from the candidate scenarios that precisely match the newly added data: The keywords of the candidate scenarios are compared with the keywords of the newly added data to obtain overlapping keywords. The number of overlapping keywords and the number of keywords in the newly added data are counted, and the number of overlapping keywords is divided by the number of keywords in the newly added data to obtain the overlap ratio. The scenario with the largest overlap ratio is selected as the primary scenario. Then, calculate the number of keywords in the primary scenario, divide the number of keywords in the primary scenario by the number of overlapping keywords to get the blank percentage, and select the scenario with the smallest blank percentage as the target scenario for the current new data.

5. A dataset management platform according to claim 4, characterized in that, Similarity analysis is performed to output the overall similarity: 5-1. Extract the dataset corresponding to the scene that matches the target scene from the database and use it as the comparison dataset for the newly added data; extract the original data from the comparison dataset and divide the original data into text data and image data according to the data types it includes; similarly, divide the newly added data into text data and image data. 5-2. Calculate the text similarity between the text data of the newly added data and the text data of the comparison dataset. Specifically, the text is converted into numerical vectors using a pre-trained natural language processing model. This yields the numerical vectors of the text data of the newly added data and the text data of the comparison dataset. The cosine similarity is then used to calculate the similarity between the numerical vectors of the two text data to obtain the text similarity. 5-3. Calculate the image similarity between the image data of the newly added data and the image data of the comparison dataset. Specifically, use a pre-trained visual model to extract the feature vector of the image. This will give you the feature vectors of the image data of the newly added data and the image data of the comparison dataset. Use cosine similarity to measure the similarity between the feature vectors of the two image data to obtain the image similarity. 5-4. Assign a weight to the text and image similarity, and the sum of the weights of the text similarity and image similarity is 1. Linearly weight the text and image similarities according to the weights assigned to the text and image similarities to obtain the overall similarity. Thus, the overall similarity between the newly added data and all comparison datasets can be obtained.

6. A dataset management platform according to claim 5, characterized in that, Based on domain knowledge, we can clearly define how to represent the newly added data and determine the target data elements accordingly. Extract the newly added data and its corresponding scenario, and use the domain knowledge table and data element template in the database that are in the same scenario as the newly added data. Call the historical task objectives and the definition of each task objective of the data element template and domain knowledge for the scenario in the domain knowledge table to determine the types of objects that need to be identified, labeled or calculated in the newly added data. Further analyze the representation of each type of object in the data, and then associate recommended labeling tool types with each representation method to generate all representation methods of the newly added data and the recommended labeling tools associated with each representation method. Each representation method is refined into several basic annotation items, i.e., data elements. From this, we can obtain all representation methods of the newly added data, as well as the data elements and annotation tools corresponding to each representation method. These are used as the target data elements of the newly added data and fed back to the data annotation module.

7. A dataset management platform according to claim 6, characterized in that, Deconstruct new user requirements and identify the data elements to be changed based on domain knowledge: The system extracts new user requirements, performs text cleaning to remove irrelevant characters, and segments the text to extract core verbs and noun phrases to output a standardized word list. An integrated large language model is used to perform deep semantic analysis on the word list to identify the user's change type, target, and expected change effect. The change type, target, and expected change effect are merged into a new requirement intent, which is then matched with a domain knowledge base to obtain the change data elements of the new requirement.

8. A dataset management platform according to claim 7, characterized in that, Then, based on the changed data features, the annotation tools are adjusted to output the target data features: Step 1: Extract all annotation tools stored in the database and match them with the changed data features to identify the supported and unsupported features; Step two: Check whether the annotation tool corresponding to the supporting element meets the expected change effect of the supporting element. If it does, record the supporting element as an effect-satisfied element; otherwise, record the supporting element as an effect-unsatisfied element. Count the number of unsupported elements, the number of effect-satisfied elements, and the number of effect-unsatisfied elements. Step 3: If the number of elements is not supported to be 0, and the effect does not meet the requirement of the number of elements being 0, then the changed data element is directly sent to the data annotation module as the target data element of the new requirement. Step 4: If the number of elements is not supported = 0, and the effect does not meet the requirement of the number of elements ≥ 1, then generate a work order for adjusting the accuracy of the annotation tool and send it to the corresponding human to adjust the accuracy so that the annotation tool can meet the requirements of the new requirements. When the instruction that the accuracy adjustment of the annotation tool is completed is received, return to step 1 until the requirements of step 3 are met, or the maximum number of loops is reached, then request human intervention. Step 5: If the conditions other than those in Steps 3 and 4 are met, load the annotation tool corresponding to the unsupported feature. When the loading completion instruction is received, return to Step 1 until the requirements of Step 3 are met, or the maximum number of loops is reached, then request manual intervention.

9. A dataset management platform according to claim 8, characterized in that, Update the database based on the output location of the target data features: In the data annotation module, annotation tasks are created based on data elements. This includes setting annotation tools, importing labels, assigning tasks, and executing the annotation process to output initial annotation results. Human annotators then adjust the initial annotation results to finally obtain the dataset.

10. A dataset management platform according to claim 9, characterized in that, Update the database based on the output location of the target data features: Identify the output location of the target data element. If the output location is the scene parsing module, there is no need to update the data element template and domain knowledge table of the scene to which the target data element belongs. Instead, compare the target data element in the dataset generated by the data annotation module with all data elements in the data element template of its scene, update the target data element in the data element template, and supplement the domain knowledge table with the representation method in the dataset generated by the newly added data. If the output location is the data parsing module, update the target data element in the data element template, and update the annotation tool of the dataset in the domain knowledge table.

Images