Rapid labeling method based on artificial intelligence large model

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a rapid labeling method based on an artificial intelligence large model, which comprises the following steps of: 1, extracting an entity set and a relationship set of a specific field through a document set of a target field, identifying nouns, terms and keywords related to the field in a document by using an entity identification technology to form an entity set, and marking the entity set and the relationship set; the method comprises the following steps of: establishing an entity set, analyzing logic and semantic relevance among entities by utilizing a relation extraction technology, generating a relation set, combining the entity set with the relation set to construct a knowledge triad, and the knowledge triad comprises an association relation among the entities. By introducing the domain knowledge graph, the entity and relation information extracted from the target domain document is structured into the knowledge triad, and the knowledge triad is matched with the to-be-annotated data to optimize the annotated semantic information, so that the domain adaptability and semantic consistency of the annotation result are realized, and higher accuracy and consistency effects of the annotation data in the professional field are obtained.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a fast labeling method based on an artificial intelligence large model. Background Art

[0002] In the context of the rapid development of artificial intelligence and big data technology, data annotation, as a basic link in deep learning model training, directly affects the performance and promotion of the model. With the widespread application of deep learning models in the fields of natural language processing, computer vision and multimodal data processing, the scale and complexity of annotation tasks have increased rapidly, and higher requirements have been placed on the efficiency and quality of annotation results. In the current technical environment, traditional annotation methods are stretched to the limit and cannot meet the growing technical needs.

[0003] Traditional manual labeling methods mainly rely on manual operations, which have certain flexibility and accuracy, but have problems such as low efficiency, high time cost, and poor consistency. In small-scale data labeling, manual labeling has certain advantages. If the labeling task is extended to large-scale data and cross-domain and multimodal scenarios, the limitations of traditional methods become increasingly prominent, as shown in the following:

[0004] The knowledge of existing large models mainly comes from a wide range of general data and lacks deep semantic understanding for specific fields, resulting in insufficient accuracy and consistency of annotation results in professional fields;

[0005] Large models cannot flexibly adapt to the needs of different labeling tasks. Existing methods lack dynamic optimization methods for labeling task prompt templates, resulting in low efficiency and insufficient adaptability in multi-scene labeling tasks.

[0006] In the labeling task, the existing automatic labeling methods do not work well for complex and ambiguous samples, and require manual review, which increases the workload and time cost. At the same time, they fail to fully utilize the manual review results to optimize the model performance;

[0007] The labeling task has expanded from a single modality to multi-modality. Existing labeling technologies have limited capabilities for collaborative labeling of multi-modal data. At the same time, there is a lack of optimization methods for the hierarchical and refined processing of labeling results.

[0008] Therefore, those skilled in the art provide a fast labeling method based on an artificial intelligence large model to solve the above-mentioned problems. Summary of the invention

[0009] In view of the deficiencies in the prior art, the present invention provides a rapid labeling method based on an artificial intelligence large model to solve the problems raised in the above background technology.

[0010] To achieve the above objectives, the present invention is implemented through the following technical solutions: a fast labeling method based on artificial intelligence large model, comprising:

[0011] Step 1: Extract entity sets and relationship sets in a specific field through the document set of the target field, wherein entity recognition technology is used to identify domain-related nouns, terms and keywords in the document to form an entity set, and relationship extraction technology is used to analyze the logical and semantic associations between entities to generate a relationship set. The entity set and the relationship set are combined to construct a knowledge triple. The knowledge triple contains the association relationship between entities. The constructed knowledge triple is stored in the graph database to form a domain knowledge graph;

[0012] Step 2: Based on the knowledge graph, perform knowledge enhancement annotation on the data to be annotated, match the data to be annotated with the domain knowledge graph, and perform association matching with the corresponding entities in the knowledge graph according to the entities appearing in the data to be annotated. Optimize the annotated semantic information by querying the contextual association relationship of the entities in the knowledge graph. In the matching process, the logic of constructing entities and relationships in the knowledge triples is used to provide domain information background for step 3.

[0013] Step 3: Based on the results of knowledge-enhanced annotation, a dynamic prompt template is designed for optimization. Combined with the historical data of the annotation task, the content and structure of the prompt template are dynamically adjusted by analyzing the historical annotation results. In the dynamic prompt template, the domain information provided by the knowledge-enhanced annotation is combined to generate prompt content for a specific annotation task, so that the artificial intelligence large model can use it in the annotation task.

[0014] Step 4: Use dynamic prompt templates and large models to preliminarily annotate the data to be annotated, and screen samples in the unlabeled data through active learning. During the screening process, the uncertainty score of the sample is calculated in combination with the classification probability distribution in the annotation results, and the samples are sorted according to the uncertainty score. Some samples are selected for manual review based on the set uncertainty threshold. For the samples after manual review, the corrected data is sent back to the model training set to update the model parameters;

[0015] Step 5: For multimodal labeling tasks, use the optimized large model to label multimodal data, and extract text feature vectors and visual feature vectors from multimodal data of text and images, and achieve collaborative labeling of multimodal data through alignment and matching of feature vectors. At the same time, for the hierarchical structure of labels, build a label system with different granularities, optimize the labels according to the requirements of the labeling task, adjust the labeling probability by calculating the correlation between label hierarchies, generate labeling results and output them to the automated labeling workflow to complete the overall rapid labeling task process.

[0016] Preferably, the generation of the relationship set in the knowledge triple in step 1 includes constructing a relationship weight by counting the co-occurrence frequency between entities in the target domain document, and the relationship weight is calculated by the following formula:

[0017]

[0018] Among them, w ij For entity e i and e j The relationship weight between

[0019] Frequency (e i , e j ) is entity e i and e j The number of times they co-occur in the document,

[0020] Frequency (e i ) and frequency (e j ) is entity e i and e j The total number of times it occurs in the document.

[0021] Preferably, the relationship weight is used to adjust the priority of entity association in the knowledge graph, and the priority calculation formula is:

[0022] Among them, p ij For entity e i To entity e j The priority of entity e i The number of all directly related entities in the knowledge graph, w ij For entity e i and e j The relationship weight between

[0023] Represents entity e i The total weight of the relationship between all related entities is used as the normalized denominator to normalize the weight of each relationship to the range of [0, 1].

[0024] k is the iterative calculation entity e i The weight of the relationships with all entities.

[0025] Preferably, the optimization of the dynamic prompt template in step 3 adopts a feedback learning-based method to update the weight of the prompt template, and the weight update formula is:

[0026] w t =w t-1 +a(Q t -Q t-1 ),

[0027] Among them, w t and w t-1 is the weight of the current and previous round prompt templates, Q t and Q t-1 is the accuracy of the current round and the previous round of labeling tasks, and a is the learning rate;

[0028] After the weight of the prompt template is adjusted, a more targeted task description is generated by screening the frequently appearing keywords, so as to improve the labeling performance of the subsequent large model.

[0029] Preferably, the uncertainty score of the sample in step 4 is calculated using the entropy value formula:

[0030]

[0031] Among them, H(x) is the uncertainty score of sample x, K is the number of categories predicted by the model,

[0032] p(y i |x) is the sample belonging to category y i probability.

[0033] Preferably, the samples are selected for manual review based on the set uncertainty threshold r, and the threshold is dynamically adjusted by the following formula:

[0034] r t =r t-1 +y·(R target -R actual ),

[0035] Among them, r t and r t-1 is the uncertainty threshold of the current round and the previous round,

[0036] R target is the target manual review sample ratio, R actual is the current proportion of manual review samples, and y is the learning rate.

[0037] Preferably, in step 5, the text and image data are annotated by a multimodal model, and the similarity between the text feature vector and the visual feature vector is calculated by cosine similarity:

[0038]

[0039] Among them, sin(v, t) is the similarity between the visual feature vector v and the text feature vector t,

[0040] ∥v∥ and ∥t∥ are the moduli of v and t.

[0041] Preferably, the hierarchical structure of the tags is optimized by using a multimodal model, and the association between the tags is adjusted by the following formula:

[0042] Among them, p′(l i |x) is the sample x belonging to label l i The optimized probability of M is the number of all labels of sample x, p(l i |x) is the initial probability of the sample.

[0043] Preferably, after the similarity between the text feature vector and the visual feature vector is calculated, the feature alignment relationship of the multimodal model is optimized by contrastive learning, and the optimization target formula is:

[0044]

[0045] Among them, L contrastive is the loss function of contrastive learning, sin(v, t) is the similarity between the visual feature vector v and the text feature vector t, R is the temperature parameter, G is the number of negative samples,

[0046] sin(v,t i ) is the visual feature vector v and the i-th negative sample text feature vector t i The similarity.

[0047] Preferably, when optimizing the association relationship between tags, the hierarchical structure depth of the tags is considered at the same time, and the depth weight is adjusted. The optimization formula is:

[0048]

[0049] Among them, w′(l i ) indicates the label (l i ),

[0050] depth(l i ) indicates label l i The depth value in the hierarchical tag structure,

[0051] B is the adjustment coefficient, which is used to control the influence of depth on weight.

[0052] The present invention provides a fast labeling method based on an artificial intelligence large model. It has the following beneficial effects:

[0053] 1. The present invention introduces a domain knowledge graph to structure the entity and relationship information extracted from the target domain documents into knowledge triples, and matches and optimizes the annotated semantic information with the data to be annotated, thereby achieving domain adaptability and semantic consistency of the annotation results, and obtaining higher accuracy and consistency of the annotated data in professional fields.

[0054] 2. The present invention uses dynamic prompt template optimization technology and combines historical data of labeling tasks to adjust prompt content and weights in real time, thereby achieving flexible adaptability of large models to different labeling task requirements, significantly improving labeling efficiency and enhancing the model's adaptability to multi-scene labeling tasks.

[0055] 3. The present invention combines active learning with uncertainty sampling technology to screen out uncertainty samples for manual review and dynamically update model parameters, thereby optimizing the performance of large models in complex sample labeling tasks, reducing the workload of manual review and improving the overall accuracy of labeling.

[0056] 4. The present invention combines multimodal feature alignment technology with a hierarchical label optimization method to achieve collaborative annotation of multimodal data such as text and images and refined processing of annotation results, so that the multimodal data annotation is more consistent, and the annotation results have a hierarchical and refined effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0058] In order to make the technical personnel in the technical field understand the scheme of the present invention, the technical scheme in the embodiment of the present invention will be clearly and completely described below in combination with the drawings in the embodiment of the present invention. Obviously, the described embodiment is a partial embodiment of the present invention, not a complete embodiment. Based on the embodiment of the present invention, other embodiments obtained by ordinary technicians in the field without creative work should fall within the scope of protection of the present invention.

[0059] The present invention is described in detail below in conjunction with the accompanying drawings:

[0060] Example:

[0061] Please see attached Figure 1 The embodiment of the present invention provides a fast labeling method based on an artificial intelligence large model, comprising:

[0062] Step 1: Extract entity sets and relationship sets in a specific field through the document set of the target field, wherein entity recognition technology is used to identify domain-related nouns, terms and keywords in the document to form an entity set, and relationship extraction technology is used to analyze the logical and semantic associations between entities to generate a relationship set. The entity set and the relationship set are combined to construct a knowledge triple. The knowledge triple contains the association relationship between entities. The constructed knowledge triple is stored in the graph database to form a domain knowledge graph;

[0063] Step 2: Based on the knowledge graph, perform knowledge enhancement annotation on the data to be annotated, match the data to be annotated with the domain knowledge graph, and perform association matching with the corresponding entities in the knowledge graph according to the entities appearing in the data to be annotated. Optimize the annotated semantic information by querying the contextual association relationship of the entities in the knowledge graph. In the matching process, the logic of constructing entities and relationships in the knowledge triples is used to provide domain information background for step 3.

[0064] Step 3: Based on the results of knowledge-enhanced annotation, a dynamic prompt template is designed for optimization. Combined with the historical data of the annotation task, the content and structure of the prompt template are dynamically adjusted by analyzing the historical annotation results. In the dynamic prompt template, the domain information provided by the knowledge-enhanced annotation is combined to generate prompt content for a specific annotation task, so that the artificial intelligence large model can use it in the annotation task.

[0065] Step 4: Use dynamic prompt templates and large models to preliminarily annotate the data to be annotated, and screen samples in the unlabeled data through active learning. During the screening process, the uncertainty score of the sample is calculated in combination with the classification probability distribution in the annotation results, and the samples are sorted according to the uncertainty score. Some samples are selected for manual review based on the set uncertainty threshold. For the samples after manual review, the corrected data is sent back to the model training set to update the model parameters;

[0066] Step 5: For multimodal labeling tasks, use the optimized large model to label multimodal data, and extract text feature vectors and visual feature vectors from multimodal data of text and images, and achieve collaborative labeling of multimodal data through alignment and matching of feature vectors. At the same time, for the hierarchical structure of labels, build a label system with different granularities, optimize the labels according to the requirements of the labeling task, adjust the labeling probability by calculating the correlation between label hierarchies, generate labeling results and output them to the automated labeling workflow to complete the overall rapid labeling task process.

[0067] Benefits of step 1: Through entity recognition and relationship extraction technology, domain-related entity and relationship information in target domain documents is extracted, knowledge triples are effectively constructed and stored in the graph database. While structuring domain knowledge, knowledge triples provide background information support for professional fields for subsequent annotation tasks. The construction of domain knowledge graphs improves the adaptability and semantic understanding ability of annotation tasks in professional scenarios, making the annotation results more accurate and consistent.

[0068] Benefits of step 2: Match the data to be annotated with the knowledge graph, use the contextual associations in the knowledge graph to optimize the annotation semantic information, and effectively improve the degree of reliance on domain knowledge for the annotation task. In the matching process, the logical relationship based on the knowledge triples further strengthens the semantic consistency of the annotation results, especially in data scenarios involving complex semantic associations, to ensure that the annotation results conform to the logic of domain knowledge.

[0069] Benefits of step 3: Dynamic prompt templates can be optimized based on historical data of annotation tasks. By adjusting prompt content and structure in real time, the adaptability of annotation tasks to dynamically changing requirements can be improved. Combined with the domain information provided by knowledge-enhanced annotation, the generated prompt content can further improve the task understanding ability of large models and adapt large models to multi-scenario annotation requirements. By dynamically optimizing the weight of prompt templates, the efficiency of annotation tasks and the annotation performance of models can be improved, and the complexity of human intervention can be reduced.

[0070] Benefits of step 4: Active learning technology is used to screen uncertain samples in unlabeled data, and complex and ambiguous samples are manually reviewed first, effectively reducing unnecessary manual review workload. The uncertainty scoring mechanism ensures the scientific nature of sample screening, so that manual review is mainly focused on high-value samples, thereby improving overall labeling efficiency. The reviewed sample data is fed back to the model training set, the model parameters are updated, and the model performance is gradually optimized, further improving the accuracy and efficiency of automatic labeling.

[0071] Benefits of step 5: By extracting text feature vectors and visual feature vectors, and using the alignment and matching of feature vectors to achieve collaborative annotation of multimodal data such as text and images, the problem of poor semantic consistency in multimodal tasks is solved. A hierarchical labeling system is constructed in the annotation results, and the demand for fine-grained label processing is met through hierarchical optimization strategies, making the annotation results hierarchical and refined. The integration of automated annotation workflows enables the automatic generation, storage, and export of annotation data, significantly reducing the complexity of manual participation and improving the automation level of annotation tasks.

[0072] The generation of the relationship set in the knowledge triple in step 1 includes constructing the relationship weight by counting the co-occurrence frequency between entities in the target domain documents. The relationship weight is calculated by the following formula:

[0073]

[0074] Among them, w ij For entity e i and e j The relationship weight between

[0075] Frequency (e i , e j ) is entity e i and ej The number of times they co-occur in the document,

[0076] Frequency (e i ) and frequency (e j ) is entity e i and e j The total number of times it appears in the document;

[0077] Relationship weight is used to adjust the priority of entity association in the knowledge graph. The priority calculation formula is:

[0078]

[0079] Among them, p ij For entity e i To entity e j The priority of entity e i The number of all directly related entities in the knowledge graph, w ij For entity e i and e j The relationship weight between

[0080] Represents entity e i The total weight of the relationship between all related entities is used as the normalized denominator to normalize the weight of each relationship to the range of [0, 1].

[0081] k is the iterative calculation entity e i The weight of the relationships with all entities.

[0082] The benefits of building relationship weights: Converting semantic associations between entities into quantifiable weight values ​​solves the problem of strong subjectivity in the representation of entity relationships. Through co-occurrence frequency statistics, the corpus data in the domain documents is fully utilized to ensure that the knowledge graph construction has a data-driven scientific basis. Effectively identify entity pairs with close core relationships in the target domain, providing high-quality semantic support for subsequent annotation tasks.

[0083] The benefits of priority calculation based on relationship weights: Normalize the entity associations in the knowledge graph to facilitate the use of subsequent annotation tasks. Ensure that the relationship most closely associated with a certain entity has a higher semantic weight during the annotation process, improving the effect of knowledge enhancement annotation. Through iterative calculation of priorities, dynamically optimize the associations of entities in the knowledge graph, making the semantic expression ability of the knowledge graph more accurate.

[0084] Benefits of iterative optimization: Through iterative optimization, the priority of entity association relationships is gradually adjusted to dynamically enhance the expressiveness of the knowledge graph. This avoids deviations in the initial data, makes the knowledge graph more globally semantically consistent, and provides a more robust semantic background for subsequent tasks.

[0085] The optimization of the dynamic prompt template in step 3 adopts a feedback learning-based approach to update the weight of the prompt template. The weight update formula is:

[0086] w t =w t-1 +a(Q t -Q t -1),

[0087] Among them, w t and w t-1 is the weight of the current and previous round prompt templates, Q t and Q t-1 is the accuracy of the current round and the previous round of labeling tasks, and a is the learning rate;

[0088] After the weight of the prompt template is adjusted, high-frequency keywords are filtered to generate more targeted task descriptions, which are used to improve the labeling performance of subsequent large models.

[0089] The benefits of weight update based on feedback learning: The weight of the prompt template is dynamically adjusted using the historical data of the annotation task, so that the prompt content closely matches the requirements of the annotation task. The weight update amplitude is controlled by the learning rate a to ensure the stability and controllability of the prompt optimization process. The adaptability of the dynamic prompt template to the changing scenarios in the annotation task is improved, so that the annotation task can maintain efficiency and accuracy in complex and diverse situations.

[0090] Benefits of high-frequency keyword screening: The screening of high-frequency keywords focuses on the core semantic content in the annotation task, effectively avoiding redundant and inefficient information interference in the template content. For the task scenario, the generated task description is more accurate and can quickly guide the large model to complete the annotation task. It reduces the time cost of manually designing prompt templates, improves the automatic generation capability of prompt templates, and provides more efficient support for complex annotation tasks.

[0091] Benefits of improving the labeling performance of large models: Improve the ability of large models to understand labeling task instructions and reduce the labeling error rate caused by improper prompts. For multi-round labeling tasks, the optimized prompt template significantly reduces the frequency of manual intervention and improves labeling efficiency. Enhance the generalization ability of large models in complex scenarios, adapt to different task requirements, and ensure that the labeling results are more reliable and consistent.

[0092] In step 4, the entropy value calculation formula is used to calculate the uncertainty score of the sample:

[0093]

[0094] Among them, H(x) is the uncertainty score of sample x, K is the number of categories predicted by the model,

[0095] p(yi |x) is the sample belonging to category y i The probability of

[0096] Samples are selected for manual review based on the set uncertainty threshold r, and the threshold is dynamically adjusted using the following formula:

[0097] r t =r t-1 +y·(R target -R actual ),

[0098] Among them, r t and r t-1 is the uncertainty threshold of the current round and the previous round,

[0099] R target is the target manual review sample ratio, R actual is the current proportion of manual review samples, and y is the learning rate.

[0100] Benefits of entropy-based uncertainty scoring: Entropy provides a quantitative measure of sample prediction uncertainty. Samples with high entropy values ​​indicate that the model has high prediction uncertainty on the sample, and samples with problematic labeling results are identified. By analyzing the classification probability of samples, samples with poor model performance can be effectively screened to avoid wasting manual review resources on samples with high confidence in model predictions. Efficiently focus on complex, ambiguous, and ambiguous samples with high uncertainty, provide clear priority processing targets for manual review, and improve review efficiency.

[0101] Benefits of dynamically adjusting the uncertainty threshold: The threshold adjustment mechanism compares the target manual review sample ratio with the current review sample ratio, and dynamically adjusts the uncertainty threshold based on the deviation to keep the review sample volume consistent with actual needs. If the current review sample ratio exceeds the target ratio, the threshold is reduced to reduce the frequency of screening uncertainty samples and control the workload of manual review. If the review ratio is insufficient, the threshold is increased to focus more high-value samples into the review process.

[0102] Benefits of reducing manual review workload: The total number of reviewed samples is significantly reduced, the manual review workload is reduced, and at the same time, manual efforts are focused on complex samples where the labeling quality needs to be improved the most, optimizing the efficiency of manual resource use. The reviewed high-quality sample data is fed back to the model to further optimize the model parameters. The performance of the model in subsequent labeling tasks is continuously improved, and the reliance on manual review is gradually reduced. The dynamic learning mechanism enables the model to gradually adapt to the distribution of complex samples, improve the model's ability to handle high-uncertainty samples, and reduce the overall labeling error.

[0103] In step 5, the text and image data are annotated by the multimodal model, and the similarity between the text feature vector and the visual feature vector is calculated by cosine similarity:

[0104]

[0105] Among them, sin(v, t) is the similarity between the visual feature vector v and the text feature vector t,

[0106] ∥v∥ and ∥t∥ are the moduli of v and t;

[0107] After calculating the similarity between the text feature vector and the visual feature vector, the feature alignment relationship of the multimodal model is optimized through contrastive learning. The optimization target formula is:

[0108]

[0109] Among them, L contrastive is the loss function of contrastive learning, sin(v, t) is the similarity between the visual feature vector v and the text feature vector t, R is the temperature parameter, G is the number of negative samples,

[0110] sin(v,t i ) is the visual feature vector v and the i-th negative sample text feature vector t i The similarity.

[0111] Benefits of using cosine similarity calculation to achieve accurate alignment of text and visual features: Use cosine similarity to measure the directional consistency of text feature vectors and visual feature vectors, and accurately capture the semantic association between modal features. By standardizing text and visual features, the impact of feature vector scale differences on similarity calculations is avoided, ensuring that similarity values ​​have stronger semantic interpretation capabilities. Improve the feature alignment capabilities between text and image modalities, ensure the collaborative consistency of multimodal data in annotation tasks, and make annotation results more reliable and accurate.

[0112] Benefits of optimizing the feature alignment relationship of the multimodal model based on contrastive learning: Contrastive learning enhances the ability of the multimodal model to distinguish the alignment of positive sample features by maximizing the similarity of positive samples while minimizing the similarity between visual features and text features of negative samples. The temperature parameter is used to adjust the similarity distribution of positive and negative samples to ensure the stability of the optimization process. In the multimodal annotation task, the optimization of feature alignment improves the model's ability to understand complex text-image semantic associations, ensuring that the annotation results are more refined and semantically consistent.

[0113] The benefits of semantic consistency and refinement of multimodal annotation: Improve the model's ability to understand multimodal inputs and enhance the performance of cross-modal semantic consistency in annotation tasks. In practical applications, the optimized multimodal model can accurately capture the semantic relationship between multimodal data and provide annotation results for complex scenarios. By optimizing the feature alignment relationship, the common modality mismatch problem in multimodal annotation tasks is solved, significantly improving the practicality and accuracy of the annotation results.

[0114] The multimodal model is used to optimize the hierarchical structure of labels, and the relationship between labels is adjusted by the following formula:

[0115]

[0116] Among them, p′(l i |x) is the sample x belonging to label l i The optimized probability of M is the number of all labels of sample x, p(l i |x) is the initial probability of the sample;

[0117] When optimizing the association between tags, the hierarchical structure depth of the tags is considered and the depth weight is adjusted. The optimization formula is:

[0118]

[0119] Among them, w′(l i ) indicates the label (l i ),

[0120] depth(l i ) indicates label l i The depth value in the hierarchical tag structure,

[0121] B is the adjustment coefficient, which is used to control the influence of depth on weight.

[0122] The benefits of optimizing the association between labels to improve the hierarchy of the labeling results are: normalize all label probabilities of samples to ensure that the label probability distribution is within the specified range, and avoid deviations in labeling results due to inconsistent probability values ​​between labels. In the label optimization process, make full use of the initial probability and improve the scientificity and rationality of the label probability distribution through normalization. By optimizing the association between labels, ensure that the labeling results have semantic accuracy, while retaining the relative hierarchy between labels, so that the labeling is more in line with the hierarchical characteristics of the data structure.

[0123] The benefits of introducing depth weights and enhancing the hierarchical expression of label structures are as follows: the label depth weights are dynamically adjusted according to the depth value of the label in the hierarchical structure. Labels with higher depths have greater weights and give priority to expressing fine-grained semantic information, while shallow labels focus more on macro classification to ensure the global and local balance of the labeling results. The adjustment coefficient controls the impact of the depth weight on the optimization process, and the adjustment size can flexibly adapt to the needs of different labeling tasks.

[0124] Supporting multi-granularity annotation and improving the refinement of annotation results: Allowing the annotation granularity to be flexibly adjusted according to the requirements of the annotation task. Improving the practicality of the annotation results, making the annotation data suitable for macro analysis and detailed research. Ensuring the semantic consistency of the annotation results at multiple granularity levels, effectively avoiding conflicts and disconnections between annotation results of different granularities.

[0125] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A fast labeling method based on artificial intelligence large model, characterized in that: include: Step 1: Extract entity sets and relationship sets in a specific field through the document set of the target field, wherein entity recognition technology is used to identify domain-related nouns, terms and keywords in the document to form an entity set, and relationship extraction technology is used to analyze the logical and semantic associations between entities to generate a relationship set. The entity set and the relationship set are combined to construct a knowledge triple. The knowledge triple contains the association relationship between entities. The constructed knowledge triple is stored in the graph database to form a domain knowledge graph; Step 2: Based on the knowledge graph, perform knowledge enhancement annotation on the data to be annotated, match the data to be annotated with the domain knowledge graph, and perform association matching with the corresponding entities in the knowledge graph according to the entities appearing in the data to be annotated. Optimize the annotated semantic information by querying the contextual association relationship of the entities in the knowledge graph. In the matching process, the logic of constructing entities and relationships in the knowledge triples is used to provide domain information background for step 3. Step 3: Based on the results of knowledge-enhanced annotation, a dynamic prompt template is designed for optimization. Combined with the historical data of the annotation task, the content and structure of the prompt template are dynamically adjusted by analyzing the historical annotation results. In the dynamic prompt template, the domain information provided by the knowledge-enhanced annotation is combined to generate prompt content for a specific annotation task, so that the artificial intelligence large model can use it in the annotation task. Step 4: Use dynamic prompt templates and large models to preliminarily annotate the data to be annotated, and screen samples in the unlabeled data through active learning. During the screening process, the uncertainty score of the sample is calculated in combination with the classification probability distribution in the annotation results, and the samples are sorted according to the uncertainty score. Some samples are selected for manual review based on the set uncertainty threshold. For the samples after manual review, the corrected data is sent back to the model training set to update the model parameters; Step 5: For multimodal labeling tasks, use the optimized large model to label multimodal data, and extract text feature vectors and visual feature vectors from multimodal data of text and images, and achieve collaborative labeling of multimodal data through alignment and matching of feature vectors. At the same time, for the hierarchical structure of labels, build a label system with different granularities, optimize the labels according to the requirements of the labeling task, adjust the labeling probability by calculating the correlation between label hierarchies, generate labeling results and output them to the automated labeling workflow to complete the overall rapid labeling task process.

2. The rapid labeling method based on artificial intelligence large model according to claim 1 is characterized in that: The generation of the relationship set in the knowledge triple in step 1 includes constructing the relationship weight by counting the co-occurrence frequency between entities in the target domain document, and the relationship weight is calculated by the following formula: Among them, w ij For entity e i and e j The relationship weight between Frequency (e i , e j ) is entity e i and e j The number of times they co-occur in the document, Frequency (e i ) and frequency (e j ) is entity e i and e j The total number of times it occurs in the document.

3. The rapid labeling method based on artificial intelligence large model according to claim 2 is characterized in that: The relationship weight is used to adjust the priority of entity association in the knowledge graph. The priority calculation formula is: Among them, p ij For entity e i To entity e j The priority of entity e i The number of all directly related entities in the knowledge graph, w ij For entity e i and e j The relationship weight between Represents entity e i The total weight of the relationship between all related entities is used as the normalized denominator to normalize the weight of each relationship to the range of [0, 1]. k is the iterative calculation entity e i The weight of the relationships with all entities.

4. The rapid labeling method based on artificial intelligence large model according to claim 1 is characterized in that: The optimization of the dynamic prompt template in step 3 adopts a feedback learning-based method to update the weight of the prompt template. The weight update formula is: w t =w t-1 +a(Q t -Q t-1 ), Among them, w t and w t-1 is the weight of the current and previous round prompt templates, Q t and Q t-1 is the accuracy of the current round and the previous round of labeling tasks, and a is the learning rate; After the weight of the prompt template is adjusted, a more targeted task description is generated by screening the frequently appearing keywords, so as to improve the labeling performance of the subsequent large model.

5. The rapid labeling method based on artificial intelligence large model according to claim 1 is characterized in that: In step 4, the entropy value calculation formula is used to calculate the uncertainty score of the sample: Among them, H(x) is the uncertainty score of sample x, K is the number of categories predicted by the model, p(y i |x) is the sample belonging to category y i probability.

6. The rapid labeling method based on artificial intelligence large model according to claim 5 is characterized in that: The samples are selected for manual review based on the set uncertainty threshold r, and the threshold is dynamically adjusted by the following formula: r t =r t-1 +y·(R target -R actual ), Among them, r t and r t-1 is the uncertainty threshold of the current round and the previous round, R target is the target manual review sample ratio, R actual is the current proportion of manual review samples, and y is the learning rate.

7. The rapid labeling method based on artificial intelligence large model according to claim 1 is characterized in that: In step 5, the text and image data are annotated by a multimodal model, and the similarity between the text feature vector and the visual feature vector is calculated by cosine similarity: Among them, sin(v, t) is the similarity between the visual feature vector v and the text feature vector t, ||v|| and ||t|| are the moduli of v and t.

8. The rapid labeling method based on artificial intelligence large model according to claim 7 is characterized in that: The multimodal model is used to optimize the hierarchical structure of the labels, and the association between the labels is adjusted by the following formula: Among them, p′(l i |x) is the sample x belonging to label l i The optimized probability of M is the number of all labels of sample x, p(l i |x) is the initial probability of the sample.

9. The rapid labeling method based on artificial intelligence large model according to claim 7 is characterized in that: After the similarity between the text feature vector and the visual feature vector is calculated, the feature alignment relationship of the multimodal model is optimized by contrastive learning, and the optimization target formula is: Among them, L contrastive is the loss function of contrastive learning, sin(v, t) is the similarity between the visual feature vector v and the text feature vector t, R is the temperature parameter, G is the number of negative samples, sin(v,t i ) is the visual feature vector v and the i-th negative sample text feature vector t i The similarity.

10. The rapid labeling method based on artificial intelligence large model according to claim 8, characterized in that: When optimizing the association relationship between tags, the hierarchical structure depth of the tags is considered at the same time, and the depth weight is adjusted. The optimization formula is: Among them, w′(l i ) indicates the label (l i ), depth(l i ) indicates label l i The depth value in the hierarchical tag structure, B is the adjustment coefficient, which is used to control the influence of depth on weight.

Images