An intelligent hybrid labeling method for aligning traditional art resource values

Abstract

The application discloses a kind of traditional art resource value alignment intelligent mixed labeling method, comprising: the modal feature of each modality content of traditional art resource is obtained, and multiple modal evidence unit is generated;With the traditional art resource as unit, establish evidence set, when there is fact conflict between evidence unit, all conflict evidence is retained and conflict association is recorded;Search only selects the identification with direct reference relationship with evidence unit, and outputs empty identification when evidence is insufficient;According to candidate identification, evidence unit, external knowledge base and application scene, the identification accuracy, fact support degree and scene adaptability of candidate identification are verified by independent verification mechanism, and each verification result is output respectively;The uncertainty degree of the candidate identification is quantified, and the candidate identification is divided into automatic pass, manual review or return correction;The candidate identification, its evidence reference and verification result are output.The application is adapted to large quantities of traditional art resource digitization labeling scene.

Description

Technical Field

[0001] This application relates to an intelligent hybrid annotation method for aligning the values ​​of traditional art resources, belonging to the field of traditional art resource annotation technology. Background Technology

[0002] Traditional art resources encompass calligraphy and painting, sculpture, opera, music, dance, ancient books, craft records, and related research materials. Their digital carriers integrate multi-dimensional information including visual, textual, audio, and temporal elements. Manually labeling each item individually makes it difficult to balance processing efficiency, evidence integrity, and standardized labeling. While current multimodal models can assist in identifying artistic elements such as patterns, inscriptions, lyrics, and actions, the output results generally lack traceable positioning evidence. Traditional art resource annotation technologies still have significant shortcomings. Most solutions rely on single-modal feature matching, failing to deeply integrate multi-dimensional art information and struggle to construct a structured, traceable evidence system. They commonly suffer from insufficient annotation evidence and an inability to link to the spatiotemporal positioning information of the original resources. Furthermore, existing technologies lack mechanisms for identifying and preserving evidence conflicts. In scenarios with diverse and contradictory art resource information, key information is easily lost, leading to annotation distortion. Moreover, the annotation judgment logic is simplistic, lacking a multi-dimensional, layered verification system. It fails to consider the annotation uncertainty caused by evidence completeness and modal conflict, making it difficult to align the values ​​of art resources across different scenarios. In addition, traditional annotation methods lack robust permission control and iteration mechanisms, easily resulting in the direct generation of official labels, modification, or deletion of original evidence. Manual correction experience cannot effectively feed back into algorithm optimization, ultimately leading to low annotation accuracy, adaptability, and credibility, severely limiting the efficient implementation of digital archiving, intelligent interpretation, and cultural dissemination of traditional art resources. Summary of the Invention

[0003] According to one aspect of this application, an intelligent hybrid annotation method for aligning the values ​​of traditional art resources is provided, which is suitable for scenarios involving the digital annotation of large volumes of traditional art resources.

[0004] A smart hybrid annotation method for aligning the values ​​of traditional art resources includes: Traditional art resources are acquired and modal features of each modality are extracted. The modal features are bound to the spatial or temporal location information of the original resources to generate multimodal evidence units with a unified data structure that can be traced back to the original location. An evidence set is established with the traditional art resources as the unit. When there is a factual conflict between different evidence units, all conflicting evidence is retained in the evidence set and the conflict association is recorded. Based on the evidence unit, a matching candidate identifier is retrieved from a preset controlled identifier system. The retrieval process only allows the selection of identifiers that have a direct reference relationship with the evidence unit. When the evidence is insufficient to support any identifier, an empty identifier is output. The generation conclusion of the candidate identifier is isolated from independent verification. An independent verification mechanism independently verifies the identification accuracy, factual support and scenario adaptability of the candidate identifier based on the candidate identifier, the evidence unit, the external knowledge base and the preset application scenario, and outputs the verification results for each item respectively. Based on the verification results, the completeness of the evidence unit referenced by the candidate identifier, and the degree of conflict between different modal evidence units, the uncertainty of the candidate identifier is comprehensively quantified, and the candidate identifier is divided into automatic approval, manual review, or return for correction according to the uncertainty. The system receives correction decisions made by the manual review process for the diverted review identifiers. After the correction decisions are approved and regression testing is completed, they are included in the feedback sample set for subsequent iterative updates of retrieval or verification strategies. Output the candidate identifier, the evidence unit it references, and the verification results; Each execution step operates within the preset permission boundaries, and each execution step does not directly release an official version or delete original evidence.

[0005] Furthermore, the extraction of modal features for each modal content includes: Each modal encoder retains the spatial, character, or temporal structure of its own mode and projects it to a unified feature space through a modal adapter: ; ; Where m represents the resource mode; i represents the i-th region, segment, or time window in that mode; Indicates the original input unit; A dedicated encoder representing mode m; This represents the native features output by the encoder; Indicates a modal adapter; The 'd' represents the adapted features; 'd' represents the unified feature dimension. Let represent the d-dimensional real space.

[0006] Furthermore, the evidence unit is represented as follows: ; in, This represents the evidence marker for the i-th modality m; the semicolon in square brackets indicates vector concatenation; Represents a region, character range, or time window; Indicates the resolution confidence level; This represents traceability information consisting of resource source, model version, and extraction time; This represents a function that encodes location information, confidence level, and source fields into a vector; D represents the uniform dimension of evidence labeling.

[0007] Furthermore, the step of retrieving matching candidate identifiers from a preset controlled identifier system based on evidence includes: All evidence units are organized into an evidence sequence according to resource ownership and temporal correlation. The evidence sequence is then fused using a masked attention mechanism, with missing modalities skipped by a mask. ; Where E represents the evidence sequence consisting of all evidence markers; N represents the number of evidence markers; The parameter is The evidence sequence fusion model; H represents the fused hidden state sequence; represents the i-th hidden state; a represents the trainable attention query vector; Let z represent the normalized weight of the i-th piece of evidence; z represents the resource-level fusion representation; i and All are evidence indexes; (·) represents an exponential function; Indicates transpose; The resource-level fused representation is projected onto the identifier vector space, and the retrieval score is calculated by combining the semantic relationships in the identifier ontology: ; ; Where j represents the standard label index; Let represent the retrieval score of the j-th standard tag; P represents the matrix that projects the fused representation onto the tag vector space; Indicates the first A standard label vector; cos(·) represents the cosine similarity; The score represents the synonym, hyponym, exclusivity, or branching relationship within the tag body; Indicates the weight of the ontology relation item; Indicates the number of candidate tags; This represents the candidate set consisting of the K highest-scoring labels; (·) indicates that the top K items are selected based on their scores.

[0008] Furthermore, the independent verification of the candidate identifier's accuracy, factual support, and scenario adaptability includes: performing joint prediction on the candidate identifier using a multi-task review model. ; ; in, (·) indicates that the parameter is A multi-task review model; This represents the prediction of the correctness of candidate label j; This indicates that cultural facts support the prediction; This indicates a prediction of the fit between the dissemination scenario and the value. This represents the probability distribution of tag hierarchy categories; This indicates the probability distribution of the processed routes; This represents the probability distribution of evidence citations. This represents the review representation of candidate label j; The parameter is The candidate label agent encoding function; This represents a cultural knowledge fragment that is related to candidate label j and retains its source. This indicates the applicable value rule activated by the current task; s represents the propagation scenario encoding; The multi-task review model is trained using a joint training loss: ; in, This indicates the total losses from the joint training exercise; , and represents the binary cross-entropy loss for label correctness, cultural fact support, and dissemination scenario adaptation tasks, respectively; The cross-entropy loss represents the classification by label hierarchy; This represents the cross-entropy loss used for route classification. Indicates the loss of evidence cited for supervision; each Represents the non-negative weight of the corresponding task, and satisfies .

[0009] Furthermore, the completeness of the evidence unit is: ; Where j represents the candidate label index; Indicates the completeness of evidence for candidate label j; Indicates whether the i-th piece of evidence is referenced by candidate label j; The calibrated confidence level of the i-th piece of evidence is represented by ; i represents the evidence index; the max(·) function is used to avoid a zero denominator when there is no evidence.

[0010] Furthermore, the uncertainty of the comprehensive quantization candidate identifier includes: The statistical dispersion of the routing probability is estimated through multiple random inactivation inferences or model ensembles, and cross-modal evidence conflicts and conflicts between individual verification results are included as independent quantification terms in the degree of uncertainty: ; in, Indicates the first The second random inactivation reasoning or the first The model provides a routing probability vector for candidate label j; L represents the number of inferences or the number of models. Indicates the average route probability; 'b' indicates uncertainty regarding candidate labels; 'b' indicates the route category during automatic approval, expert review, or rejection of revisions. Indicates the weight of the probability variance term; Represents the L2 norm; Indicates the degree of conflict between cross-modal evidence or the conclusions of sub-examinations; Indicates the weight of the conflicting terms.

[0011] Furthermore, the step of triaging candidate identifiers into automatic approval, manual review, or rejection for correction based on the degree of uncertainty includes: ; in, This indicates the action to take for candidate label j; AUTO means automatic approval; REVIEW means submit to experts; REVISE means reject the current candidate and return for revision. This represents the minimum value of the three prediction results in the joint prediction: label accuracy, cultural fact support, and adaptation to the communication scenario. , and These represent the sub-item quality threshold, evidence completeness threshold, and uncertainty threshold, respectively; This indicates a rule that cannot be automatically passed. This indicates a miss; Indicates that the manual review criteria have been met; 1 indicates that at least one manual review condition has been met.

[0012] Furthermore, the expert-weighted adoption rate of the candidate identifiers is: ; in, The expert-weighted adoption rate of candidate label j is represented by t; the expert index is represented by t. This indicates a non-negative weight predetermined based on the expert's professional field and historical consistency, and at least one of these weights is greater than zero. These represent whether expert t rejects or accepts candidate label j, respectively; expert decisions regarding modification, merging, or splitting should be recorded using separate structured fields, not compressed into a single file. The binary value of .

[0013] Furthermore, automatic processing will cease and manual review will commence when any of the following conditions occur: There is no evidence unit that can pinpoint the specific location of the original resource to support the judgment; The factual conflicts or contradictions in the sources of knowledge among different evidentiary units exceed a predetermined threshold; Applicable rules that cannot be automatically passed; New identifiers or new identifier levels have appeared. The preset maximum number of retries has been reached and automatic repair has failed.

[0014] The beneficial effects that this application can produce include: This application provides an intelligent hybrid annotation method for aligning the values ​​of traditional art resources. By constructing multimodal evidence units bound to spatiotemporal location information, it forms a structured, traceable, and standardized data structure. This solves the problems of missing evidence, lack of traceability, and inconsistent labels in traditional art resource annotation, improving the authenticity and traceability of multimodal annotation results. By fully retaining and recording conflict evidence and relationships between resources, it avoids the drawbacks of traditional annotation methods, such as losing contradictory information and causing annotation distortion, thus fully preserving the diverse and complex cultural information of traditional art resources. Furthermore, it employs a matching rule that limits retrieval to direct reference markers and uses empty markers as a fallback. It employs an independent, layered verification mechanism that considers accuracy, factual support, and scenario adaptability. By combining evidence completeness and modal conflict quantification to measure annotation uncertainty, it achieves a streamlined process of automatic approval, manual review, and correction of annotation results. This effectively improves annotation accuracy, scenario adaptability, and value alignment, addressing the shortcomings of traditional annotation judgment logic being singular and unreliable. By limiting the permission boundaries of each operation step, it prevents the tampering, deletion, and informal publication of original evidence. The manual review and correction results are iteratively optimized for retrieval and verification strategies, continuously improving the intelligence level of the annotation system and adapting it to large-scale digital annotation scenarios for traditional art resources. Attached Figure Description

[0015] Figure 1 A flowchart of an intelligent hybrid annotation method for aligning the values ​​of traditional art resources in one embodiment of this application; Figure 2 This is a diagram illustrating the division of roles among intelligent agents and the human-machine organizational relationship in one embodiment of this application; Figure 3 This is a diagram illustrating multimodal coding, modality adaptation, and evidence sequence fusion in one embodiment of this application. Figure 4 This application provides an independent value review, uncertainty assessment, and manual triage diagram for one implementation method. Detailed Implementation

[0016] The present application is described in detail below with reference to the embodiments, but the present application is not limited to these embodiments.

[0017] See Figure 1-4,like Figure 1 As shown, an intelligent hybrid annotation method for aligning the values ​​of traditional art resources includes: Traditional art resources are acquired and modal features of each modality are extracted. The modal features are bound to the spatial or temporal location information of the original resources to generate multimodal evidence units with a unified data structure that can be traced back to the original location. An evidence set is established with the traditional art resources as the unit. When there is a factual conflict between different evidence units, all conflicting evidence is retained in the evidence set and the conflict association is recorded. Based on the evidence unit, a matching candidate identifier is retrieved from a preset controlled identifier system. The retrieval process only allows the selection of identifiers that have a direct reference relationship with the evidence unit. When the evidence is insufficient to support any identifier, an empty identifier is output. The generation conclusion of the candidate identifier is isolated from independent verification. An independent verification mechanism independently verifies the identification accuracy, factual support and scenario adaptability of the candidate identifier based on the candidate identifier, the evidence unit, the external knowledge base and the preset application scenario, and outputs the verification results for each item respectively. Based on the verification results, the completeness of the evidence unit referenced by the candidate identifier, and the degree of conflict between different modal evidence units, the uncertainty of the candidate identifier is comprehensively quantified, and the candidate identifier is divided into automatic approval, manual review, or return for correction according to the uncertainty. The system receives correction decisions made by the manual review process for the diverted review identifiers. After the correction decisions are approved and regression testing is completed, they are included in the feedback sample set for subsequent iterative updates of retrieval or verification strategies. Output the candidate identifier, the evidence unit it references, and the verification results; Each execution step operates within the preset permission boundaries, and each execution step does not directly release an official version or delete original evidence.

[0018] Specifically, traditional art resources refer to traditional paintings, calligraphy, operas, music, dances, crafts, sculptures, ancient books, and related explanatory materials preserved in the form of images, text, audio, video, or a combination thereof. Value alignment is determined based on the cultural context, narrative identity, dissemination scenario, and applicable norms of the resources, judging the appropriateness of labels and subsequent use. This judgment is stored separately from cultural fact verification and label correctness. Intelligent hybrid labeling is completed by intelligent agents, which can automatically parse, retrieve, generate, and verify the data. Experts handle disputes, high risks, new labels, and version changes. Intelligent agents are processing units that execute tasks under clearly defined role instructions, input fields, allowed tools, output formats, retry counts, stopping conditions, and permission boundaries. Intelligent agents can consist of one or more models and deterministic tools. The evidence ledger is a collection of evidence organized by resource identifiers. Each record includes modality, location, extraction result, confidence level, source, model version, timestamp, and subsequent usage relationship. The dissemination scenario refers to the specific environment in which the resources are planned to be used, such as in-museum research, educational exhibitions, public recommendations, activities for minors, or commercial dissemination. The same content can trigger different rules in different scenarios. Traditional art resources encompass multiple modalities, including images, audio / video, and text. Feature extraction must be inseparable from the original material; each piece of evidence must be bound to a feature value and its original location (image coordinates / video timestamp / text character range) to ensure that any subsequent annotation can directly jump to the corresponding segment in the original material. All evidence from the same art resource constitutes an evidence set. In cases of conflicting facts across multiple pieces of evidence, such as conflicting historical records regarding the dating of an artifact, all evidence is retained, recording only the relationship between the conflict and the evidence itself, preventing the machine from directly discarding valuable dissenting historical materials. The controlled labeling system involves a pre-defined standardized label library, including ontological labels, hierarchical relationships, and constraint rules within the traditional art field; mandatory... Candidate labels must be directly supported by evidence units; models are not allowed to create labels without basis. If the existing evidence is insufficient to support any label, an empty label is output, and no guessing results are forced to avoid machine illusions. The candidate label generation module is only responsible for producing candidates and cannot verify its own output. This is done by a completely independent verification subsystem. The label accuracy is judged by whether the label itself matches the resource object; the factual support is judged by whether there is cultural and historical data or physical evidence to support it; and the scenario adaptability is judged by whether the label is suitable for the target application scenario. Among these, the value requirements of different scenarios such as popular science dissemination, academic research, and cultural and creative development are different. The results of the three verifications are output separately, and the interpretability of each item is preserved.

[0019] Furthermore, by comprehensively considering the independent verification output of the input, the completeness of the cited evidence, and the degree of conflict between multimodal evidence, an uncertainty index is calculated. Based on the uncertainty, candidate identifiers are divided into three flow channels: automatic approval, manual review, and return for correction, achieving intelligent division of labor between humans and machines. Manual review provides decisions such as modification, addition, and deletion. After approval and regression testing, qualified candidates enter the feedback sample set to back-optimize the retrieval model and verify the model parameters, continuously improving the quality of subsequent annotations. The output is not just a simple label but also includes the candidate identifier, the evidence units it references, and the complete set of verification results to ensure auditability. No algorithm module has the right to directly launch the official version or delete the original art resource evidence. The official version release requires upper-level approval, and the original evidence is permanently retained to prevent the loss of original material information due to algorithm errors.

[0020] The modal features extracted from each modal content include: Each modal encoder retains the spatial, character, or temporal structure of its own mode and projects it to a unified feature space through a modal adapter: ; ; Where m represents the resource mode; i represents the i-th region, segment, or time window in that mode; Indicates the original input unit; A dedicated encoder representing mode m; This represents the native features output by the encoder; Indicates a modal adapter; The 'd' represents the adapted features; 'd' represents the unified feature dimension. Let represent the d-dimensional real space.

[0021] Specifically, for image resources, the encoding objects include the overall image, local regions, patterns, inscriptions, and seals. Visual Transformers, hybrid encoders combining convolutional neural networks and Transformers can be used, combined with OCR and region detection tools to output visual features, object attributes, recognized text, and their bounding boxes, polygons, or page number information. For text resources, the encoding objects include exhibition labels, inscriptions, ancient books, explanatory texts, and research materials. Contextual language encoders can be used, combined with layout parsing, entity recognition, relation extraction, and citation parsing tools to output semantic features, entity relationships, time and location information, source information, page numbers, paragraphs, and character ranges. For audio resources… The encoding objects include speech, singing style, musical events, timbre, and rhythm. Acoustic Transformer, Conformer, or audio-text aligned encoders can be used, combined with automatic speech recognition and audio segmentation tools, to output acoustic features, transcribed text, character or event information, start and end times, and channel information. For video resources, the encoding objects include shots, keyframes, character actions, subtitles, speech, and events. Video Transformer or spatiotemporal encoders can be used, combined with shot segmentation, keyframe extraction, OCR, automatic speech recognition, and audio-visual alignment tools, to output spatiotemporal features, action events, subtitles, speech, and corresponding shot numbers, frame numbers, and time windows. The original feature spaces of different modalities are not the same. This application does not require the modal encoders to share parameters, nor does it directly splice unprocessed original features. Each encoder first retains the spatial, character, or temporal structure of its own modality, and then projects it to a unified dimension through a modal adapter. When m is an image modality, i is a local region cropped from the image; when m is a text modality, i is a character range of a colophon or historical material; when m is an audio / video modality, i is a time window segment. A dedicated encoder for modality m. Different modalities use models suitable for themselves. The encoder must preserve the original structure of the modality: spatial structure for images, character order for text, and temporal sequence for audio and video. It should not completely erase or lose structural information and should output original features. ; The modality adapter is a projection network that maps native features of different modalities (with varying dimensions) to the same d-dimensional real feature space, resulting in adapted features. This application preserves the corresponding modal structure through a dedicated encoder and then unifies the feature dimensions through a modal adapter. Image regions, text intervals, audio segments, and video windows can have different numbers, but the adapted feature vectors have the same dimension, which is used for aligning the feature dimensions of different modalities. This facilitates cross-modal comparison and fusion, while not destroying the spatial / temporal semantic information of each modality itself, thus providing underlying features for constructing evidence units.

[0022] The evidence unit is represented as follows: ; in, This represents the evidence marker for the i-th modality m; the semicolon in square brackets indicates vector concatenation; Represents a region, character range, or time window; Indicates the resolution confidence level; This represents traceability information consisting of resource source, model version, and extraction time; This represents a function that encodes location information, confidence level, and source fields into a vector; D represents the uniform dimension of evidence labeling.

[0023] Specifically, the evidence ledger agent converts the parsing results of each modality into evidence tags with unified fields. The agent does not change the parsed content, only adding fields such as location, confidence level, source, and version, and verifies the evidence citation relationship. Each piece of evidence must be able to return its specific location in the original resource. Model descriptions that cannot be located can serve as subsequent search clues, but must not be the sole basis for automatically passing candidate tags. The encoding function encodes three types of non-semantic meta-information into vectors: It represents a region, character range, or time window, specifically positioning information such as image coordinates, text start and end positions, and video timestamps, used to jump back to the original footage; The confidence score represents the degree of certainty the model has when extracting this piece of evidence; This represents the source information, consisting of the resource origin, model version, and extraction time, which can distinguish which evidence comes from older model versions; ";" represents vector concatenation, which concatenates the semantic feature vector and the meta-information encoding vector to obtain a complete evidence unit of dimension D. .

[0024] As shown in Table 1, the fields of the evidence ledger are explained. The evidence ledger uses resource_id as the primary key and stores at least the fields evidence_id, modality, locator, content, confidence, source, extractor_version, created_at, and conflict_with. When the contents of two pieces of evidence are inconsistent, the evidence ledger retains both pieces of evidence and records the conflict relationship through the conflict_with field. The evidence ledger agent does not delete or overwrite one of them automatically.

[0025] Table 1 Explanation of Evidence Ledger Fields

[0026] Among them, `resource_id` is used to establish the association between evidence records and traditional art resources, `evidence_id` is used to uniquely identify a single piece of evidence, and `locator` uses different positioning structures according to the evidence modality. For multiple pieces of evidence in the same resource, independent evidence records are established for each. When different pieces of evidence are inconsistent in terms of content, time, space, identity of the person, or source of knowledge, a two-way or traceable conflict reference is established through the `conflict_with` field, without overwriting or deleting the original evidence. `confidence` only represents the calibration confidence level of the evidence extraction process and is not directly used as the basis for automatic approval of candidate labels.

[0027] The step of retrieving matching candidate identifiers from a preset controlled identifier system based on evidence includes: All evidence units are organized into an evidence sequence according to resource ownership and temporal correlation. The evidence sequence is then fused using a masked attention mechanism, with missing modalities skipped by a mask. ; Where E represents the evidence sequence consisting of all evidence markers; N represents the number of evidence markers; The parameter is The evidence sequence fusion model; H represents the fused hidden state sequence; represents the i-th hidden state; a represents the trainable attention query vector; Let z represent the normalized weight of the i-th piece of evidence; z represents the resource-level fusion representation; i and All are evidence indexes; (·) represents an exponential function; Indicates transpose; The resource-level fused representation is projected onto the identifier vector space, and the retrieval score is calculated by combining the semantic relationships in the identifier ontology: ; ; Where j represents the standard label index; Let represent the retrieval score of the j-th standard tag; P represents the matrix that projects the fused representation onto the tag vector space; Indicates the first A standard label vector; cos(·) represents the cosine similarity; The score represents the synonym, hyponym, exclusivity, or branching relationship within the tag body; Indicates the weight of the ontology relation item; Indicates the number of candidate tags; This represents the candidate set consisting of the K highest-scoring labels; (·) indicates that the top K items are selected based on their scores.

[0028] Specifically, all evidence units under the same traditional art resource are assembled into sequence E; different resource modalities may be missing, for example, some intangible cultural heritage materials only have pictures and no text, and missing modal evidence is masked and not included in the calculation; The evidence fusion model outputs a hidden state sequence H; a query vector a can be trained for attention scoring. The normalized weights represent the importance of each piece of evidence to the entire resource; z is the global fusion vector representation of the entire art resource, obtained by weighted summation of all evidence. The P matrix projects the global resource vector z onto the label vector space; Calculate the semantic similarity between the resource and the j-th tag in the controlled library; Ontology relationship scoring incorporates ontology rules from a tag knowledge base, including synonym tags, parent-child hierarchy, mutually exclusive tags, and branch constraints; μ modulates the influence of ontology knowledge on the score; for example, if the "Peking Opera" tag has been selected, the ontology rules will increase the score of the "Opera" higher-level tag and decrease the score of the "Kunqu" mutually exclusive tag; hard rules of domain knowledge are integrated into the retrieval score, rather than simply relying on semantic vectors. Based on the overall search scores, the candidate set is obtained by selecting the tags with the highest Top-K scores. This refers to the candidate identifier that will be verified later.

[0029] In this application, evidence tags are arranged into evidence sequences according to resources, time, and association. The evidence sequence fusion model can use a masked Transformer to eliminate the need to fill in missing modalities with fictitious content. The model's attention weights are used to select evidence related to the current annotation target, but the attention weights themselves cannot replace the location reference of the original evidence. The candidate label agent first retrieves labels from the controlled label system, and then generates, merges, or proposes new labels from the top K labels with the highest scores. The retrieval process considers both vector similarity and label ontology relationships to avoid mixing labels from different branches into one category simply because they are similar in wording.

[0030] For each candidate label, construct a review representation:

[0031] in, This represents the review representation of candidate label j; The parameter is The candidate label agent encoding function; the semicolon in square brackets indicates feature concatenation; This represents a cultural knowledge fragment that is related to candidate label j and retains its source. s represents the applicable value rule activated by the current task; s represents the propagation scenario code.

[0032] The independent verification of the candidate identifier's accuracy, factual support, and scenario adaptability includes: performing joint prediction on the candidate identifier using a multi-task review model. ; ; in, (·) indicates that the parameter is A multi-task review model; This represents the prediction of the correctness of candidate label j; This indicates that cultural facts support the prediction; This indicates a prediction of the fit between the dissemination scenario and the value. This represents the probability distribution of tag hierarchy categories; This indicates the probability distribution of the processed routes; This represents the probability distribution of evidence citations. This represents the review representation of candidate label j; The parameter is The candidate label agent encoding function; This represents a cultural knowledge fragment that is related to candidate label j and retains its source. This indicates the applicable value rule activated by the current task; s represents the propagation scenario encoding; The multi-task review model is trained using a joint training loss: ; in, This indicates the total losses from the joint training exercise; , and represents the binary cross-entropy loss for label correctness, cultural fact support, and dissemination scenario adaptation tasks, respectively; The cross-entropy loss represents the classification by label hierarchy; This represents the cross-entropy loss used for route classification. Indicates the loss of evidence cited for supervision; each Represents the non-negative weight of the corresponding task, and satisfies .

[0033] Specifically, the independent value review agent does not read the final conclusion field of the candidate label agent, but only reads the candidate label name, location evidence, cultural knowledge source, applicable rules and dissemination scenario. By isolating the candidate generation conclusion from the independent review process, the probability of the two agents repeating the same error is reduced. The independent value review result includes at least the label correctness, cultural fact support, dissemination scenario and value suitability, label hierarchical position and the need for manual review. The above results are saved separately.

[0034] Training samples are organized into units of associated records consisting of resources, candidate labels, evidence sets, dissemination scenarios, and expert decisions. Positive samples come from records confirmed by experts and with complete evidence. Difficult negative samples can be constructed through methods such as mismatch of synonymous labels, cross-branch label replacement, mismatch of cultural origins, confusion between role stance and work stance, time window offset, and insufficient evidence. The training set, validation set, and test set are grouped and segmented according to the source of the work or resource to avoid different segments of the same work being assigned to the training set and test set respectively, which would cause data leakage.

[0035] This application shares underlying features across multiple tasks, simultaneously performing verification, hierarchical checks, route prediction, and evidence attribution. During training, the weights can be optimized by adjusting λ between label accuracy, rigorous cultural facts, and alignment with scene values, thus adapting to the humanistic priority requirements of traditional art annotation.

[0036] The completeness of the evidence unit is: ; Where j represents the candidate label index; Indicates the completeness of evidence for candidate label j; Indicates whether the i-th piece of evidence is referenced by candidate label j; The calibrated confidence level of the i-th piece of evidence is represented by ; i represents the evidence index; the max(·) function is used to avoid a zero denominator when there is no evidence.

[0037] Specifically, such as Figure 4 As shown, the uncertainty-splitting agent calculates the evidence completeness of candidate labels and estimates the instability of the processing route probability through multiple random inactivation inferences, model sets, or other calibration methods. Cross-modal conflicts are counted separately, and conflicts in other modalities are not offset by increasing the confidence of one modality.

[0038] Completeness of evidence It is used to measure how many high-confidence evidence units support the current candidate identifier, and is an important input indicator for uncertainty calculation and automatic pass judgment; Binary value: 1 represents that evidence i is cited by label j; 0 represents that it is not cited. The calibration confidence level of the i-th piece of evidence is derived from the metadata within the evidence unit. This represents the sum of the confidence levels of all cited evidence for the candidate label. This represents the total number of cited pieces of evidence. If no evidence is cited, the denominator is forced to equal 1 to prevent division by zero errors. The completeness of evidence is equivalent to the average confidence level of the cited evidence, ranging from 0 to 1. For example, if a label cites 3 pieces of evidence with confidence levels of 0.9, 0.8, and 0.2 respectively, then the completeness = (0.9 + 0.8 + 0.2) / 3 = 0.63. When the completeness is very low, it means that there is little evidence supporting the label or that the evidence itself is not credible. This candidate is more likely to undergo manual review and cannot be automatically approved.

[0039] The uncertainty of the comprehensive quantization candidate identifier includes: The statistical dispersion of the routing probability is estimated through multiple random inactivation inferences or model ensembles, and cross-modal evidence conflicts and conflicts between individual verification results are included as independent quantification terms in the degree of uncertainty: ; in, Indicates the first The second random inactivation reasoning or the first The model provides a routing probability vector for candidate label j; L represents the number of inferences or the number of models. Indicates the average route probability; 'b' indicates uncertainty regarding candidate labels; 'b' indicates the route category during automatic approval, expert review, or rejection of revisions. Indicates the weight of the probability variance term; Represents the L2 norm; Indicates the degree of conflict between cross-modal evidence or the conclusions of sub-examinations; Indicates the weight of the conflicting terms.

[0040] Specifically, the uncertainty of candidate labels As the core basis for separating human and machine traffic; sources of uncertainty include insufficient confidence in the model's predictions, unstable model outputs, and conflicting evidence / verification conclusions.

[0041] To obtain the statistical distribution of the model output, use Dropout to randomly deactivate the model multiple times during inference, or use multiple models to form a model ensemble, run it L times, and obtain L sets of route probability vectors. ; take the average to get .

[0042] It consists of the sum of three terms: The complement of the largest category in the average route probability. If the model is very certain about a certain route category, the max value is close to 1, and this term is close to 0; if the model is volatile, this term has a large value. The variance between multiple inference outputs measures the stability of the model's internal output; η is the weight; the greater the difference between multiple inference results, the higher the uncertainty. The degree of conflict between cross-modal evidence or sub-item review conclusions includes two types of conflict: contradictions between different modal evidence units such as images, texts, and audio / video, and conflicts between the three verification results, such as high accuracy but low cultural factual support; ξ controls the contribution of conflict to the total uncertainty.

[0043] The larger the overall size, the less reliable the candidate identifier is, and the more it needs to be handled manually. Traditional art often presents historical contradictions, therefore, conflicts are specifically addressed. Incorporate uncertainty, rather than just looking at model scores.

[0044] The process of categorizing candidate identifiers into automatic approval, manual review, or rejection for correction based on the degree of uncertainty includes: ; in, This indicates the action to take for candidate label j; AUTO means automatic approval; REVIEW means submit to experts; REVISE means reject the current candidate and return for revision. This represents the minimum value of the three prediction results in the joint prediction: label accuracy, cultural fact support, and adaptation to the communication scenario. , and These represent the sub-item quality threshold, evidence completeness threshold, and uncertainty threshold, respectively; This indicates a rule that cannot be automatically passed. This indicates a miss; Indicates that the manual review criteria have been met; 1 indicates that at least one manual review condition has been met.

[0045] Specifically, the conditions for manual review include newly added labels, conflicting evidence, high-risk rule hits, contradictory knowledge sources, or uncertain label hierarchy positions. Each threshold is determined jointly by the validation set results and acceptable risk, and is not directly set using training set scores. For high-risk propagation scenarios, this can improve... and and reduce The hit result of the rule cannot be automatically passed and takes priority over the model score.

[0046] The priority of the judgment logic includes: Automatic passage: All conditions must be met simultaneously: If all conditions are met simultaneously, the machine will directly output the label. (The following conditions must be met: ≥sub-threshold, evidence completeness ≥ threshold, uncertainty ≤ threshold, rule prohibiting automatic passage not met, no manual review triggered.) Manual review: as long as 1. (If any manual review trigger condition is met, the process will directly enter the expert review channel without considering other indicators;) Return for correction: If all of the above two conditions are not met, return to the upstream module and request a new search for candidate identifiers or supplementary evidence units.

[0047] Therefore, traditional art annotation cannot rely on a single score. Instead, it should use multiple conditions and AND constraints. If any one condition is not met, the annotation should not be automatically uploaded.

[0048] The expert-weighted adoption rate of the candidate identifier is: ; in, The expert-weighted adoption rate of candidate label j is represented by t; the expert index is represented by t. This indicates a non-negative weight predetermined based on the expert's professional field and historical consistency, and at least one of these weights is greater than zero. These represent whether expert t rejects or accepts candidate label j, respectively; expert decisions regarding modification, merging, or splitting should be recorded using separate structured fields, not compressed into a single file. The binary value of .

[0049] Specifically, for samples requiring manual review, the expert interface presents the original resources, location evidence, candidate labels, sub-scores, cultural knowledge sources, applicable rules, dissemination scenarios, and agent processing trajectories in a unified order. Experts can make decisions to adopt, modify, merge, split, reject, or suggest adding labels. Each decision should be recorded with reasons. Adding labels should also include the parent label, label definition, scope of application, counterexamples, and evidence requirements. The same high-risk sample should be reviewed by at least two independent experts. If experts disagree, the sample should be submitted to a third expert or a special meeting for processing. The weighted voting of experts can only be used to summarize opinions and cannot replace the review of expert reasons.

[0050] The expert-weighted adoption rate metric is used to measure the collective attitude of multiple domain experts towards the same candidate identifier. Not all experts are given equal weight. Non-negative weights are assigned based on the relevance of experts' research fields and the consistency of their historical annotations. Experts specializing in intangible cultural heritage research have higher weights for intangible cultural heritage resources. Experts who frequently produce contradictory annotations have lower weights. Binary values ​​are defined as follows: 1 = expert t accepts candidate j from the machine output; 0 = expert t does not accept. Experts perform complex operations such as modifying labels, merging labels, and splitting labels. These operations should not be forcibly filled into 0-1 binary variables but must be stored in independent structured fields. This only indicates whether to directly adopt the machine's output candidates as is; any modifications will be recorded separately to avoid information loss. Weighted adoption rate, with a value of 0-1, can be used to statistically analyze feedback samples and provide supervision signals for subsequent model iterations. For example, if multiple highly weighted experts do not adopt the sample, the sample will be preferentially added to the training set as a negative sample.

[0051] The domain expert group can only confirm the results of a single labeling or submit change suggestions. The label governance committee decides whether to issue a formal change after verifying the experts' reasons, affected samples, regression tests, version differences, and rollback plans. Public opinions can serve as clues to discover disputes, but do not directly determine the official label version.

[0052] Automatic processing will stop and manual review will begin if any of the following conditions are met: There is no evidence unit that can pinpoint the specific location of the original resource to support the judgment; The factual conflicts or contradictions in the sources of knowledge among different evidentiary units exceed a predetermined threshold; Applicable rules that cannot be automatically passed; New identifiers or new identifier levels have appeared. The preset maximum number of retries has been reached and automatic repair has failed.

[0053] Specifically, this application mandates human intervention in certain scenarios to mitigate the risk of omissions in model quantification metrics. If any one criterion is met, regardless of the model's score or uncertainty value, automatic approval is not permitted. All evidence cannot be mapped to the specific location of the original material, and all judgments lack physical / textual basis, preventing the machine from outputting labels. Multimodal evidence and multiple historical sources often contradict each other significantly. Traditional art faces numerous academic controversies, and the machine cannot automatically select a conclusion; expert evaluation is required. Domain blacklists and rules for highly sensitive topics can be configured, mandating human intervention upon detection. The controlled label library expands. New ontology entries cannot be automatically generated by the algorithm; expert review of label definitions is necessary. If upstream retrieval and evidence extraction fail to produce qualified candidates after multiple iterations, the automated process reaches its limit and is handled manually. This prevents the algorithm from outputting unreliable labels in scenarios with insufficient evidence, academic controversies, or changes in the label library, ensuring the humanistic rigor of traditional art labeling.

[0054] It is worth noting that any formal changes to new tags, tag levels, cultural knowledge, applicable rules, or prompt templates should be approved by the Tag Governance Committee. When approving changes, the committee should at least check whether the reasons for the changes are sufficient; whether there is evidence of positioning and knowledge sources; whether it affects existing tags; whether regression testing has been completed; whether the scope of application is clearly defined; and whether there are version rollback materials. After approval, a new version should be released, and the original version, changes, approver, release time, and rollback path should all be retained. If the approval is not approved, the original official version should continue to be used.

[0055] Human decisions confirmed by experts are first entered into the candidate feedback set and are not directly written into the formal labeling system. When updating the prompt template, the original template, the reason for modification, the affected tasks, and the regression test results should be recorded; when updating the retrieval index, the source and expiration time of the knowledge entries should be retained; when updating training data, the sample source, expert decision, and data partitioning information should be retained.

[0056] The process of updating data and formal objects is represented as follows:

[0057] ; In the formula, k represents the version number; This represents the k-th version of training or evaluation data; This represents the newly added and confirmed feedback samples in version k; This indicates the next version of data after the merger; This represents the set of official objects for version k, including tags, cultural knowledge, applicable rules, and prompt templates; Indicates a proposed change; Publish(·) indicates a release operation that will only be performed if the committee approves, the regression test passes, and the rollback materials are complete; This indicates the new version after its release.

[0058] The two update paths should be executed separately. The inclusion of a feedback sample in the candidate dataset does not necessarily mean that the corresponding label, knowledge, rule, or prompt template has become the official version. If any update fails, the current version should still be used. The system records the reasons for update failures and outputs resource information, candidate labels, official labels, location evidence, cultural knowledge sources, sub-judgment results, uncertainties, processing actions, propagation scenarios, expert decisions, agent trajectories, and version information. Each output should simultaneously save the data structure version and method version. In actual implementation, JSON Schema is used to constrain field types, enumeration ranges, required fields, and conditional relationships.

[0059] Example This embodiment uses a traditional Chinese opera video containing lyrics and performance actions as input resource, and the predetermined dissemination scenario is public screening, to illustrate the implementation process of the present invention in terms of multimodal resource processing and independent review of value-based tags.

[0060] The modality parsing agent identifies the resource as a video modality and simultaneously initiates three processing paths: video encoding, audio encoding, and text encoding. Video encoding: A spatiotemporal encoder was used to extract shot, keyframe, and character motion features. The video was divided into 157 shots, and 872 keyframes were extracted. Character identity (such as "painted face character" or "dan role"), action posture, and stage choreography information were extracted through character detection and skeletal keypoint recognition. Each element was accompanied by shot number and frame number position.

[0061] Audio encoding: Acoustic features were extracted using an audio Transformer encoder, and the lyrics were transcribed using Automatic Speech Recognition (ASR). The transcribed text is "I was originally a carefree person from Wolonggang...", with start and end timestamps [00:32:15, 00:45:22] and vocal channel information.

[0062] Text encoding: The transcribed text of the lyrics is encoded using a contextual language encoder. Entity recognition and relation extraction are used to identify “Wolonggang” (place name), “Sandanren” (character’s self-reference), “Zhuge Liang” (character’s related person), and the reference relationship between the character identity “Zhuge Liang” and the lines.

[0063] like Figure 3 As shown, the evidence ledger agent records evidence units of three modalities separately and projects them uniformly into a 512-dimensional feature space. Evidence units include: keyframe image evidence (character wearing a Bagua robe and holding a feather fan), audio lyrics evidence (acoustic features of the character's singing and transcribed text), and textual evidence of character relationships (character identity and narrative relationship). The confidence level of the lyrics transcription result is 0.89, and the confidence level of the video subtitle OCR result is 0.93. The two differ in a few words in the sentence "I was originally a carefree person from Wolonggang." The evidence ledger retains both pieces of evidence and records the conflict.

[0064] The candidate label agent integrates three types of evidence: video, audio, and text, and retrieves candidate labels such as "recluse", "recluse", "image of Zhuge Liang", and "character expression".

[0065] An independent value-reviewing agent independently reviewed the candidate tag "Seclusion". The agent checked the character speaking in the lyrics "I am a carefree person from Wolonggang" and confirmed that it was "Zhuge Liang" played by a "painted face" actor, rather than a narrator or the author speaking directly. Checking the plot, the agent found that the song was Zhuge Liang's self-monologue in "The Empty City Stratagem", which is a self-positioning of the character. Checking the narrative relationship, the statement is a character expression (the character Zhuge Liang's introduction of himself in the play), rather than a direct advocacy by the provider or disseminator of the work (i.e., TV series / opera works do not directly promote "everyone should live in seclusion").

[0066] The independent value-reviewing agents output the following respectively: Correct prediction of candidate label j =0.95, "seclusion" accurately describes Zhuge Liang's self-positioning in the drama; Cultural facts support predictions =0.92, in the original novel of Romance of the Three Kingdoms, Zhuge Liang did indeed live in seclusion at Wolonggang, which is consistent with cultural facts; Communication Scenarios and Value Adaptation Prediction =0.98. In public screening settings, labeling Zhuge Liang as "recluse" helps the audience understand the character's background and aligns with the communication objectives. All three sub-values ​​were high and exceeded the automatic approval threshold for public screening scenarios (the scenario adaptation threshold was set to 0.85). The candidate tag "Hidden" was diverted to automatic approval.

[0067] After the candidate label "Image of Zhuge Liang" was generated by the candidate label agent, during the independent review process, it was found that the confidence level of the word "Wolonggang" in the audio transcription result was only 0.78, while the OCR recognition result of the video subtitles was "Wolonggang" (confidence level 0.95). The two are inconsistent in a few words (such as the difference between "gang" and "gang"). This conflict has been recorded in the evidence ledger. The uncertainty diversion agent will determine the cross-modal evidence conflict level. As an independent quantification term, it is included in the degree of uncertainty. =0.45, combining the completeness of evidence and the dispersion of routing probability, we obtain the degree of uncertainty. =0.28, which is higher than the automatic pass threshold of 0.25 but lower than the manual review threshold of 0.40, so it is diverted to the return correction.

[0068] The candidate label agent re-searched based on the diversion results, corrected it to the candidate label "Wolonggang·Zhuge Liang", and resubmitted it. There were no conflicts in the second review. The three sub-values ​​were 0.91, 0.89, and 0.93, respectively. The evidence completeness was 0.94, and the uncertainty was 0.11. The rule was not hit, and it was automatically approved.

[0069] The final structured output includes resource information, the official tags "Seclusion" and "Zhuge Liang of Wolonggang" and their location evidence (keyframe number, audio timestamp), sub-scores, uncertainty, dissemination scenario (public screening), review status (released), and agent trajectory. Specifically, the output clearly distinguishes between "character expression" and "work advocacy." The candidate tag "Seclusion" is marked as "character expression," and a review comment explains that "this statement pertains to Zhuge Liang's self-positioning as a character, rather than a lifestyle advocated by the work's disseminator for viewers to emulate." This distinction is generated and independently recorded by an independent value review agent during the review process.

[0070] The above description is merely a few embodiments of this application and is not intended to limit this application in any way. Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any changes or modifications made by those skilled in the art without departing from the scope of the technical solution of this application using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A smart hybrid annotation method for aligning the values ​​of traditional art resources, characterized in that, include: Traditional art resources are acquired and modal features of each modality are extracted. The modal features are bound to the spatial or temporal location information of the original resources to generate multimodal evidence units with a unified data structure that can be traced back to the original location. An evidence set is established with the traditional art resources as the unit. When there is a factual conflict between different evidence units, all conflicting evidence is retained in the evidence set and the conflict association is recorded. Based on the evidence unit, a matching candidate identifier is retrieved from a preset controlled identifier system. The retrieval process only allows the selection of identifiers that have a direct reference relationship with the evidence unit. When the evidence is insufficient to support any identifier, an empty identifier is output. The generation conclusion of the candidate identifier is isolated from independent verification. An independent verification mechanism independently verifies the identification accuracy, factual support and scenario adaptability of the candidate identifier based on the candidate identifier, the evidence unit, the external knowledge base and the preset application scenario, and outputs the verification results for each item respectively. Based on the verification results, the completeness of the evidence unit referenced by the candidate identifier, and the degree of conflict between different modal evidence units, the uncertainty of the candidate identifier is comprehensively quantified, and the candidate identifier is divided into automatic approval, manual review, or return for correction according to the uncertainty. The system receives correction decisions made by the manual review process for the diverted review identifiers. After the correction decisions are approved and regression testing is completed, they are included in the feedback sample set for subsequent iterative updates of retrieval or verification strategies. Output the candidate identifier, the evidence unit it references, and the verification results; Each execution step operates within the preset permission boundaries, and each execution step does not directly release an official version or delete original evidence.

2. The intelligent hybrid annotation method for aligning the values ​​of traditional art resources according to claim 1, characterized in that, The modal features extracted from each modal content include: Each modal encoder retains the spatial, character, or temporal structure of its own mode and projects it to a unified feature space through a modal adapter: ； ； Where m represents the resource mode; i represents the i-th region, segment, or time window in that mode; Indicates the original input unit; A dedicated encoder representing mode m; This represents the native features output by the encoder; Indicates a modal adapter; The 'd' represents the adapted features; 'd' represents the unified feature dimension. Let represent the d-dimensional real space.

3. The intelligent hybrid annotation method for aligning the values ​​of traditional art resources according to claim 1, characterized in that, The evidence unit is represented as follows: ； in, This represents the evidence marker for the i-th modality m; the semicolon in square brackets indicates vector concatenation; Represents a region, character range, or time window; Indicates the resolution confidence level; This represents traceability information consisting of resource source, model version, and extraction time; This represents a function that encodes location information, confidence level, and source fields into a vector; D represents the uniform dimension of evidence labeling.

4. The intelligent hybrid annotation method for aligning the values ​​of traditional art resources according to claim 1, characterized in that, The step of retrieving matching candidate identifiers from a preset controlled identifier system based on evidence includes: All evidence units are organized into an evidence sequence according to resource ownership and temporal correlation. The evidence sequence is then fused using a masked attention mechanism, with missing modalities skipped by a mask. ； Where E represents the evidence sequence consisting of all evidence markers; N represents the number of evidence markers; The parameter is The evidence sequence fusion model; H represents the fused hidden state sequence; represents the i-th hidden state; a represents the trainable attention query vector; Let z represent the normalized weight of the i-th piece of evidence; z represents the resource-level fusion representation; i and All are evidence indexes; (·) represents an exponential function; Indicates transpose; The resource-level fused representation is projected onto the identifier vector space, and the retrieval score is calculated by combining the semantic relationships in the identifier ontology: ； ； Where j represents the standard label index; Let represent the retrieval score of the j-th standard tag; P represents the matrix that projects the fused representation onto the tag vector space; Indicates the first A standard label vector; cos(·) represents the cosine similarity; The score represents the synonym, hyponym, exclusivity, or branching relationship within the tag body; Indicates the weight of the ontology relation item; Indicates the number of candidate tags; This represents the candidate set consisting of the K highest-scoring labels; (·) indicates that the top K items are selected based on their scores.

5. The intelligent hybrid annotation method for aligning the values ​​of traditional art resources according to claim 1, characterized in that, The independent verification of the candidate identifier's accuracy, factual support, and scenario adaptability includes: performing joint prediction on the candidate identifier using a multi-task review model. ； ； in, (·) indicates that the parameter is A multi-task review model; This represents the prediction of the correctness of candidate label j; This indicates that cultural facts support the prediction; This indicates a prediction of the fit between the dissemination scenario and the value. This represents the probability distribution of tag hierarchy categories; This indicates the probability distribution of the processed routes; This represents the probability distribution of evidence citations. This represents the review representation of candidate label j; The parameter is The candidate label agent encoding function; This represents a cultural knowledge fragment that is related to candidate label j and retains its source. This indicates the applicable value rule activated by the current task; s represents the propagation scenario encoding; The multi-task review model is trained using a joint training loss: ； in, This indicates the total losses from the joint training exercise; , and represents the binary cross-entropy loss for label correctness, cultural fact support, and dissemination scenario adaptation tasks, respectively; The cross-entropy loss represents the classification by label hierarchy; This represents the cross-entropy loss used for route classification. Indicates the loss of evidence cited for supervision; each Represents the non-negative weight of the corresponding task, and satisfies .

6. The intelligent hybrid annotation method for aligning the values ​​of traditional art resources according to claim 5, characterized in that, The completeness of the evidence unit is: ； Where j represents the candidate label index; Indicates the completeness of evidence for candidate label j; Indicates whether the i-th piece of evidence is referenced by candidate label j; The calibrated confidence level of the i-th piece of evidence is represented by ; i represents the evidence index; the max(·) function is used to avoid a zero denominator when there is no evidence.

7. The intelligent hybrid annotation method for aligning the values ​​of traditional art resources according to claim 6, characterized in that, The uncertainty of the comprehensive quantization candidate identifier includes: The statistical dispersion of the routing probability is estimated through multiple random inactivation inferences or model ensembles, and cross-modal evidence conflicts and conflicts between individual verification results are included as independent quantification terms in the degree of uncertainty: ； in, Indicates the first The second random inactivation reasoning or the first The model provides a routing probability vector for candidate label j; L represents the number of inferences or the number of models. Indicates the average route probability; 'b' indicates uncertainty regarding candidate labels; 'b' indicates the route category during automatic approval, expert review, or rejection of revisions. Indicates the weight of the probability variance term; Represents the L2 norm; Indicates the degree of conflict between cross-modal evidence or the conclusions of sub-examinations; Indicates the weight of the conflicting terms.

8. The intelligent hybrid annotation method for aligning the values ​​of traditional art resources according to claim 7, characterized in that, The process of categorizing candidate identifiers into automatic approval, manual review, or rejection for correction based on the degree of uncertainty includes: ； in, This indicates the action to take for candidate label j; AUTO means automatic approval; REVIEW means submit to experts; REVISE means reject the current candidate and return for revision. This represents the minimum value of the three prediction results in the joint prediction: label accuracy, cultural fact support, and adaptation to the communication scenario. , and These represent the sub-item quality threshold, evidence completeness threshold, and uncertainty threshold, respectively; This indicates a rule that cannot be automatically passed. This indicates a miss; Indicates that the manual review criteria have been met; 1 indicates that at least one manual review condition has been met.

9. The intelligent hybrid annotation method for aligning the values ​​of traditional art resources according to claim 1, characterized in that, The expert-weighted adoption rate of the candidate identifier is: ； in, The expert-weighted adoption rate of candidate label j is represented by t; the expert index is represented by t. This indicates a non-negative weight predetermined based on the expert's professional field and historical consistency, and at least one of these weights is greater than zero. These represent whether expert t rejects or accepts candidate label j, respectively; expert decisions regarding modification, merging, or splitting should be recorded using separate structured fields, not compressed into a single file. The binary value of .

10. The intelligent hybrid annotation method for aligning the values ​​of traditional art resources according to claim 1, characterized in that, Automatic processing will stop and manual review will begin if any of the following conditions are met: There is no evidence unit that can pinpoint the specific location of the original resource to support the judgment; The factual conflicts or contradictions in the sources of knowledge among different evidentiary units exceed a predetermined threshold; Applicable rules that cannot be automatically passed; New identifiers or new identifier levels have appeared. The preset maximum number of retries has been reached and automatic repair has failed.

Images