LU601246B1 - Methods and systems for interpreting artificial intelligence model algorithms - Google Patents

Abstract

The present invention provides a method and system for interpreting artificial intelligence model algorithms and relates to the field of interpretable artificial intelligence technology. The method and system for interpreting the artificial intelligence model algorithm comprises: S1: modeling intermodal feature associations on the input data based on multimodal input data using an algorithm for matching crossmodal attention; quantifying the contribution of each modal feature to the model output by means of a gradient class activation map with a locally interpretable, model-independent interpretation; and generating results of the analysis of the meaning of the multimodal features. By constructing the correlation matrix between local image regions and text embeddings, combining gradient-like activation maps with the calculation of the contribution of text features, quantifying the crossmodal interaction weights to achieve accurate positioning of the multimodal feature meaning analysis, screening the subset of features causally related to the model output based on the causal detection algorithm, and using a counterfactual sample generator to perturb the non-causal features in order to validate the robustness of the features and exclude pseudocorrelation perturbations.

Description

Method and system for interpreting model algorithms of artificial intelligence LU601246

intelligence

Technical area

The present invention relates to the technical field of interpretable artificial intelligence, in particular to the method and system for interpreting

Artificial intelligence model algorithms.

Technology in the background

Interpretive AI is an important branch of AI whose main goal is to...

To make the decision logic of complex models transparent and understandable for humans. In this area, the internal working mechanisms of models are explored through

Algorithm design, visualization tools, and interactive interpretation systems were disclosed, and the credibility and compliance problems of black-box models in scenarios with high reliability requirements, such as medical diagnosis, financial risk management, and autonomous driving, were solved. The interpretation method and system for AI model algorithms refers to the solution that dynamically reveals the model's decision logic and ensures its consistency through multimodal techniques.

Feature correlation analysis, causal modification, semanticized

Inference chain extraction and hierarchical interpretation generation are monitored. The goal is to provide transparent decision support for medical image analysis, financial risk assessment, and other scenarios, helping users to...

to understand key features and lines of reasoning on which the models are based, to verify the fairness and robustness of the algorithms, and to ensure the long-term reliability of the interpretation results during the model iteration process.

Existing methods focus on calculating the contribution of unimodal

Features in multimodal feature analysis, and a quantitative modeling of the weights of intermodal interaction is lacking, leading to insufficient accuracy in the

Identifying the relevance of key features in scenarios involving data fusion from multiple sources. Feature selection does not strictly distinguish between statistical and statistical criteria.

Correlation and causality are distinguished, thereby preserving pseudo-correlated features and impairing the credibility of the decision logic. The generation of the interpretation relies on the direct mapping of the hidden layer of the model and is not combined with the mapping of domain knowledge, resulting in a low degree of

Semanticization of the interpretive results leads to difficulties in meeting the requirements for argumentation and validation in professional scenarios. The uniformity

The output granularity of the explanatory content cannot be adapted to the cognitive differences of

adapting to users with different levels of expertise reduces the practical application value of the explanatory information.

Content of the invention

In view of the shortcomings of the state of the art, the present

The invention provides a method and a system for interpreting model algorithms of artificial intelligence, which solves the problem that the prior art is limited to the

The calculation of the unimodal feature contribution in the multimodal feature analysis is concentrated, and the quantitative modeling of the crossmodal interaction weights is lacking, leading to insufficient accuracy in identifying the relevance of key features in the scenario of data fusion from multiple sources.

To achieve the above-mentioned purpose, the present invention is achieved by the following technical solution: a method for interpreting artificial intelligence model algorithms, comprising the following steps:

S1: Based on the multimodal input data, the correlation matrix between the local regions of the image and the text word embedding is constructed, which

Key regions of the image are located, the contribution of the text features is calculated, the multimodal feature importance is weighted, and the results of the multimodal analysis are presented.

Feature importance analyses are created;

S2: Based on the results of the multimodal feature weight analysis, the subset of features that have a causal relationship to the model output is filtered out; the non-causal features are skewed to generate comparison samples.

The robustness of the features is checked, the pseudo-relevant disturbance elements are filtered out, and the subset of causal correction features is generated;

S3: Based on the subset of causal correction features, the representation of the hidden layer of the model is aligned with the embedding space of the domain knowledge graph, which represents the association paths between the key neurons in the

The model's decision process is activated, and the node of the knowledge graph is extracted and the semanticized decision path mapping table is created;

S4: Train based on the semanticized decision path assignment table

They adapt the competency assessment model, adjust the output granularity according to the primary, intermediate, and expert levels, and generate adaptive hierarchical

Interpretive content;

SS: Based on the adaptive hierarchical interpretation content, the

Calculate JS dispersion of the same sample before and after model literation, which new

Identify the subset of data that causes the difference to exceed the threshold, correlate its causal relationship with the parameter update, and the

Prepare a report on monitoring the consistency of interpretations.

Preferably, the generation of the results of the multimodal process includes

Feature weight analysis based on the results of the multimodal

Feature weight analysis in S1 involves the following steps:

S101: Based on the multimodal input data, constructing a

Association matrix of local image regions and text word embeddings, quantification of intermodal interaction weights and generation of a crossmodal association matrix;

S102: based on the cross-modal correlation matrix, localizing the

Image key region, merging the text feature contribution calculation, weighting the output multimodal feature meaning score, and generating the multimodal

Feature-meaning-analysis-result field.

Preferably, generating a subset of causal correction features based on S2 comprises the following steps:

S201: Based on the results of the multimodal feature meaning analysis, a subset of causal features is screened using a conditional independence test and causal graph structure learning, and a subset of causal

Feature candidates are generated;

S202: Based on the subgroup of causal feature candidates,

Comparative samples were generated to verify the robustness of the features, pseudo-relevant

Interference terms were filtered out and a subset of causal correction features was generated. LU601246

Preferably, the generation of a semanticized

Decision path mapping table based on the semanticized

Decision path mapping table in S3 the following steps:

S301: Based on a subset of causal correction features, aligning the

Representation of the hidden layer of the model with the embedding space of the

Knowledge graph entity, minimizing semantic distance and generating the

Figure table of the hidden layer and the knowledge entity;

S302: Extracting the data based on the comparison of the figure table

Association paths between key neurons and knowledge graph nodes, generating the semanticized decision path mapping table.

Preferably, the generation of the adaptive hierarchical interpretation content based on the content in S4 comprises the following steps:

S401: based on the semanticized decision path mapping table, training the professionalism scoring model through the user's historical interaction data to generate the user professionalism portrait;

S402: based on the user professionalism profile, customizing the

Output granularity according to primary, intermediate and expert levels to generate the adaptive hierarchical interpretation content.

Preferably, generating a report for monitoring the

Explanatory consistency based on S5 involves the following steps:

S501: Based on the adaptive hierarchical interpretation content, calculating the difference in the interpretation distribution of the same sample before and after the

Model update and generation of an interpretation discrepancy scale;

S502: Based on the explanation discrepancy scoring table, locating the added data subset, associating its causal relationship with the

Parameter update and generation of the explanation consistency monitoring report.

The system for interpreting artificial intelligence model algorithms, wherein the interpretation system for the artificial intelligence model algorithm comprises the following

The module includes: a multimodal causal analysis module, a semantic hierarchical

Interpretation module and a module for monitoring the consistency of the interpretation;

The multimodal causal analysis module constructs an association matrix of images and text, quantifies feature contributions, filters a subset of causal features, filters out pseudo-relevant noise, and generates a subset of causal

Correction features;

The multimodal causal analysis module, which is a submodule for the crossmodal

Includes matching and a sub-module for causal correction;

The semantic hierarchical interpretation module, which represents the hidden layer of the

The model compares with the entity of the knowledge graph, extracts the decision inference chain, identifies the user's level of expertise, outputs the interpretation by level, and generates adaptive hierarchical interpretation content;

The semantic hierarchical interpretation module, which is a semantic

It contains a mapping sub-module and a hierarchical generation sub-module;

The module for monitoring interpretation consistency, in order to detect differences in the

To quantify the distribution of interpretations and to locate the new data subsets,

to assign parameters to update causality and generate the monitoring report 601246 for interpretation consistency; the module for monitoring the

Interpretation consistency includes the submodule for the discrepancy metric and the

Submodule for drift traceability.

Preferably, a submodule for cross-modal matching constructs a

Association matrix from local regions of the image and text/word embeddings, generating a cross-modal association matrix;

The submodule for causal corrections locates the key regions of the image, calculates the contribution of the text features, filters the subset of causal features, filters the pseudo-relevant noise, and generates the subset of causal correction features.

Preferably the semantic mapping submodule that matches the representation of the hidden layer of the model with the embedding space of the knowledge graph, generates the hidden layer and the knowledge entity mapping table;

The submodule for hierarchical generation extracts semanticized inference chains, identifies the user's expertise, and dynamically creates hierarchical

Explanatory content is generated, and adaptive hierarchical explanatory content is generated.

Preferably a submodule for the discrepancy metric that accounts for the difference in the

The distribution of explanations before and after the model update was calculated, and a

Evaluation sheet created for the explanation discrepancy;

The drift traceability submodule, which includes the added subset of

Data localized, parameters correlated to update causality, and a report on

Monitoring of the consistency of the declarations was created.

The present invention presents the method and system for the interpretation of

Artificial intelligence model algorithms are ready.

It has the following beneficial effects:

The present invention quantifies cross-modal interaction weights by the

Construction of an association matrix of local image regions and text word embeddings that

Combination of gradient class activation maps with calculation of the contribution of

Text features and the realization of precise positioning for the analysis of

Significance of multimodal features. Based on the causal detection algorithm, we examine the subset of features that have a causal relationship to the model output and use the counterfactual sampling generator to identify the non-causal features.

to disrupt features in order to test the robustness of the features and the

To eliminate pseudocorrelation errors, thereby improving the reliability of feature selection and the credibility of the decision logic. Adapting the representation of the hidden layer of the model to the embedding space of the domain knowledge graph.

Extraction of association pathways between key neurons and knowledge nodes and

Construction of the semanticized inference chain to improve comprehensibility and the

To improve the adaptability of the interpretation results. Through training the

In the model for assessing user expertise, the granularity of the explanatory content is dynamically adjusted to meet the needs of users with different professional backgrounds.

To do justice to the background and improve the practical usefulness of the explanatory information, we calculate the distribution difference of the interpretation results before and after model literatureation, determine the causal relationship between the new data subset and the parameter update, and continuously monitor the consistency of the interpretation.

to ensure the stability of the interpretation logic during the long-term iteration process4 V601 246. The processing logic systematically solves the problems of distortion of the

Feature interpretation, pseudocorrelation interference, semantic separation and dynamic adaptability in multimodal scenarios through cross-modal fusion, causal modification, knowledge graph embedding, hierarchical generation and incremental monitoring.

Description of the attached drawings

Figure 1 shows a block diagram of the system of the present invention;

Figure 2 shows a schematic diagram of the main steps of the present

Invention;

Figure 3 shows a schematic diagram of the S1 refinement of the present invention;

Figure 4 shows a schematic diagram of the S2 refining process of the present invention;

Figure 5 shows a schematic diagram of the S3 refinement of the present invention;

Figure 6 shows a schematic diagram of the S4 refinement of the present

Invention;

Figure 7 shows a schematic diagram of the S5 refinement of the present

Invention.

Detailed description

The technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the accompanying drawings, and it is obvious that the described embodiments represent only a part of the embodiments of the present invention.

This does not represent the invention, but rather all its embodiments. Based on the

In the present invention, all other embodiments that can be achieved by a person skilled in the art without creative work fall within the scope of protection of the present invention.

embodiment

As shown in Figures 1-7, one embodiment of the present invention represents

Interpretation procedure of an algorithm for an artificial intelligence model, comprising the following steps:

S1: Based on the multimodal input data, an association matrix of local image regions and text word embeddings is constructed, the image key regions are localized, the text feature contribution level is calculated, the = multimodal

Feature importance points are weighted and the multimodal

Feature importance analysis results are generated;

Based on the multimodal input data, the mechanism of cross-modal attention is used to construct the association matrix between local

to construct image regions and text word embeddings, to locate the key regions of the image through gradient-weighted class activation assignment, the

to calculate text feature contribution based on the Shapley score and to weight the output of the multimodal feature importance points to evaluate the results of the multimodal

To generate a feature importance analysis.

S2: Based on the results of the multimodal feature meaning analysis, the subset of features that have a causal relationship to the model output is filtered out; the non-causal features are perturbed to generate comparison samples.

The robustness of the features is checked, the pseudo-relevant interference terms are filtered out, and a subset of the causal correction features is generated;

Based on the results of SI, use the causal

Structural learning algorithm to filter the causal feature subset, generate counterfactual samples through a generative adverse network to verify feature robustness, filter the pseudorelevant interference terms, and generate the causal

Correction feature subset.

S3: Based on the causally modified feature subset, the

Representation of the hidden layer of the model with the embedding space of the

Domain knowledge graphs were compared, identifying association paths between the data in the

The decision-making process of the model activated key neurons and the nodes of the

Knowledge graphs extracted and the semanticized decision path mapping table created;

Based on the subset S2, the algorithm for embedding the

Knowledge graphs are used to represent the hidden layer of the model with the

to match the entity of the knowledge graph, and the association paths between the

Key neurons and the knowledge node are extracted by pathfinding of the neural network of the graph to generate the semanticized table of the decision path mapping.

S4: Train based on the semanticized decision path assignment table

They use the professionalism assessment model, adjusting the output granularity accordingly.

primary, intermediate, and expert levels are accessed and generate adaptive hierarchical

Interpretive content;

Based on the S3 assignment table, the

Train the Support Vector Machine Expertise Assessment Model with which

Combine a multigranularity module for natural language generation to dynamically adjust the granularity of the interpreted content and the adaptive hierarchical

Generate interpretive content.

SS: Based on the adaptive hierarchical interpretation content, the

JS dispersion of the same sample before and after model literation is calculated, the new subset of data that causes the difference to exceed the threshold is located, its causal relationship is correlated with the parameter update, and the report for monitoring interpretation consistency is generated.

Based on the content of S4, the incremental JS scattering algorithm is used to determine the difference in the interpretation distribution before and after the

To calculate model literation, find the subset of data that is the source of the

The difference, through the analysis of the feature's meaning, lies in relating it to the causal relationship to

to correlate parameter updates and the report for monitoring the

to create consistency in interpretation.

The generation of results from the multimodal feature meaning analysis on the

The basis of S1 comprises the following steps:

S101: Based on the multimodal input data, construct an association matrix of local image regions and text embeddings, quantify the intermodal interaction weights, and generate a crossmodal association matrix;

Constructing an image-text association matrix using an intermodal

Attention-direction algorithm to quantify interaction weights.

S102: Based on the cross-modal correlation matrix, the key region of the V601246 is identified.

Image localized, calculation of text feature contribution merged, the multimodal

Feature importance weighted and the result field of the multimodal

Feature importance analysis generated.

Based on Grad-CAM, locate the key regions of the image, fuse them.

Use LIME to calculate the text contribution and enter the multimodal

Importance assessment.

Generating a subset of causal modification features based on S2 involves the following steps:

S201: Generating a subset of causal feature candidates by screening the

Subset of causal features tested using a conditional independence test with causal

Graph structure learning based on the results of multimodal

Feature meaning analysis;

Screening the subset of causal feature candidates using a conditional

Independence test using the FCI algorithm.

S202: Based on the subset of causal candidate features,

Comparative samples were generated to verify the robustness of the features, pseudo-relevant

To filter out interference concepts and generate a subset of causal correction features.

Non-causal ~GAN perturbation features are used to counterfactual

to generate samples and filter the interference terms.

The generation of a semanticized decision path mapping table on the

The foundation of S3 comprises the following steps:

S301: Based on the subset of causal correction features, the

Representation of the hidden layer of the model in the embedding space of the

Knowledge graph entity adapted, semantic distance minimized, and mapping table created between hidden layer and knowledge entity;

Aligning the hidden layer representations with the knowledge mapping entities using the TransE algorithm, minimizing the semantic distance.

S302: Based on the alignment mapping table, the association paths between key neurons and knowledge graph nodes are extracted and the semanticized

Decision path mapping table generated.

Extracting critical associated paths based on GNN pathfinding and

Generating semantic mapping tables.

The generation of adaptive hierarchical interpretation content based on S4 includes the following steps:

S401: Based on the semanticized decision path mapping table, training the professionalism assessment model with user history interaction data to create a

To create a user professionalism profile;

Training an SVM expertise assessment model using

User history data is used to create a user portrait.

S402: Based on the user professionalism profile, the

Output granularity is adjusted according to the primary, middle, and expert levels to generate adaptive hierarchical interpretation content.

Dynamic generation of multigranular explanatory content based on the

NLG module. LU601246

Generating a report to monitor the consistency of interpretation on the

The foundation of S5 comprises the following steps:

S501: Based on the adaptive hierarchical interpretation content, calculation of the difference in the interpretation distribution of the same sample before and after the

Model update and generation of an interpretation discrepancy score table;

Calculating the JS dispersion before and after model literatureation, generating a

Explanatory discrepancy score table.

S502: Based on the table for explanatory discrepancies, the new

Data subset located, its causal relationship correlated with parameter updates, and a report created to monitor explanatory consistency.

Locate the source data of the discrepancy through backtracking analysis and correlate the causal chain of parameter updates.

The system for interpreting artificial intelligence model algorithms, wherein the interpretation system for the artificial intelligence model algorithm comprises the following

The module includes: a multimodal causal analysis module, a semantic hierarchical

Interpretation module and a module for monitoring the consistency of the interpretation;

The multimodal causal analysis module constructs an association matrix of images and text, quantifies feature contributions, filters a subset of causal features, filters out pseudo-relevant noise, and generates a subset of causal

Correction features;

The multimodal causal analysis module, which is a submodule for the crossmodal

Includes matching and a sub-module for causal correction;

The semantic hierarchical interpretation module, which represents the hidden layer of the

The model compares with the entity of the knowledge graph, extracts the decision inference chain, identifies the user's level of expertise, outputs the interpretation by level, and generates adaptive hierarchical interpretation content;

The semantic hierarchical interpretation module comprises a semantic

Mapping submodule and a hierarchical generation submodule;

The module for monitoring interpretation consistency, in order to detect differences in the

to quantify the distribution of interpretations, to locate the added data subsets, to correlate the parameters to update the causality, and to prepare the report for

To create a system for monitoring the consistency of interpretations;

The module for monitoring interpretation consistency includes a submodule for the

Discrepancy metric and a submodule for drift traceability.

Preferably, the submodule for cross-modal matching constructs a

Association matrix from local image regions and text word embeddings, generating a cross-modal association matrix;

A cross-modal attention mechanism is used to...

To construct an association matrix of local image regions and text word embeddings, to quantify inter-modal interaction weights, and to generate a cross-modal association matrix.

The submodule for causal corrections, localization of image key regions,

Calculation of the contribution of text features, filtering of the subset of causal features,

Filtering of pseudocorrelation noise, generation of a subset of causal

Correction features. LU601246

Based on the cross-modal correlation matrix, localization of the key region of the image by gradient-weighted class activation assignment, fusion of locally interpretable, model-irrelevant interpretations to calculate the

Text feature contribution, combined with the causal recognition algorithm for filtering the causal feature subset, using a generative adverse network for

Disturbance of the non-causal features to generate counterfactual samples, filtering of the pseudo-relevant noise and generation of the causal correction feature subset.

The semantic mapping submodule, to represent the hidden layer of the

to align the model with the space of the knowledge graph embedding in order to uncover the hidden

to generate the layer and the knowledge entity table;

Application of the knowledge graph embedding algorithm to align the representation of the model's hidden layer with the knowledge graph embedding space, minimize the semantic distance, and align the hidden layer and the

To generate a knowledge entity mapping table.

The submodule for hierarchical generation extracts semanticized inference chains, identifies user expertise, and dynamically creates hierarchical

It generates interpretive content and generates adaptive hierarchical interpretive content.

Based on the mapping table, the semanticized inference chain is processed by the

Path search of the graphical neural network extracted and combined with the data from the

Support vector machine classifier identified user know-how portrait combined.

The multi-granular natural language generation module is used for the dynamic generation of interpretation content at primary, intermediate and expert levels, as well as for the generation of adaptive hierarchical interpretation content.

The discrepancy measurement submodule calculates the difference in the distribution of explanations before and after the model update and creates a table for the

Explanatory discrepancy;

Based on the adaptive hierarchical interpretation content, the incremental JS dispersion algorithm calculates the difference in the interpretation distribution of the same sample before and after model literation and generates the

Interpretation discrepancy scale.

The drift traceability submodule locates new subsets of data and correlates them.

Parameters to update causality, and generates reports to monitor the

Consistency of interpretation.

Use the algorithm for analyzing feature meaning (feature)

Importance Backtracking) to find the new data subset where the difference is the

If the threshold is exceeded, combine the method for estimating the causal effect to link the causal relationship with the parameter update and create

See the report on monitoring the consistency of declarations.

Although embodiments of the present invention have been shown and described, the person skilled in the art will recognize that a multitude of changes, modifications, substitutions and variations can be made to these embodiments without altering the fundamental design.

to deviate from the principle and spirit of the present invention, the scope of which is limited by the attached claims and their equivalents.

Claims (10)

Claims LU601246 1. The procedure for interpreting artificial intelligence model algorithms, characterized in that it comprises the following steps: S1: Based on the multimodal input data, constructing an association matrix of local image regions and text word embeddings, locating key image regions, calculating the contribution of text features, weighting the output multimodal feature importance rating, and generating results of the multimodal feature importance analysis; S2: Based on the results of the analysis of the importance of multimodal features, the subset of features that have a causal relationship with the model output is filtered, the non-causal features are perturbed to generate comparison samples, the robustness of the features is checked, the pseudo-relevant perturbation elements are filtered, and the subset of causal correction features is generated; S3: Based on the subset of causal correction features, the representation of the hidden layer of the model is aligned with the embedding space of the domain knowledge graph, the association paths between the key neurons activated in the model's decision-making process and the nodes of the knowledge graph are extracted, and the semanticized decision path mapping table is created; S4: Based on the semanticized decision path mapping table, you train the competency assessment model, adjust the output granularity according to the primary, intermediate, and expert levels, and generate adaptive hierarchical interpretation content; SS: Based on the adaptive hierarchical interpretation content, you calculate the JS dispersion of the same sample before and after model literation, identify the new subset of data that causes the difference to exceed the threshold, correlate its causal relationship with the parameter update, and create the report for monitoring interpretation consistency. 2. The method for interpreting artificial intelligence model algorithms according to claim 1, characterized in that: the generation of a result of the importance analysis of the multimodal features based on the multimodal features in S1 comprises the following steps: S101: Based on the multimodal input data, construct an association matrix of local image regions and text word embeddings, quantify the intermodal interaction weights, and generate a crossmodal association matrix; S102: Based on the cross-modal correlation matrix, the key region of the image is localized, the text feature contribution calculation is fused, the multimodal feature meaning output evaluation is weighted, and the multimodal feature meaning analysis result field is generated. 3. The method for interpreting artificial intelligence model algorithms according to claim 1, characterized in that: generating a subset of causal modification features based on the causal modification features in S2 comprises the following steps: S201: Based on the results of the multimodal feature meaning analysis, screening the subset of causal features by conditional independence test and causal graph structure learning, and generating the subset of causal feature candidates; S202: Generating a subset of causal feature candidates based on the subset of causal features, generating comparison samples to verify feature robustness, filtering pseudo-relevant interference terms, and generating a subset of causal correction features. 4. The method for interpreting artificial intelligence model algorithms according to claim 1, characterized in that: generating a semanticized decision path mapping table based on the semanticized decision path mapping table in S3 comprises the following steps: S301: Based on the subset of causal correction features, aligning the hidden layer representation of the model with the knowledge graph entity embedding space, minimizing the semantic distance, and generating the hidden layer knowledge entity mapping table; S302: Based on the alignment of the mapping table, extracting the association paths between key neurons and knowledge graph nodes, and generating the semantic decision path mapping table. 5. The method for interpreting artificial intelligence model algorithms according to claim 1, characterized in that: the generation of adaptive hierarchical interpretation content based on S4 comprises the following steps: S401: Training the professionalism assessment model based on the semanticized decision path mapping table using the user's historical interaction data to generate the user's professionalism portrait; S402: Adjusting the output granularity according to primary, intermediate, and expert levels based on the user's professionalism portrait to generate the adaptive hierarchical interpretation content. 6. The method for interpreting artificial intelligence model algorithms according to claim 1, characterized in that, based on: generating a report for monitoring the consistency of the interpretation based on S5, comprises the following steps: S501: calculating the difference in the interpretation distribution of the same sample before and after the model update based on the adaptive hierarchical interpretation content and generating an interpretation discrepancy scale; S502: locating a subset of the added data, assigning its causal relationship to the parameter update, and generating an explanation consistency monitoring report based on the explanation discrepancy scoring table. 7. The system for interpreting artificial intelligence model algorithms, characterized in that the interpretation system for the artificial intelligence model algorithm comprises the following modules: a multimodal causal analysis module, a semantic hierarchical interpretation module, and a module for monitoring the consistency of the interpretation; The multimodal causal analysis module constructs an association matrix of images and text, quantifies the feature contribution, filters a subset of causal features, filters pseudocorrelation noise, and generates a subset of causal Correction features; LU601246 The multimodal causal analysis module, including a sub-module for cross-modal matching and a sub-module for causal correction; The semantic hierarchical interpretation module, which matches the hidden layer of the model with the entity of the knowledge graph, extracts the decision inference chain, identifies the user's level of expertise, outputs the interpretation by level, and generates adaptive hierarchical interpretation content; The semantic hierarchical interpretation module, which includes a semantic mapping sub-module and a hierarchical generation sub-module; The interpretation consistency monitoring module to quantify differences in the distribution of interpretations, locate new data subsets, assign parameters to update causality, and generate the interpretation consistency monitoring report; The module for monitoring interpretation consistency includes the sub-module for the discrepancy metric and the sub-module for drift traceability. 8. The system for interpreting model algorithms of artificial intelligence according to claim 7, characterized in that: a submodule for cross-modal matching constructs an association matrix of local regions of the image and text word embeddings and generates a cross-modal association matrix; the submodule for causal correction, which localizes key regions of the image, calculates contributions of text features, filters a subset of causal features, filters pseudocorrelated noise, and generates a subset of causal correction features. 9. The system for interpreting model algorithms of artificial intelligence according to claim 7, characterized in that: a semantic mapping submodule that aligns the hidden layer representation of the model with the knowledge graph entity embedding space, generates a hidden layer, knowledge entity mapping table; the submodule for hierarchical generation that extracts a semanticized inference chain, identifies the user's expertise, dynamically generates a hierarchical interpretation content and generates an adaptive hierarchical interpretation content. 10. The system for interpreting artificial intelligence model algorithms according to claim 7, characterized in that: a submodule for a discrepancy metric that calculates the difference in the distribution of explanations before and after the model update and generates a table for evaluating the explanation discrepancy; the submodule for drift traceability that locates a subset of the added data, correlates parameters to update causal relationships, and generates a report for monitoring explanation consistency.