Ffact verification method based on multi-modal evidence fusion and discourse relation enhancement

Through the fact verification method of multimodal evidence fusion and enhanced discourse relationship, the problem of dissemination of false information in social media is solved, the accuracy and reliability of the judgment of authenticity of statements is improved, and the effective identification and reduction of false information is achieved. .

CN119089381AActive Publication Date: 2024-12-06BEIHANG UNIV

Patent Information

Application Number
CN202411162163.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-23
Publication Date
2024-12-06
Estimated Expiration
2044-08-23

AI Technical Summary

Technical Problem

In the rapid spread of false information in social media, the fact-proof model has poor accuracy and reliability, and the multimodal evidence and discourse relationship are not fully utilized, resulting in insufficient accuracy and reliability when judging the authenticity of the statement. .

Method used

Using a fact verification method based on multimodal evidence fusion and discourse relationship enhancement, the multimodal evidence characteristics and discourse dialogue are integrated through evidence retrieval, constructing discourse relationship diagrams and authenticity judgment steps, combining graph neural networks and attention mechanisms, the multimodal evidence characteristics and discourse Relationships to improve the accuracy of judgments on the authenticity of statements.

Benefits of technology

It improves the accuracy and reliability of the multimodal fact verification model in judging the authenticity of the statement, effectively deals with the problem of dissemination of false information in social media, and reduces the dissemination of false information.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089381A_ABST
    Figure CN119089381A_ABST
Patent Text Reader

Abstract

The invention belongs to the field of computer science and technology, and particularly relates to a fact checking method based on multi-modal evidence fusion and discourse relation enhancement, which comprises the following steps: step 1, evidence retrieval: retrieving related evidences from an evidence library for a to-be-detected statement; 2, constructing an evidence relation graph: extracting a discourse relation from the related evidence by using a natural language processing tool, constructing a discourse relation heterogeneous graph based on the extracted discourse relation, and fusing and extracting the features of the discourse relation heterogeneous graph through a graph neural network to form the evidence features of the discourse relation heterogeneous graph; and step 3, authenticity judgment: adopting a complete related evidence as a unit, fusing the declaration characteristics and the evidence characteristics, and analyzing whether the related evidence supports the declaration to be judged or not. According to the method, the accuracy and the reliability of the multi-modal fact verification model in the aspect of judging the authenticity of the declaration are improved, and the declaration to be detected can be effectively judged.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of computer science and technology, and specifically relates to a fact-checking method based on multimodal evidence fusion and discourse relationship enhancement. Background Art

[0002] With the rapid development of the Internet and the rise of social media, the speed of information dissemination is increasing. Due to the fact that everyone can participate in the dissemination of social media, various information spreads virally in social media at an extremely fast speed. However, due to the imperfection and lack of the review mechanism, social media has become a hotbed for the dissemination of various false information. The widespread dissemination of such false information makes it difficult for people to discover the truth behind the incident and may even seriously affect social order.

[0003] In order to deal with the growing rumors on social media platforms, two tasks, fake news detection and fact verification, have been proposed. The former is usually a binary classification task for the authenticity of the news to be detected, while the latter requires the system to retrieve evidence related to the statement in a reliable knowledge base and judge whether the statement is true, false or insufficient information (NEI) based on the evidence.

[0004] For fact verification tasks, some researchers are working on improving the evidence retrieval process to increase the relevance of the retrieved evidence to the claim. Some researchers are working on improving the fact verification process so that the model can better learn the representation of the claim and evidence, and thus make judgments on the authenticity of the claim.

[0005] Despite this, most of the related research in the past was limited to the text modality, and only simply spliced ​​the features of the statement and the features of the evidence. Although a small number of studies explored the interactive relationship between statements and evidence through the attention mechanism, they ignored the interactive relationship between evidence and evidence; other studies used graph neural networks for fact verification tasks, using the information transmission of graph nodes to study the interaction between evidence and evidence, but did not fully explore the semantic structure information contained in the evidence, resulting in poor accuracy and reliability in judging the authenticity of the statement, and unable to effectively judge the statement to be tested. Summary of the invention

[0006] The purpose of the present invention is to provide a fact-checking method based on multimodal evidence fusion and discourse relationship enhancement, which realizes automatic fact-checking based on multimodal evidence fusion and discourse relationship enhancement, improves the accuracy and reliability of the multimodal fact-checking model in judging the authenticity of statements, and can effectively make judgments on the statements to be tested.

[0007] The technical solution adopted by the present invention is as follows:

[0008] A fact-checking method based on multimodal evidence fusion and discourse relationship enhancement includes the following steps:

[0009] Step 1: Evidence retrieval: Retrieve relevant evidence from the evidence library for the claim to be tested;

[0010] Step 2: Construct an evidence relationship graph: Use natural language processing tools to extract discourse relations from relevant evidence, construct a discourse relationship heterogeneous graph based on the extracted discourse relations, and fuse and extract the features of the discourse relationship heterogeneous graph through a graph neural network to form the evidence features of the discourse relationship heterogeneous graph;

[0011] Step 3: Authenticity judgment: Use complete relevant evidence as a unit, merge the claim features and evidence features, and analyze whether the relevant evidence supports the claim to be judged.

[0012] Furthermore, the relevant evidence in step 1 includes text evidence and image evidence.

[0013] Furthermore, the step 2 comprises the following steps:

[0014] Step 101: construct relevant indexes for text evidence to form a new index library;

[0015] Step 102: training a semantic similarity model and a multimodal fact-checking model for retrieving relevant text evidence;

[0016] Step 103: training a multimodal fact-checking model for retrieving claim-related image evidence;

[0017] Step 104: Use the trained semantic similarity model to extract features of the statement and the evidence index, calculate the cosine similarity, sort them according to the similarity, and select the top 5 most relevant evidence as the retrieved text evidence;

[0018] Step 105: Use the trained multimodal fact-checking model to extract features of the claim and features of the image evidence, calculate their cosine similarity, sort them according to the similarity, and select the top 5 most relevant images as the retrieved image evidence.

[0019] Furthermore, the semantic similarity model and the multimodal fact verification model in step 1 are trained using a contrastive learning method. During the training process, evidence related to the statement is selected as a positive sample, and relevant evidence is randomly selected from evidence unrelated to the statement as negative samples for contrastive learning to participate in the training.

[0020] Furthermore, the step 2 comprises the following steps:

[0021] Step 201: Analyze the discourse relations of text evidence using a text relation extraction tool;

[0022] Step 202: constructing a directed heterogeneous graph as an evidence relationship graph based on the extracted discourse relations to form a discourse relationship heterogeneous graph;

[0023] Step 203: Connect the statement and related image evidence to the discourse relationship heterogeneous graph nodes as node features, use the attention mechanism of the graph neural network to control message transmission, use the attention mechanism to perform weighted summation on the node features, complete the fusion and extraction of the discourse relationship heterogeneous graph node features, and obtain the final discourse relationship heterogeneous graph evidence features.

[0024] Furthermore, the discourse relationship heterogeneous graph includes statement nodes, text evidence nodes and image evidence nodes, as well as five types of edges: selection relationship, parallel relationship, causal relationship, inverse causal relationship and ordinary connection relationship.

[0025] Furthermore, the step 3 comprises the following steps:

[0026] Step 301: Sentence processing is performed on the text evidence, and the sentence-processed text evidence and the image evidence are input into the multimodal fact verification model, and the multimodal fact verification model is used to extract text evidence features, image evidence features and features of the statement to be detected;

[0027] Step 302: Calculate the attention distribution of the claim and evidence through the cross attention mechanism, where the claim feature is used as the query and the evidence feature is used as the key and value for feature fusion;

[0028] Step 303: The fused features are input into the pooling layer and the linear layer to judge the authenticity of the statement, and the categories include true, false and insufficient information.

[0029] The technical effects achieved by the present invention are:

[0030] The fact-checking method based on multimodal evidence fusion and discourse relationship enhancement of the present invention realizes automatic fact-checking based on multimodal evidence fusion and discourse relationship enhancement, solves the deficiency of multimodal automatic fact-checking, improves the accuracy and reliability of the multimodal fact-checking model in judging the authenticity of statements, can effectively make judgments on statements to be tested, effectively addresses the problem of false information dissemination in social media, and is of great significance to reducing the spread of false information. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0032] In order to make the purpose and advantages of the present invention more clearly understood, the present invention is specifically described below in conjunction with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the present invention, and does not strictly limit the scope of protection of the specific claims of the present invention.

[0033] Embodiment 1:

[0034] like Figure 1 As shown, a fact-checking method based on multimodal evidence fusion and discourse relationship enhancement includes the following steps:

[0035] Step 1: Evidence retrieval: Retrieve relevant evidence from the evidence database for the statement to be tested; use the large semantic similarity model to extract important information from the evidence text and eliminate the interference of redundant content on the retrieval.

[0036] Among them, effectively retrieving relevant evidence from the evidence library is to retrieve evidence related to the statement by calculating the semantic similarity between the statement to be detected and the evidence in the evidence library, that is, for the statement to be retrieved, relevant evidence that may be helpful in judging the authenticity is retrieved in the corresponding evidence knowledge base.

[0037] The statement to be tested refers to a sentence or paragraph published in social media, news articles or other information sources whose authenticity needs to be verified; a statement usually contains a specific statement, claim or opinion, and it is necessary to retrieve and analyze relevant evidence to determine whether it is true, false or cannot be determined by existing information;

[0038] The relevant evidence in step 1 includes text evidence and image evidence;

[0039] Textual evidence refers to textual materials related to the claim to be verified. These materials can include news articles, research reports, official statements, blog posts, social media posts, etc. Textual evidence provides detailed information and background information to support or refute the claim, and is usually analyzed and relevant features are extracted through natural language processing technology;

[0040] Image evidence refers to pictures or images related to the claim to be verified. These images can include photos, screenshots, charts, illustrations, etc., providing visual support or rebuttal information; image evidence is analyzed and feature extracted through computer vision technology to assist in verifying the authenticity of the claim.

[0041] Among them, in the step 1, the large semantic similarity model is used to extract important information from the evidence text, and the interference of redundant content on the retrieval is excluded. The large semantic similarity model is used to extract important information from the text evidence, and a summary is made. The summary is regarded as an index of relevant evidence. When performing similarity calculation, the summary content generated by the large semantic similarity model is used to exclude redundant content to reduce interference with the retrieval results.

[0042] Specifically, the step 1 includes the following steps:

[0043] Step 101: construct relevant indexes for text evidence to form a new index library;

[0044] During the index construction process, a large semantic similarity model is used to summarize the text evidence and predict the evidence statements that may be related to the evidence text. The prompt prefix used is "which statements the following evidence text may be related to", and the final extracted index is built into an index library.

[0045] Step 102: Train a semantic similarity model and a multimodal fact verification model for retrieving relevant text evidence. The training is performed based on the pre-trained model SBERT for calculating semantic similarity. The training adopts a contrastive learning method to select evidence related to each statement as a positive example, and selects some evidence from the remaining corpus as negative examples for contrastive learning. During the training process, the degree of correlation between the evidence and the statement is determined by calculating the cosine similarity between the statement features and the evidence features, and InfoNceLoss is used as the loss function used in the retrieval training.

[0046] Step 103: Training a multimodal fact verification model CLIP for retrieving claim-related image evidence, using a multimodal pre-trained model as a basis for training, and the training method is similar to the semantic similarity model for retrieving claim-related text evidence.

[0047] Step 104: Use the trained semantic similarity model to extract features of the statement and the features of the evidence index, calculate their cosine similarity, sort them according to the similarity, and select the top 5 most relevant evidence as the retrieved text evidence.

[0048] Step 105: Use the trained multimodal fact-checking model to extract features of the claim and features of the image evidence, calculate their cosine similarity, sort them according to the similarity, and select the top 5 most relevant images as the retrieved image evidence.

[0049] Here, when the semantic similarity model and the multimodal fact verification model CLIP are used in step 1 and combined with the contrastive learning method for training, evidence related to the statement is selected as a positive sample during the training process, and relevant evidence is randomly selected from evidence unrelated to the statement as a negative sample for contrastive learning to participate in the training, so as to improve the accuracy of retrieving relevant text evidence and image evidence from the evidence library.

[0050] Step 2: Construct the evidence relationship graph: Introduce the discourse relationship enhancement module, use the existing natural language processing tool (Graphene) as the discourse relationship enhancement module to extract the corresponding discourse relations from the relevant evidence, construct the discourse relationship heterogeneous graph based on the extracted discourse relations, and fuse and extract the features of the discourse relationship heterogeneous graph through the graph neural network to form the evidence features of the discourse relationship heterogeneous graph;

[0051] The step 2 comprises the following steps:

[0052] Step 201: Use the text relationship extraction tool Graphene to analyze the discourse relationship of the text evidence. Graphene can convert complex sentences into simple independent sentences and identify the relationship between sentences to complete text cleaning and discourse analysis.

[0053] The relationships extracted by Graphene include BACKGROUND, CAUSE, CONTRAST, etc. In order to simplify the representation of the discourse relationship graph, they are simplified into selection relationship, parallel relationship and causal relationship. The corresponding relationships are as follows: selection relationship: DISJUNCTION; parallel relationship: LIST, CONTRAST; causal relationship: BACKGROUND, CAUSE, CONDITION, PURPOSE, CAUSE C, RESULT, RESULT C.

[0054] Step 202: Based on the extracted discourse relations, a directed heterogeneous graph as an evidence relation graph is constructed to form a discourse relation heterogeneous graph. The discourse relation heterogeneous graph includes statement nodes, text evidence nodes, and image evidence nodes, as well as five types of edges: selection relations, parallel relations, causal relations, inverse causal relations, and ordinary connection relations. The specific construction of the discourse relation heterogeneous graph is completed using DGL. DGL (Deep Graph Library) is an open source library for graph neural networks (GNNs) that aims to simplify the construction, training, and application of graph neural networks.

[0055] Step 203: Connect the statement and related image evidence to the nodes of the discourse relation heterogeneous graph as node features, use the attention mechanism of the graph neural network to control message transmission, alleviate the long-distance dependency problem, enhance the ability of the multimodal fact verification model to understand evidence and judge the authenticity of the statement, use the attention mechanism to perform weighted summation on the node features, complete the fusion and extraction of the node features of the discourse relation heterogeneous graph, and obtain the final discourse relation heterogeneous graph evidence features.

[0056] In step 2, CLIP is used to encode the text evidence nodes and image evidence nodes in the constructed discourse relationship heterogeneous graph to obtain the initial feature representation of the nodes. Different from conventional graph neural networks, heterogeneous graph networks have multiple types of nodes and edges. Therefore, the influence of different types of nodes and edges needs to be considered when passing messages. The nodes (i, j) associated by the meta-path φ are represented by α i,j Represents the importance of node i to node j; the attention weight is calculated by the following formula

[0057]

[0058] Among them, h i and h j Represent the feature representation of nodes i and j respectively, α Φ represents the attention vector, || represents concatenation, represents the set of adjacent nodes of node i under the meta-path φ, α ij represents the attention weight, h k Represents the features of the neighboring nodes of node i.

[0059] Then, this study superimposed two layers of discourse relationship heterogeneous graph attention network and input its output node features into the bidirectional GRU module. Through its gating mechanism, it can effectively process the dependencies in long sequences, making up for the shortcomings of graph neural networks in capturing long-range dependencies and obtaining the final node representation.

[0060] Finally, the attention mechanism is used to perform weighted summation on the node features to obtain the final discourse relationship heterogeneous graph features. The formula is as follows:

[0061] α i =Softmax(LeakyReLU(H i ||C))

[0062] G=∑ i∈N α i H i

[0063] Among them, H i represents the feature of the i-th node in the graph, C represents the declared feature, α i is the weight coefficient of the ith node, N represents all nodes in the graph, and G represents the final discourse relation heterogeneous graph feature.

[0064] The module for judging the authenticity of the statement to be tested classifies whether the statement is true, false or cannot be judged based on the extracted statement and evidence features. By integrating multimodal evidence and analyzing the discourse relationship structure, the model's performance in judging the authenticity of the statement is ultimately improved.

[0065] The step 2 uses the natural language processing tool Graphene as an evidence discourse relationship enhancement module to analyze relevant text evidence and extract discourse relationships, and constructs a multimodal heterogeneous graph based on the extracted discourse relationships, and classifies and summarizes the extracted discourse relationships, and finally simplifies them into four relationships: selection, parallel, causal, and reverse causal. The multimodal heterogeneous graph finally obtained contains three types of nodes: statement nodes, text evidence nodes, and image evidence nodes, and five types of edges: selection relationships, parallel relationships, causal relationships, reverse causal relationships, and ordinary connection relationships. CLIP is used to extract features from the text nodes and image nodes in the constructed discourse relationship heterogeneous graph to obtain the initial feature representation of the nodes.

[0066] In the process of training the multimodal fact-verification model, the heterogeneous graph attention network in the graph neural network is used to fuse and extract the features of the multimodal heterogeneous graph. The heterogeneous graph attention network is stacked with two layers, and the node features it outputs are input into the bidirectional GRU module. The dependency relationship in the long sequence is processed through its gating mechanism, and the node features are weighted summed up using the attention mechanism to obtain the final discourse heterogeneous graph features, so as to enhance the multimodal fact-verification model's understanding of evidence and its ability to judge the authenticity of the statement.

[0067] Step 3: Authenticity judgment: Using complete relevant evidence as a unit, merge the claim features and evidence features to analyze whether the relevant evidence supports the claim to be judged;

[0068] Here, in step 3, the evidence of different modalities is comprehensively utilized through feature extraction and fusion methods to improve the overall judgment ability of the model; the heterogeneous graph features are combined with the features of the statement and evidence to enhance the model's understanding of the evidence and the ability to judge the authenticity of the statement.

[0069] In the process of evidence fusion and analysis in step 3, the statement features and evidence features are fused through the attention mechanism to improve the overall judgment ability of the model; the attention distribution of the statement and evidence is calculated through the cross-attention mechanism, the statement features are used as the query, and the features of the evidence are used as the key and value, and finally the semantic understanding features are obtained by calculating the average features. The heterogeneous graph features are combined with the features of the statement and evidence to enhance the model's understanding of the evidence and the ability to judge the authenticity of the statement. The final classification results include true, false and uncertain.

[0070] By analyzing the relevant evidence of the statement to be tested, the relevant evidence can be visualized and added as a new feature to the training of the multimodal fact verification model. The evidence discourse relationship diagram is constructed and used for the training of the neural network, thereby improving the multimodal fact verification model's ability to judge the authenticity of the statement.

[0071] Specifically, the step 3 includes the following steps:

[0072] Step 301: Use NLTK to perform sentence processing on the retrieved text evidence, input the sentence-processed text evidence and image evidence together into the multimodal fact verification model CLIP, and use the multimodal fact verification model CLIP to extract text evidence features, image evidence features and features of the statement to be detected. The features used here are the average features of the output of the last hidden layer of the multimodal fact verification model CLIP after removing the first [CLS].

[0073] Step 302: Calculate the attention distribution of the claim and evidence through the cross-attention mechanism, where the claim feature is used as the query and the evidence feature is used as the key and value for feature fusion, and capture the interactive information and gap by calculating the difference and product between the claim feature and the evidence feature.

[0074] Step 303: The fused features are input into the pooling layer and the linear layer to determine the authenticity of the claim. The categories include supported, refuted, and insufficient information (NEI). Insufficient information means that it cannot be determined. The loss function uses cross entropy loss.

[0075] Through the above technical scheme, the present invention realizes automatic fact-checking with multimodal evidence fusion and discourse relationship enhancement, solves the lack of multimodal automatic fact-checking, improves the accuracy and reliability of the multimodal fact-checking model in judging the authenticity of statements, and can effectively make judgments on the statements to be tested, effectively addressing the problem of false information dissemination in social media, which is of great significance to reducing the spread of false information.

[0076] The above is only a preferred embodiment of the present invention. It should be noted that, for those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications should also be considered as the protection scope of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention shall be implemented according to the conventional means in the art unless otherwise specified and limited.

Claims

1. A fact-checking method based on multimodal evidence fusion and discourse relationship enhancement, characterized by: The following steps are involved: Step 1: Evidence retrieval: Retrieve relevant evidence from the evidence library for the claim to be tested; Step 2: Construct an evidence relationship graph: Use natural language processing tools to extract discourse relations from relevant evidence, construct a discourse relationship heterogeneous graph based on the extracted discourse relations, and fuse and extract the features of the discourse relationship heterogeneous graph through a graph neural network to form the evidence features of the discourse relationship heterogeneous graph; Step 3: Authenticity judgment: Use complete relevant evidence as a unit, merge the claim features and evidence features, and analyze whether the relevant evidence supports the claim to be judged.

2. The fact-checking method based on multimodal evidence fusion and discourse relationship enhancement according to claim 1, characterized in that: The relevant evidence in step 1 includes text evidence and image evidence.

3. The fact-checking method based on multimodal evidence fusion and discourse relationship enhancement according to claim 1, characterized in that: The step 2 comprises the following steps: Step 101: construct relevant indexes for text evidence to form a new index library; Step 102: training a semantic similarity model and a multimodal fact-checking model for retrieving relevant text evidence; Step 103: training a multimodal fact-checking model for retrieving claim-related image evidence; Step 104: Use the trained semantic similarity model to extract features of the statement and the evidence index, calculate the cosine similarity, sort them according to the similarity, and select the top 5 most relevant evidence as the retrieved text evidence; Step 105: Use the trained multimodal fact-checking model to extract features of the claim and features of the image evidence, calculate their cosine similarity, sort them according to the similarity, and select the top 5 most relevant images as the retrieved image evidence.

4. The fact-checking method based on multimodal evidence fusion and discourse relationship enhancement according to claim 3 is characterized by: The semantic similarity model and the multimodal fact verification model in step 1 are trained using a contrastive learning method. During the training process, evidence related to the statement is selected as a positive sample, and relevant evidence unrelated to the statement is randomly selected from evidence as negative samples for contrastive learning to participate in the training.

5. The fact-checking method based on multimodal evidence fusion and discourse relationship enhancement according to claim 3 is characterized by: The step 2 comprises the following steps: Step 201: Analyze the discourse relations of text evidence using a text relation extraction tool; Step 202: constructing a directed heterogeneous graph as an evidence relationship graph based on the extracted discourse relations to form a discourse relationship heterogeneous graph; Step 203: Connect the statement and related image evidence to the discourse relationship heterogeneous graph nodes as node features, use the attention mechanism of the graph neural network to control message transmission, use the attention mechanism to perform weighted summation on the node features, complete the fusion and extraction of the discourse relationship heterogeneous graph node features, and obtain the final discourse relationship heterogeneous graph evidence features.

6. The fact-checking method based on multimodal evidence fusion and discourse relationship enhancement according to claim 5, characterized in that: The text relationship extraction tool is Graphene.

7. The fact-checking method based on multimodal evidence fusion and discourse relationship enhancement according to claim 5 is characterized by: The discourse relationship heterogeneous graph includes statement nodes, text evidence nodes and image evidence nodes, as well as five types of edges: selection relationship, parallel relationship, causal relationship, reverse causal relationship and ordinary connection relationship.

8. The fact-checking method based on multimodal evidence fusion and discourse relationship enhancement according to claim 5, characterized in that: The step 3 comprises the following steps: Step 301: Sentence processing is performed on the text evidence, and the sentence-processed text evidence and the image evidence are input into the multimodal fact verification model, and the multimodal fact verification model is used to extract text evidence features, image evidence features and features of the statement to be detected; Step 302: Calculate the attention distribution of the claim and evidence through the cross attention mechanism, where the claim feature is used as the query and the evidence feature is used as the key and value for feature fusion; Step 303: The fused features are input into the pooling layer and the linear layer to judge the authenticity of the statement, and the categories include true, false and insufficient information.

Citation Information

Patent Citations

  • Multimodal fusion Mongolian rumor detection method based on knowledge perception attention network

    CN114925682A

  • Multi-modal attention rumor detection method based on causal atlas

    CN116501877A

  • Multi-modal fact checking method based on entity enhancement and vertical check

    CN118364124A

Cited By

  • Self-adaptive multi-modal fact checking method and device with interpretability

    CN122196253A

  • Adaptive multi-modal fact checking method and device with explainability

    CN122196253B