Data processing method and device based on multi-agent cooperation and electronic equipment

Abstract

Embodiments of the present disclosure disclose a data processing method and device based on multi-agent collaboration and an electronic device. A specific implementation of the method includes: obtaining multi-modal business data to be processed; performing collaborative information extraction on the multi-modal business data based on multi-agents to obtain business information; performing structured event extraction on the business information to obtain structured business information; performing preliminary logical verification on the structured business information based on a business time sequence knowledge graph to obtain a logical verification result; in response to determining that the logical verification result meets a complex condition, performing deep logical verification on the structured business information based on a logical verification agent in the multi-agents and the business time sequence knowledge graph to obtain a deep logical verification result as the logical verification result; and injecting the structured business information into the business time sequence knowledge graph according to the logical verification result to obtain an updated business knowledge graph. The implementation improves the efficiency of data processing.

Description

Technical Field

[0001] The embodiments of this disclosure relate to the field of computer technology, and more specifically to data processing methods, apparatuses, and electronic devices based on multi-agent collaboration. Background Technology

[0002] The increasing demands for precision in processing and analyzing business data in sectors such as industrial chains and finance have made accurately extracting effective business information from multi-source, heterogeneous data a key challenge in data processing. Currently, the common approach to data processing involves using different modules to independently process data from different sources, and then storing the results in a database. For example, when processing multimodal business data including visual images, audio recordings, and text reports, separate processing modules need to be built for visual, audio, text, and structured data, and the extraction results are then validated manually or through simple rules.

[0003] However, when using the above methods, a common technical problem is that industry chain and financial data not only include structured data, but also unstructured data such as research reports, news pictures, and conference call recordings. Moreover, the data quality varies greatly and often contains a lot of noise and false information. A large amount of data can reduce the real-time performance of data processing and easily lead to a decrease in data processing efficiency. Summary of the Invention

[0004] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0005] Some embodiments of this disclosure propose data processing methods, apparatuses, and electronic devices based on multi-agent collaboration to address the technical problems mentioned in the background section above.

[0006] In a first aspect, some embodiments of this disclosure provide a data processing method based on multi-agent collaboration. The method includes: acquiring multimodal business data to be processed, wherein the multimodal business data includes visual data, audio data, and text data; extracting collaborative information from the multimodal business data based on preset multi-agents to obtain business information, wherein the multi-agents include a visual agent, an audio agent, and a semantic agent; extracting structured events from the business information to obtain structured business information; performing preliminary logical verification on the structured business information based on a preset business temporal knowledge graph to obtain a logical verification result; in response to determining that the logical verification result satisfies preset complex conditions, performing deep logical verification on the structured business information based on a preset logical verification agent and the business temporal knowledge graph to obtain a deep logical verification result, which is used as the logical verification result; and injecting the structured business information into the business temporal knowledge graph according to the logical verification result to obtain an updated business knowledge graph.

[0007] Secondly, some embodiments of this disclosure provide a data processing apparatus based on multi-agent collaboration. The apparatus includes: an acquisition unit configured to acquire multimodal business data to be processed, wherein the multimodal business data includes visual data, audio data, and text data; a collaborative information extraction unit configured to extract collaborative information from the multimodal business data based on preset multi-agents to obtain business information, wherein the multi-agents include a visual agent, an audio agent, and a semantic agent; and a structured event extraction unit configured to extract structured events from the business information to obtain structured business information. The system comprises: a preliminary logic verification unit, configured to perform preliminary logic verification on the structured business information based on a preset business temporal knowledge graph, and obtain a logic verification result; a deep logic verification unit, configured to perform deep logic verification on the structured business information based on a preset logic verification agent and the business temporal knowledge graph in response to determining that the logic verification result satisfies preset complex conditions, and obtain a deep logic verification result as the logic verification result; and an injection unit, configured to inject the structured business information into the business temporal knowledge graph according to the logic verification result, and obtain an updated business knowledge graph.

[0008] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any implementation of the first aspect above.

[0009] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0010] The above-described embodiments of this disclosure have the following beneficial effects: The data processing efficiency is improved through the multi-agent collaborative data processing method of some embodiments of this disclosure. Specifically, the reason for insufficient data processing efficiency is that industry chain and financial related data not only include structured data but also unstructured data such as research reports, news images, and conference call recordings. Furthermore, the data quality varies greatly, often containing a large amount of noise and false information, and a large amount of data reduces the real-time performance of data processing. Based on this, the multi-agent collaborative data processing method of some embodiments of this disclosure first acquires the multimodal business data to be processed. This multimodal business data includes visual data, audio data, and text data. Then, based on preset multi-agent agents, collaborative information extraction is performed on the multimodal business data to obtain business information. This multi-agent agent includes a visual agent, an audio agent, and a semantic agent. This avoids the bias caused by a single model processing data and improves the real-time performance and quality of information extraction through parallel processing. Secondly, structured event extraction is performed on the business information to obtain structured business information. Converting the business information into machine-readable structured data improves the processing efficiency of logical verification. Next, based on a pre-defined business temporal knowledge graph, a preliminary logical verification is performed on the structured business information to obtain the logical verification result. The temporal knowledge graph allows for rapid determination of any contradictions between the structured business information and existing knowledge, improving the real-time performance of the logical verification. Then, in response to the determination that the logical verification result satisfies pre-defined complex conditions, a deep logical verification is performed on the structured business information based on a pre-defined logical verification agent and the aforementioned business temporal knowledge graph, yielding a deep logical verification result, which serves as the final logical verification result. For complex information, a dedicated logical verification agent is activated for deeper reasoning and verification, avoiding feeding all data into a complex inference model and effectively balancing computational resource consumption and accuracy. Finally, based on the logical verification result, the structured business information is injected into the aforementioned business temporal knowledge graph, resulting in an updated business knowledge graph. This allows verified information to be updated in the knowledge graph, achieving efficient data processing. Attached Figure Description

[0011] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0012] Figure 1 This is a flowchart of some embodiments of the data processing method based on multi-agent collaboration according to the present disclosure; Figure 2 This is an event extraction flowchart based on some embodiments of the multi-agent collaborative data processing method disclosed herein; Figure 3 This is a schematic diagram of knowledge graph updating based on the multi-agent collaborative data processing method disclosed herein; Figure 4 This is a schematic diagram of the structure of some embodiments of the data processing device based on multi-agent collaboration according to the present disclosure; Figure 5 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0013] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0014] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0015] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0016] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0017] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0018] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] Figure 1A flow 100 of some embodiments of a multi-agent cooperative data processing method according to the present disclosure is shown. This multi-agent cooperative data processing method includes the following steps: Step 101: Obtain the multimodal business data to be processed.

[0020] In some embodiments, the execution entity (e.g., a computing device) of the multi-agent collaborative data processing method can acquire multimodal business data to be processed. This multimodal business data can be data related to industry chain analysis. The multimodal business data can include visual data, audio data, and text data. In practice, the execution entity can pull multimodal data from a target port in real time through a pre-deployed real-time data access system as the multimodal business data to be processed. This real-time data access system can be Kafka. The target port can be a port for publishing business data.

[0021] As an example, the aforementioned execution entity can pull multimodal business data from the target port via Kafka. This multimodal business data will be temporarily stored in a Kafka message queue for downstream agents to read. In financial business scenarios, the target port can include, but is not limited to: exchange market data servers, corporate announcement publishing servers, financial information platforms, and web crawler collection servers. The multimodal business data can include visual data, audio data, and text data from the industry chain or financial sector. The visual data can be unstructured data in the financial sector in the form of images and videos, such as company financial report images, stock price K-line charts, and financial news illustrations. The audio data can be unstructured data in the financial sector in the form of sound and voice, such as conference call recordings and financial interview audio. The text data can be data in the financial sector in the form of text, such as financial news and policy documents.

[0022] Step 102: Based on the preset multi-agent system, perform collaborative information extraction on the multimodal business data to obtain business information.

[0023] In some embodiments, the aforementioned execution entity can collaboratively extract information from the aforementioned multimodal business data based on a preset multi-agent system to obtain business information. The aforementioned multi-agent system includes: a visual agent, an audio agent, and a semantic agent. The visual agent can be an agent specifically designed to process visual data, such as the VisionTransformer (ViT). The visual agent can be used to convert visual data into text descriptions. The audio agent can be an agent specifically designed to process audio data, such as the Whisper speech recognition model. The audio agent can be used to convert the audio data into text descriptions. The semantic agent can be an agent specifically designed to process text data, such as the Longformer-Encoder-Decoder Model (LED). The semantic agent can be used to perform deep semantic understanding on the text data to obtain corresponding text information.

[0024] In practice, the aforementioned executing entities can use the visual agent, audio agent, and semantic agent among the aforementioned multi-agent entities to process the visual data, audio data, and text data in the aforementioned multimodal business data in parallel, and combine the output text information into business information.

[0025] Optionally, the aforementioned execution entity, based on a pre-defined multi-agent system, performs collaborative information extraction on the aforementioned multimodal business data to obtain business information, which may include the following steps: The first step, in response to the determination that the multimodal service data includes visual data, is to extract information from the visual data based on the visual agent included in the multi-agent system, thereby obtaining visual features. In practice, when the multimodal service data includes visual data, the visual data can be input into the visual encoder included in the visual agent to obtain visual features. These visual features can be 512-dimensional feature vectors. The visual agent can include a visual encoder and a visual decoder. The visual encoder can be used to extract features from image or video data. For example, when the visual agent is a Vision Transformer, the visual encoder can be a TransformerEncoder.

[0026] The second step, in response to the determination that the multimodal service data includes audio data, involves extracting information from the audio data based on the audio agent included in the multi-agent system to obtain audio features. In practice, when the multimodal service data includes audio data, the audio data can be input into the audio encoder included in the audio agent to obtain audio features. These audio features can be 512-dimensional feature vectors. The audio agent can include an audio encoder and an audio decoder. The audio encoder can be used to extract features from the audio data. For example, when the audio agent is Whisper, the audio encoder can be a Whisper Encoder.

[0027] The third step, in response to the determination that the multimodal business data includes text data, is to extract information from the text data based on the semantic agent included in the multi-agent system, thereby obtaining text features. In practice, when the multimodal business data includes text data, the text data can be input into the semantic encoder included in the semantic agent to obtain text features. The text features can be 512-dimensional feature vectors. The semantic agent can include a semantic encoder and a semantic decoder. The semantic encoder can be used to extract features from the text data. The semantic decoder can be used to predict the text content corresponding to the text features. For example, when the semantic agent is a Longformer-Encoder-Decoder Model, the semantic encoder can be a Longformer Encoder, and the semantic decoder can be a Longformer Decoder.

[0028] The fourth step involves multimodal embedding and fusion of the aforementioned visual, audio, and text features to generate business information. In practice, firstly, layer normalization (LayerNorm) is applied to the visual, audio, and text features to obtain layer-normalized visual, audio, and text features. Next, these layer-normalized visual, audio, and text features are input into a ReLU activation function to obtain non-linearly activated visual, audio, and text features. Secondly, these non-linearly activated visual, audio, and text features are concatenated along their channel dimensions to obtain a multimodal fusion feature vector. Finally, the semantic decoder is used to decode the multimodal fusion feature vector to obtain the business information.

[0029] Step 103: Extract structured events from the business information to obtain structured business information.

[0030] In some embodiments, the aforementioned executing entity can perform structured event extraction on the aforementioned business information to obtain structured business information. This structured business information may include a set of business facts. The business facts in the set may include head entities, relationships, tail entities, timestamps, data sources, etc. In practice, structured event extraction can be performed on the aforementioned business information using a pre-defined entity relationship identification technology. This entity relationship identification technology can be the CasRel model.

[0031] As an example, suppose the business information includes: "Company A raised the price of memory cards (December 16, 2025)" and "Media B published an article stating that artificial intelligence technology increased the demand for memory chips (January 24, 2026)". The structured business information obtained by extracting the above business information into structured events could include: {Artificial intelligence technology, driving demand, memory chips, 2026-1-24, Media B}, {Company A, raising prices, memory cards, 2025-12-16, Company A}.

[0032] In practice, a common technical challenge in structured event extraction from business information is that this information often contains numerous implicit relationships. Extracting binary relationships makes cross-sentence reasoning difficult, leading to missed or false positives in event extraction and thus reducing its accuracy. Therefore, the following solution is proposed.

[0033] Optionally, the aforementioned executing entity may extract structured events from the aforementioned business information to obtain structured business information, which may include the following steps: The first step is to perform word segmentation on the above business information to obtain a business word segmentation sequence. The business word segments in this sequence can be individual words. The business word segmentation sequence can be a sequence of words arranged in order. In practice, a pre-defined word segmentation model can be used to segment the above business information to obtain the business word segmentation sequence. This word segmentation model can be the JieBa word segmentation model. This word segmentation model can be a word segmentation model specifically designed for the business domain. For example, a JieBa word segmentation model customized based on a financial dictionary.

[0034] The second step involves encoding each business segment in the aforementioned business segmentation sequence to generate business segmentation features, resulting in a business segmentation feature sequence. These business segmentation features can be represented by vectors. In practice, a pre-defined text encoder can be used to encode each business segment in the sequence to generate business segmentation features. This text encoder can be trained using a dataset from the business domain. Alternatively, it can be a FinBERT model.

[0035] The third step involves assigning self-attention weights to each business segmentation feature in the aforementioned business segmentation feature sequence to generate segmentation attention features, resulting in a segmentation attention feature sequence. These segmentation attention features can be represented by vectors. In practice, a pre-defined self-attention module can be used to assign self-attention weights to each business segmentation feature in the aforementioned business segmentation feature sequence to generate segmentation attention features, thus obtaining the segmentation attention feature sequence. This self-attention module can be a multi-head self-attention module.

[0036] Specifically, for each business word segmentation feature in the aforementioned business word segmentation feature sequence, firstly, the correlation between this business word segmentation feature and the other business word segmentation features in the sequence can be determined using the aforementioned self-attention module. This attention module can be a multi-head self-attention module comprising eight attention heads. The correlation can be a value between 0 and 1. Next, the correlation can be multiplied by the aforementioned business word segmentation feature to obtain the word segmentation attention feature.

[0037] The fourth step involves performing head entity recognition on the aforementioned word segmentation attention feature sequence to obtain a head entity information set. This head entity information set includes the head entity location and head entity features. The head entity location can be the position of the corresponding business word segment within the aforementioned business word segmentation sequence. The head entity features can be the corresponding word segmentation attention features. In practice, a pre-defined head entity recognition decoder can be used to perform head entity recognition on the aforementioned word segmentation attention feature sequence to obtain the head entity information set. The head entity recognition decoder may include a fully connected layer and a softmax activation function.

[0038] Specifically, for each word segmentation attention feature in the above word segmentation attention feature sequence, firstly, the word segmentation attention feature can be feature-mapped using the fully connected layer to obtain the mapped feature. The mapped feature can be represented by a vector, for example, a 256-dimensional vector. Next, the head entity category probability corresponding to the mapped feature can be determined using the Softmax activation function. The head entity category probability can be a percentage. The head entity category probability represents the probability that the business word segmented by the word segmentation attention feature is a head entity. Finally, when the head entity category probability is higher than a preset probability value, the position corresponding to the word segmentation attention feature can be determined as the head entity position, and the word segmentation attention feature can be determined as a head entity feature, thus obtaining the head entity information. The probability value can be a preset value between 0 and 1, for example, 0.75, without specific limitation here.

[0039] Fifth, for each header entity in the above header entity information set, perform the following steps: The first sub-step involves fusing the head entity features from the aforementioned head entity information with each word segmentation attention feature sequence from the aforementioned word segmentation attention feature sequence to obtain a fused feature sequence. In practice, the fused feature sequence can be obtained using the following formula: .

[0040] in, This represents the i-th fusion feature. This represents the attention feature of the i-th word segmentation. This represents the characteristics of the head entity. and Both represent learnable parameter matrices. , , , , All of these represent parameters, and no specific limitations are specified here.

[0041] The second sub-step involves performing relation prediction and tail entity prediction on the aforementioned fused feature sequence to obtain relation information and tail entity information. The relation information can include relation position and relation features. The relation position can be the position of the corresponding business word segment within the aforementioned business word segmentation sequence. The relation features can be the segmentation attention features of the corresponding business word segment. The tail entity information can include tail entity position and tail entity features. The tail entity position can be the position of the corresponding business word segment within the aforementioned business word segmentation sequence. The tail entity features can be the segmentation attention features of the corresponding business word segment. In practice, a pre-defined relation decoder can be used to predict relations on the aforementioned fused feature sequence to obtain relation information. This relation decoder can include a fully connected layer and a Softmax activation function. Secondly, a pre-defined tail entity decoder can be used to predict tail entities on the aforementioned fused feature sequence to obtain tail entity information. This tail entity decoder can include a fully connected layer and a Softmax activation function.

[0042] The third sub-step generates business facts based on the aforementioned head entity information, relationship information, and tail entity information. In practice, business facts can be generated by combining the corresponding business segments in the aforementioned business segmentation sequence according to the head entity position in the head entity information, the relationship position in the relationship information, and the tail entity position in the tail entity information. The event extraction flowchart corresponding to the first to sixth steps can be shown below. Figure 2 As shown. Figure 2 The blank boxes in the text can represent features, including: business word segmentation features, word segmentation attention features, and fusion features.

[0043] The sixth step is to combine the generated business facts to obtain structured business information. This structured business information may include the aforementioned business facts.

[0044] To improve the accuracy of event extraction, firstly, a dedicated word segmentation model is used to convert business information into a sequence of business word segments. This avoids the fragmentation of specialized terms, reduces ambiguous segmentation, and improves the accuracy of entity boundary segmentation. Next, a dedicated text encoder encodes each business word segment into a vector feature, reducing semantic drift. Then, a pre-defined self-attention module determines the association weight of the global word segment with the current word segment, capturing implicit relationships across different dimensions in parallel to avoid missed or false detections in event extraction. Secondly, relationship prediction is automatically triggered by using the head entity as a condition. Next, tail entities are retrieved across the entire sequence, enabling cross-sentence association and avoiding the single-sentence limitations of binary relation extraction. Finally, joint decoding ensures consistency among head entities, relationships, and tail entities, further improving the accuracy of event extraction.

[0045] Step 104: Based on the preset business time sequence knowledge graph, perform preliminary logical verification on the structured business information to obtain the logical verification results.

[0046] In some embodiments, the aforementioned executing entity can perform preliminary logical verification on the structured business information based on a preset business temporal knowledge graph to obtain logical verification results. The aforementioned business temporal knowledge graph can be a temporal knowledge graph specifically for the business domain. The aforementioned logical verification results can include logical tags corresponding to each business fact in the aforementioned business fact set. The aforementioned logical tags can indicate whether the business fact is trustworthy or untrustworthy. In practice, the aforementioned executing entity can use Link Prediction Modeling (LPM) technology, based on the aforementioned temporal knowledge graph, to perform logical verification on each business fact included in the aforementioned structured business information to generate logical tags, obtaining a logical tag set as the logical verification result. The logical tags in the aforementioned logical tag set correspond one-to-one with the business facts in the aforementioned business fact set.

[0047] Optionally, the aforementioned execution entity performs preliminary logical verification on the aforementioned structured business information based on a preset business time-series knowledge graph to obtain the logical verification result, which may include the following steps: The first step, for each business fact in the above set of business facts, is to perform the following steps: The first sub-step involves standardizing and aligning the head entities, relationships, and tail entities in the aforementioned business facts based on a pre-defined business time-series knowledge graph to generate standardized business facts. These standardized business facts include: standard head entities, standard relationships, standard tail entities, and timestamps. The aforementioned business time-series knowledge graph can include an existing entity set and an existing relationship set. The existing entity set can contain all entities in the aforementioned business time-series knowledge graph. The existing relationship set can contain all relationships in the aforementioned business time-series knowledge graph. In practice, for the head entity in the aforementioned business fact, semantic retrieval can be used to determine whether the head entity exists in the existing entity set of the business time-series knowledge graph. If it exists, the corresponding existing entity is determined as the standard head entity. If it does not exist, the head entity is determined as a new entity, serving as the standard head entity. Next, the standard relationship corresponding to the aforementioned relationship and the standard tail entity corresponding to the aforementioned tail entity can be determined through the steps of determining the standard head entity, while the timestamp remains unchanged. Finally, the combination of the aforementioned standard head entities, standard relationships, standard tail entities, and timestamps can be determined as the standardized business fact.

[0048] As an example, suppose the business fact is {Head Entity: New Energy Company, Relationship: Made, Tail Entity: Solid State Battery, Timestamp: 2024-12}. After standardization and alignment, the head entity is mapped to "New Energy Company" in the existing entity set, the relationship is mapped to "Mass Production" in the existing relationship set, and the tail entity has no corresponding existing entity. Therefore, the tail entity is directly determined as the standard tail entity "Solid State Battery". The final standardized business fact is: {Standard Head Entity: New Energy Company, Standard Relationship: Mass Production, Standard Tail Entity: Solid State Battery, Timestamp: 2024-12}.

[0049] The second sub-step involves entity embedding of the aforementioned standard header and tail entities to obtain standard header and tail entity features. Both standard header and tail entity features can be features that incorporate knowledge graph topological relationships. Both standard header and tail entity features can be in vector form; for example, the vector dimension can be 512. In practice, entity embedding of the aforementioned standard header and tail entities can be performed using a pre-defined entity embedding module included in the student logic verification model to obtain standard header and tail entity features.

[0050] The aforementioned logic verification student model can correspond to a logic verification teacher model. Both the aforementioned logic verification student model and the aforementioned logic verification teacher model can be neural network models used to determine the credibility of business facts. The aforementioned logic verification student model can be a lightweight model obtained by knowledge distillation of the aforementioned logic verification teacher model. The aforementioned logic verification student model can have the same structure as the aforementioned logic verification teacher model, but different parameter dimensions. The aforementioned logic verification student model can include: an entity embedding module, a relation embedding module, a bilinear transformation module, and a decoder. The aforementioned entity embedding module can be used to convert entities from textual form to vector form. The aforementioned entity embedding module can be a T-GNN model. The aforementioned relation embedding module can be used to convert relations from textual form to vector form. The aforementioned relation embedding module can be a TransE model. The aforementioned bilinear transformation module is used to combine head entity features and relation features. The aforementioned decoder is used to determine the confidence level of whether the pre-combined features and the tail entity features match.

[0051] The third sub-step involves embedding the aforementioned standard relationships to obtain standard relationship features. These standard relationship features can be features with the same dimensions as the aforementioned standard header entity features and standard tail entity features. In practice, the aforementioned relationship embedding module can be used to embed the aforementioned standard relationships to obtain standard relationship features.

[0052] The fourth sub-step involves performing a bilinear transformation on the aforementioned standard header entity features and standard relation features to obtain the pre-combined features. These pre-combined features can be feature vectors representing the combination logic of the header entity and relation. In practice, the executing entity can use the bilinear transformation module to perform a bilinear transformation on the aforementioned standard header entity features and standard relation features to obtain the pre-combined features. Specifically, a preset bilinear transformation matrix can be used to perform matrix operations on the standard header entity features and standard relation features to obtain the pre-combined features. The bilinear transformation matrix can be a square matrix. For example, the bilinear transformation matrix can be a 512×512 dimensional matrix. Specifically, the product of the aforementioned standard header entity features, the aforementioned bilinear transformation matrix, and the transpose of the aforementioned standard relation features can be determined as the pre-combined features.

[0053] The fifth sub-step involves determining the factual confidence levels corresponding to the aforementioned pre-combined features and the aforementioned standard tail entity features. The factual confidence level characterizes the degree to which the aforementioned business fact is true. The factual confidence level can be a value between 0 and 1. In practice, the factual confidence levels between the aforementioned pre-combined features and the aforementioned standard tail entity features can be determined based on the aforementioned decoder. Specifically, firstly, the aforementioned pre-combined features and the aforementioned standard tail entity features can be concatenated to obtain concatenated features. Next, the concatenated features can be feature-mapped using the fully connected layer included in the aforementioned decoder to obtain mapped features. Finally, the factual confidence levels corresponding to the mapped features can be determined using the activation function included in the aforementioned decoder.

[0054] The second step is to generate logical verification results based on the determined confidence levels of each fact. In practice, firstly, for each determined confidence level, if the confidence level is less than a preset confidence threshold, the logical label of the business fact corresponding to that confidence level can be determined as untrustworthy. The confidence threshold can be a preset value between 0 and 1, such as 0.8, without specific limitation here. When the confidence level is greater than or equal to the confidence threshold, the logical label of the business fact corresponding to that confidence level can be determined as trustworthy. Then, the combination of the determined logical labels can be used to determine the logical verification result.

[0055] Step 105: In response to determining that the logic verification result meets the preset complex conditions, deep logic verification is performed on the structured business information based on the preset logic verification agent and business time sequence knowledge graph to obtain the deep logic verification result, which is used as the logic verification result.

[0056] In some embodiments, the execution entity may, in response to determining that the logical verification result satisfies a preset complex condition, perform deep logical verification on the structured business information based on a preset logical verification agent and the aforementioned business time-series knowledge graph, obtaining a deep logical verification result as the logical verification result. The aforementioned complex condition may be the presence of logically untrusted tags in the logical verification result. In practice, for each logical tag in the logical verification result, when the logical tag is untrusted, deep logical verification can be performed on the business fact corresponding to the logical tag to generate a deep logical tag. Then, the logical tags can be replaced using the deep logical tags to update the logical tag set. Finally, the finally updated logical tag set can be determined as the logical verification result.

[0057] Optionally, the execution entity performs deep logical verification on the structured business information based on a preset logical verification agent and the aforementioned business time-series knowledge graph to obtain the deep logical verification result, which may include the following steps: The first step, for each business fact in the set of business facts included in the above structured business information, is to perform the following steps: The first sub-step involves performing deep fact verification on the standardized business facts corresponding to the aforementioned business facts, based on a pre-defined logic verification agent and the aforementioned business time-series knowledge graph, to obtain deep fact confidence. The aforementioned logic verification agent can be the aforementioned logic verification teacher model. In practice, the aforementioned logic verification teacher model can be used, referring to the operations corresponding to the second to fifth sub-steps in the steps, to perform deep fact verification on the standardized business facts corresponding to business facts logically labeled as untrustworthy, and obtain deep fact confidence.

[0058] In practice, the logic verification teacher model has more parameters than the student model, and can more accurately identify the credibility of business facts.

[0059] The second step is to generate deep logic verification results based on the obtained confidence levels of each deep fact. In practice, firstly, for each determined confidence level of a deep fact, if the confidence level is less than the aforementioned confidence threshold, the deep logic label of the business fact corresponding to that confidence level can be determined as untrustworthy. If the confidence level is greater than or equal to the aforementioned confidence threshold, the deep logic label of the business fact corresponding to that confidence level can be determined as trustworthy. Then, the combination of the determined deep logic labels can be used to determine the logic verification result.

[0060] Step 106: Based on the logical verification results, inject the structured business information into the business time sequence knowledge graph to obtain the updated business knowledge graph.

[0061] In some embodiments, the execution entity can inject the structured business information into the business temporal knowledge graph based on the logical verification results to obtain an updated business knowledge graph. In practice, for each logical tag included in the logical verification results, when the logical tag is reliable, the business fact corresponding to the logical tag can be added to the knowledge graph. Specifically, the nodes of the head entity and tail entity in the business fact can be determined in the knowledge graph by searching or creating, and then the head entity and tail entity are connected through the aforementioned relationship. Here, the aforementioned relationship can be used as an edge between the head entity and the tail entity. Here, the knowledge graph update diagram corresponding to the process of injecting structured business information into the business temporal knowledge graph can be as follows: Figure 3 As shown. Figure 3In the diagram, <Artificial Intelligence, Driving Demand, Memory Chip> represents the business facts to be injected. Since the original business time-series knowledge graph already contains "Artificial Intelligence" and "Memory Chip" nodes, a "Driving Demand" relationship can be created, connecting the "Artificial Intelligence" and "Memory Chip" nodes through this relationship.

[0062] Optionally, the execution entity injects the structured business information into the business time-series knowledge graph based on the logical verification results to obtain an updated business knowledge graph, which may include the following steps: The first step, based on the above logical verification results, is to identify each business fact in the set of business facts included in the structured business information that satisfies the logical verification condition as the target business fact, thus obtaining the target business fact set. The logical verification condition is that the logical label is trustworthy. In practice, for each business fact in the set of business facts included in the structured information, when the logical label corresponding to the business fact is untrustworthy, the business fact can be identified as the target business fact.

[0063] The second step is to perform the following steps for each target business fact in the above target business fact set: The first sub-step involves performing concept matching on the head entities, relations, and tail entities in the target business fact based on the aforementioned business timeline knowledge graph, obtaining a matching result set. This matching result set includes: head entity matching concepts and matching degrees, relation matching concepts and matching degrees, and tail entity matching concepts and matching degrees. In practice, for the head entities in the target business fact, a cosine similarity algorithm can be used to determine the similarity between the head entity and each existing entity in the existing entity set included in the aforementioned business timeline knowledge graph, obtaining a similarity set. Then, the highest similarity value in the similarity set can be determined as the matching degree, and the existing entity corresponding to the similarity can be determined as the head entity matching concept. Finally, the head entity matching concept and the matching degree can be determined as the matching result. Similarly, the relation matching concept and matching degree corresponding to the aforementioned relations can be determined, as well as the tail entity matching concept and matching degree corresponding to the aforementioned tail entities.

[0064] The second sub-step, in response to determining that each matching result in the aforementioned matching result set satisfies the preset matching conditions, involves conceptually aligning the aforementioned target business facts based on the aforementioned matching result set to obtain aligned business facts. In practice, when each matching degree included in the aforementioned matching result set is greater than or equal to a preset matching threshold, the head entity matching concepts included in the aforementioned matching result set can be used as aligned head entities, the relation matching concepts as aligned relations, and the tail entity matching concepts as aligned tail entities to obtain aligned business facts. The aforementioned matching threshold can be a preset value, such as 0.9, and is not specifically limited here.

[0065] The third sub-step, in response to determining that there are matching results in the aforementioned matching result set that do not meet the aforementioned matching conditions, generates a new ontology information set based on the aforementioned matching result set and the aforementioned target business facts, and performs concept alignment on the aforementioned target business facts based on the aforementioned new ontology information set to obtain aligned business facts. In practice, for each matching result in the aforementioned matching result set, when the matching degree corresponding to the aforementioned matching result is less than the aforementioned matching threshold, firstly, when the aforementioned matching degree is a matching result of a head entity, the aforementioned head entity can be identified as new ontology information. Then, this new ontology information is added to the existing entity set of the aforementioned business temporal knowledge graph. Finally, the aforementioned new ontology information can be identified as an aligned head entity. This process continues until the alignment relationship and aligned tail entity are determined, resulting in aligned business facts.

[0066] The third step is to add each aligned business fact to the aforementioned business temporal knowledge graph, resulting in an updated business knowledge graph. In practice, for each aligned business fact, the nodes corresponding to the head and tail entities in the aforementioned business temporal knowledge graph can be determined separately, and the nodes corresponding to the head and tail entities can be connected through the relationships in the aforementioned aligned business facts to update the aforementioned business knowledge graph, thus obtaining the updated business knowledge graph.

[0067] Optionally, the aforementioned business time-series knowledge graph may include a set of quadruples. First, the executing entity can combine the alignment header entity, alignment relation, alignment tail entity, and corresponding timestamp from the alignment business facts into a quadruple. This quadruple can be: <alignment header entity, alignment relation, alignment tail entity, timestamp>. Next, this quadruple can be added to the set of quadruples corresponding to the aforementioned business time-series knowledge graph to obtain the updated set of quadruples, which serves as the updated business knowledge graph.

[0068] Optionally, after step 206 above, the executing entity may also perform the following steps: The first step is to determine the set of supply chain transmission factors corresponding to the structured business information mentioned above, based on the updated business knowledge graph. These supply chain transmission factors can be numerical indicators that quantify the strength of risk and return transmission between upstream and downstream enterprises in the supply chain. For example, the development of artificial intelligence technology has led to increased demand and prices for chips. In this case, chip demand and prices are supply chain transmission factors.

[0069] In practice, a common technical challenge in mining transmission factors from knowledge graphs is that, due to the large number of nodes and edges, it is difficult to effectively learn the deep implicit relationships between nodes when dealing with complex or entirely new business scenarios. Furthermore, this leads to increased computational resource consumption, thus reducing the real-time responsiveness and accuracy of factor mining. Therefore, the following solution is proposed.

[0070] Optionally, based on the updated business knowledge graph, the set of industry chain transmission factors corresponding to the structured business information is determined, including: The first sub-step involves extracting a target industry chain subgraph from the structured business information based on the updated business knowledge graph. This target industry chain subgraph can be a graph structure representing the entities and relationships between them within the structured business information. In practice, firstly, the head and tail entities in the aligned business facts corresponding to the structured business information can be merged into a target entity set. For each target entity in this target entity set, a breadth-first search (BFS) algorithm can be used to perform a breadth-first search on the updated business knowledge graph, starting from the target entity, to obtain a set of associated entities. Each associated entity in this set can be an entity directly or indirectly connected to the target entity via edges. Next, the determined sets of associated entities and the corresponding quadruples of each entity can be used to define the target industry chain subgraph.

[0071] The second sub-step involves performing similar subgraph matching on the target industry chain subgraph based on the updated business knowledge graph, resulting in a set of similar subgraphs. These similar subgraphs are those highly similar to the target industry chain subgraph in terms of topological structure, entity relationships, and industry attributes. In practice, a subgraph matching algorithm can be used to calculate the structural and semantic similarity between the target industry chain subgraph and historical industry chain subgraphs and industry chain subgraphs in the same domain already stored in the updated business knowledge graph, thus obtaining the set of similar subgraphs.

[0072] The third sub-step involves performing network encoding on the target industry chain subgraph to obtain a multi-level grid feature set. This multi-level grid feature set includes: a node embedding feature set, a relationship embedding feature set, and subgraph embedding features. The node embedding features in the node embedding feature set can be vector-form features corresponding to each entity node in the target industry chain subgraph. The relationship embedding features can be vector-form features corresponding to each relationship edge in the target industry chain subgraph. The subgraph embedding features can be vector-form features corresponding to the target industry chain subgraph. The dimensions of the node embedding feature set, the relationship embedding feature set, and the subgraph embedding features can be the same, for example, 512 dimensions. In practice, firstly, based on a preset graph attention module, feature encoding can be performed on each entity node in the target industry chain subgraph to obtain the node embedding feature set. Then, feature encoding can be performed on each relationship edge in the target industry chain subgraph to obtain the relationship embedding feature set. Next, a preset graph pooling module can be used to concatenate the node embedding features from the node embedding feature set and the relationship embedding features from the relationship embedding feature set, followed by global pooling to obtain the subgraph embedding features.

[0073] The fourth sub-step involves determining the transmission coefficient corresponding to each node in the target industry chain sub-graph based on the aforementioned similar sub-graph set and multi-level grid feature set, thus obtaining a transmission coefficient set. The transmission coefficients in this set characterize the transmission capacity of a node within the industry chain, i.e., the node's ability to be influenced by its upstream or downstream nodes. The values ​​of these transmission coefficients can range from 0 to 1, with values ​​closer to 1 indicating stronger transmission capacity. In practice, a pre-defined industry chain transmission model can be used, taking the multi-level network feature set as input and outputting the transmission coefficient set.

[0074] Specifically, for each node included in the aforementioned target industry chain subgraph, firstly, the relationship embedding features of each edge directly connected to the node can be concatenated to obtain adjacency relationship features. Next, based on a graph attention mechanism (GAT), the aggregated neighbor features of the node in the target industry chain subgraph can be determined. Then, the node embedding features, adjacency relationship features, and aggregated neighbor features corresponding to the node can be concatenated to obtain multi-level node features. Secondly, for each similar subgraph in the aforementioned similar subgraph set, the subgraph matching degree between the similar subgraph and the target industry chain subgraph can be determined based on a preset meta-path attention mechanism. The determined subgraph matching degrees are defined as a subgraph matching degree set. Then, the subgraph matching degree set can be used as weights to perform weighted aggregation of the multi-level node features to obtain aggregated features. Finally, the aggregated features can be input into a sigmoid activation function to obtain transmission coefficients.

[0075] The fifth sub-step involves performing cross-subgraph calibration on each transmission coefficient in the aforementioned transmission coefficient set based on the aforementioned similar subgraph set, to generate calibrated transmission coefficients and obtain a calibrated transmission coefficient set. In practice, for each transmission coefficient in the aforementioned transmission coefficient set, firstly, an entity type matching algorithm can be used to determine the similar nodes in each similar subgraph of the similar subgraph set corresponding to the transmission coefficient, obtaining a set of similar nodes. These similar nodes can be those with entity type matching or attribute similarity to the aforementioned node. Next, the transmission coefficients corresponding to each similar node in the aforementioned set of similar nodes can be obtained, obtaining a set of similar transmission coefficients. Each similar node in the aforementioned set of similar nodes can have a corresponding transmission coefficient. Then, based on the aforementioned subgraph matching degree set, a weighted average can be performed on each similar transmission coefficient in the aforementioned set of similar transmission coefficients to obtain a cross-subgraph correction value. Finally, the weighted sum between the aforementioned transmission coefficients and the aforementioned cross-subgraph correction value can be determined as the calibrated transmission coefficient.

[0076] The sixth sub-step involves identifying the node corresponding to each time-series calibration transmission coefficient in the aforementioned time-series calibration transmission coefficient set that exceeds a preset transmission threshold as a supply chain transmission factor, thus obtaining a supply chain transmission factor set. The aforementioned transmission threshold can be a preset value, such as 0.85, and is not specifically limited here.

[0077] To improve the real-time response capability and accuracy of factor mining, firstly, subgraphs related to structured business information are identified from the overall knowledge graph, reducing computational resource consumption compared to full graph computation. Secondly, a set of similar subgraphs with similar topological structures and entity relationships to the target industry chain subgraph is matched from the knowledge graph. This allows for the extraction of implicit patterns across industries and domains from the similar subgraph set. Next, multi-level network encoding is performed on the target industry chain subgraph to obtain nodes, relationships, and sub-graphs. Figure 3 This involves embedding features across multiple dimensions. This enables cross-level feature interaction, improving feature learning efficiency. Subsequently, by learning the transmission factors in similar subgraphs, the transmission coefficient of each node in the target industry chain subgraph is generated, achieving quantitative analysis of industry chain transmission factors. Then, through temporal attention weighting, the transmission coefficient of each node is corrected, suppressing random noise interference and ensuring the selected factors possess temporal stability and reliability, thereby improving the accuracy of factor mining.

[0078] The second step involves issuing risk warnings to related industry information terminals based on the aforementioned set of industry chain transmission factors. These industry information terminals can be mobile terminal devices that are communicatively connected to the aforementioned executing entity. In practice, the executing entity can send each industry chain transmission factor from the aforementioned set of industry chain transmission factors to the aforementioned industry information terminals for risk warning purposes.

[0079] Further reference Figure 4 As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a data processing device based on multi-agent collaboration, and these device embodiments are similar to... Figure 1 Corresponding to the method embodiments shown, this multi-agent collaborative data processing device can be specifically applied to various electronic devices.

[0080] like Figure 4 As shown, a data processing device 400 based on multi-agent collaboration in some embodiments includes: an acquisition unit 401, a collaborative information extraction unit 402, a structured event extraction unit 403, a preliminary logic verification unit 404, a deep logic verification unit 405, and an injection unit 406. The acquisition unit 401 is configured to acquire multimodal business data to be processed, wherein the multimodal business data includes visual data, audio data, and text data; the collaborative information extraction unit 402 is configured to extract collaborative information from the multimodal business data based on preset multi-agents to obtain business information, wherein the multi-agents include a visual agent, an audio agent, and a semantic agent; the structured event extraction unit 403 is configured to extract structured events from the business information to obtain structured business information; the preliminary logic verification unit 404 is configured to... The structured business information is initially logically verified using a preset business time-series knowledge graph to obtain a logical verification result. The deep logical verification unit 405 is configured to perform a deep logical verification on the structured business information based on a preset logical verification agent and the business time-series knowledge graph in response to determining that the logical verification result satisfies a preset complex condition, and obtain a deep logical verification result as the logical verification result. The injection unit 406 is configured to inject the structured business information into the business time-series knowledge graph according to the logical verification result to obtain an updated business knowledge graph.

[0081] It is understandable that the units described in the multi-agent collaborative data processing device 400 and the reference Figure 1 The steps in the described method correspond to each other. Therefore, the operations, features, and beneficial effects described above for the method also apply to the data processing device 400 based on multi-agent cooperation and the units contained therein, and will not be repeated here.

[0082] The following is for reference. Figure 5 It shows a schematic diagram of the structure of an electronic device (e.g., a computing device) 500 suitable for implementing some embodiments of the present disclosure. Figure 5 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0083] like Figure 5As shown, the electronic device 500 may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 501, which can perform various appropriate actions and processes according to a program stored in a read-only memory 502 or a program loaded from a storage device 508 into a random access memory 503. The random access memory 503 also stores various programs and data required for the operation of the electronic device 500. The processing unit 501, the read-only memory 502, and the random access memory 503 are interconnected via a bus 504. An input / output interface 505 is also connected to the bus 504.

[0084] Typically, the following devices can be connected to the input / output interface 505: input devices 506 including, for example, a touchscreen, touchpad, keyboard, mouse, camera, microphone, accelerometer, gyroscope, etc.; output devices 507 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 508 including, for example, magnetic tape, hard disk, etc.; and communication devices 509. Communication device 509 allows electronic device 500 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 5 An electronic device 500 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 5 Each box shown can represent a device or multiple devices as needed.

[0085] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 509, or installed from a storage device 508, or installed from a read-only memory 502. When the computer program is executed by the processing device 501, it performs the functions defined above in the methods of some embodiments of this disclosure.

[0086] It should be noted that, in some embodiments of this disclosure, the computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0087] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0088] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to: acquire multimodal business data to be processed, wherein the multimodal business data includes: visual data, audio data, and text data; perform collaborative information extraction on the multimodal business data based on preset multi-agents to obtain business information, wherein the multi-agents include: a visual agent, an audio agent, and a semantic agent; perform structured event extraction on the business information to obtain structured business information; perform preliminary logical verification on the structured business information based on a preset business temporal knowledge graph to obtain a logical verification result; in response to determining that the logical verification result satisfies preset complex conditions, perform deep logical verification on the structured business information based on a preset logical verification agent and the aforementioned business temporal knowledge graph to obtain a deep logical verification result, which serves as the logical verification result; and inject the structured business information into the aforementioned business temporal knowledge graph according to the logical verification result to obtain an updated business knowledge graph.

[0089] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0090] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0091] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0092] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A data processing method based on multi-agent collaboration, characterized in that, include: Acquire multimodal business data to be processed, wherein the multimodal business data includes: visual data, audio data, and text data; Based on a preset multi-agent system, collaborative information extraction is performed on the multimodal business data to obtain business information. The multi-agent system includes: a visual agent, an audio agent, and a semantic agent. The business information is subjected to structured event extraction to obtain structured business information; Based on a preset business time-series knowledge graph, the structured business information is subjected to preliminary logical verification to obtain the logical verification result; In response to determining that the logical verification result meets preset complex conditions, based on the preset logical verification agent and the business time sequence knowledge graph, deep logical verification is performed on the structured business information to obtain the deep logical verification result, which is used as the logical verification result. Based on the logical verification results, the structured business information is injected into the business time-series knowledge graph to obtain the updated business knowledge graph; The process of extracting structured events from the business information includes: The business information is segmented into words to obtain a business word segmentation sequence; Each business word in the business word segmentation sequence is encoded to generate business word segmentation features, resulting in a business word segmentation feature sequence; Self-attention weights are assigned to each business segmentation feature in the business segmentation feature sequence to generate segmentation attention features, resulting in a segmentation attention feature sequence. Head entity recognition is performed on the segmented attention feature sequence to obtain a head entity information set, wherein the head entity information in the head entity information set includes head entity location and head entity features; For each header entity in the header entity information set, perform the following steps: The head entity features in the head entity information are fused with each word segmentation attention feature sequence in the word segmentation attention feature sequence to obtain a fused feature sequence; Relationship prediction and tail entity prediction are performed on the fused feature sequence to obtain relationship information and tail entity information, wherein the relationship information includes relationship position and relationship features; Based on the head entity information, the relationship information, and the tail entity information, business facts are generated; The generated business facts are combined to obtain structured business information; The structured business information includes a set of business facts, whereby the business facts in the set include head entities, relationships, tail entities, and timestamps; and The structured business information is preliminarily logically verified based on a preset business time-series knowledge graph to obtain the logical verification results, including: For each business fact in the set of business facts, perform the following steps: Based on a preset business time-series knowledge graph, the head entity, relationship, and tail entity in the business fact are standardized and aligned to generate standardized business facts, wherein the standardized business facts include: standard head entity, standard relationship, standard tail entity, and timestamp; Entity embedding is performed on the standard header entity and the standard tail entity to obtain standard header entity features and standard tail entity features; The standard relation is embedded to obtain the standard relation features; A bilinear transformation is performed on the standard header entity features and the standard relation features to obtain the pre-combined features; Determine the fact confidence level corresponding to the preceding combined features and the standard tail entity features; Based on the determined confidence levels of each fact, generate logical verification results; The step of injecting the structured business information into the business temporal knowledge graph based on the logical verification result to obtain the updated business knowledge graph includes: Based on the logical verification results, each business fact in the set of business facts included in the structured business information that satisfies the logical verification conditions is identified as a target business fact, thus obtaining a target business fact set; For each target business fact in the target business fact set, perform the following steps: Based on the business time sequence knowledge graph, concept matching is performed on the head entity, relation and tail entity in the target business fact to obtain a matching result set, wherein the matching result set includes: head entity matching concept and matching degree, relation matching concept and matching degree, tail entity matching concept and matching degree; In response to determining that each matching result in the matching result set satisfies the preset matching conditions, the target business fact is conceptually aligned based on the matching result set to obtain aligned business facts; In response to determining that there are matching results in the matching result set that do not meet the matching conditions, a new ontology information set is generated based on the matching result set and the target business fact, and the target business fact is conceptually aligned based on the new ontology information set to obtain aligned business facts; The obtained aligned business facts are added to the business time-series knowledge graph to obtain the updated business knowledge graph.

2. The method according to claim 1, characterized in that, The method further includes: Based on the updated business knowledge graph, determine the set of industry chain transmission factors corresponding to the structured business information; Based on the aforementioned set of industry chain transmission factors, risk warnings are issued for related industry information terminals.

3. The method according to claim 1, characterized in that, The pre-defined multi-agent system performs collaborative information extraction on the multimodal business data to obtain business information, including: In response to determining that the multimodal service data includes visual data, information is extracted from the visual data based on the visual agents included in the multi-agent system to obtain visual features; In response to determining that the multimodal service data includes audio data, information is extracted from the audio data based on the audio agent included in the multi-agent system to obtain audio features; In response to determining that the multimodal service data includes text data, information is extracted from the text data based on the semantic agents included in the multi-agent system to obtain text features; The visual features, audio features, and text features are multimodal embedded and fused to generate business information.

4. The method according to claim 1, characterized in that, The logic verification agent, based on a preset logic verification mechanism and the business time-series knowledge graph, performs deep logic verification on the structured business information to obtain deep logic verification results, including: For each business fact in the set of business facts included in the structured business information, perform the following steps: Based on the logical verification agent and the business time-series knowledge graph, deep fact verification is performed on the standardized business facts corresponding to the business facts to obtain the deep fact confidence. Based on the obtained confidence scores of each deep fact, a deep logic verification result is generated.

5. A data processing device based on multi-agent collaboration, characterized in that, include: The acquisition unit is configured to acquire multimodal business data to be processed, wherein the multimodal business data includes: visual data, audio data, and text data; The collaborative information extraction unit is configured to perform collaborative information extraction on the multimodal business data based on a preset multi-agent system to obtain business information. The multi-agent system includes a visual agent, an audio agent, and a semantic agent. A structured event extraction unit is configured to extract structured events from the business information to obtain structured business information; wherein, extracting structured events from the business information includes: The business information is segmented into words to obtain a business word segmentation sequence; Each business word in the business word segmentation sequence is encoded to generate business word segmentation features, resulting in a business word segmentation feature sequence; Self-attention weights are assigned to each business segmentation feature in the business segmentation feature sequence to generate segmentation attention features, resulting in a segmentation attention feature sequence. Head entity recognition is performed on the segmented attention feature sequence to obtain a head entity information set, wherein the head entity information in the head entity information set includes head entity location and head entity features; For each header entity in the header entity information set, perform the following steps: The head entity features in the head entity information are fused with each word segmentation attention feature sequence in the word segmentation attention feature sequence to obtain a fused feature sequence; Relationship prediction and tail entity prediction are performed on the fused feature sequence to obtain relationship information and tail entity information, wherein the relationship information includes relationship position and relationship features; Based on the head entity information, the relationship information, and the tail entity information, business facts are generated; The generated business facts are combined to obtain structured business information; A preliminary logic verification unit is configured to perform preliminary logic verification on the structured business information based on a preset business time-series knowledge graph, and obtain a logic verification result; wherein, the structured business information includes a business fact set, and the business facts in the business fact set include head entities, relationships, tail entities, and timestamps; and The structured business information is preliminarily logically verified based on a preset business time-series knowledge graph to obtain the logical verification results, including: For each business fact in the set of business facts, perform the following steps: Based on a preset business time-series knowledge graph, the head entity, relationship, and tail entity in the business fact are standardized and aligned to generate standardized business facts, wherein the standardized business facts include: standard head entity, standard relationship, standard tail entity, and timestamp; Entity embedding is performed on the standard header entity and the standard tail entity to obtain standard header entity features and standard tail entity features; The standard relation is embedded to obtain the standard relation features; A bilinear transformation is performed on the standard header entity features and the standard relation features to obtain the pre-combined features; Determine the fact confidence level corresponding to the preceding combined features and the standard tail entity features; Based on the determined confidence levels of each fact, generate logical verification results; The deep logic verification unit is configured to, in response to determining that the logic verification result meets preset complex conditions, perform deep logic verification on the structured business information based on a preset logic verification agent and the business time sequence knowledge graph, and obtain a deep logic verification result as the logic verification result. An injection unit is configured to inject the structured business information into the business temporal knowledge graph according to the logical verification result, thereby obtaining an updated business knowledge graph; wherein, the step of injecting the structured business information into the business temporal knowledge graph according to the logical verification result to obtain an updated business knowledge graph includes: Based on the logical verification results, each business fact in the set of business facts included in the structured business information that satisfies the logical verification conditions is identified as a target business fact, thus obtaining a target business fact set; For each target business fact in the target business fact set, perform the following steps: Based on the business time sequence knowledge graph, concept matching is performed on the head entity, relation and tail entity in the target business fact to obtain a matching result set, wherein the matching result set includes: head entity matching concept and matching degree, relation matching concept and matching degree, tail entity matching concept and matching degree; In response to determining that each matching result in the matching result set satisfies the preset matching conditions, the target business fact is conceptually aligned based on the matching result set to obtain aligned business facts; In response to determining that there are matching results in the matching result set that do not meet the matching conditions, a new ontology information set is generated based on the matching result set and the target business fact, and the target business fact is conceptually aligned based on the new ontology information set to obtain aligned business facts; The obtained aligned business facts are added to the business time-series knowledge graph to obtain the updated business knowledge graph.

6. An electronic device, characterized in that, include: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1 to 4.

7. A computer-readable medium, characterized in that, It stores a computer program thereon, wherein the computer program, when executed by a processor, implements the method as described in any one of claims 1 to 4.

Images