Business intelligence assisted processing method, apparatus, device, and medium

Abstract

The application relates to the technical field of intelligent decision-making, and discloses a business intelligent auxiliary processing method, device, equipment and medium, the method comprising the following steps: determining the modal type of modal business data of a target user for processing a target business; extracting initial modal features based on the modal type and fusing the initial modal features to obtain target modal features; recognizing operation intention and business entities based on the target modal features; performing logical reasoning on structured business information combined by the business entities and the operation intention, judging whether the structured business information has logical conflicts based on a logical reasoning result; when the structured business information does not have logical conflicts, extracting a target decision field and a confidence degree; when the confidence degree is greater than or equal to a confidence degree threshold, automatically auditing and passing the target business; and when the confidence degree is lower than the confidence degree threshold, transferring the target business to manual auditing processing. The application can be applied to a business system platform such as financial technology and medical health, and can improve business processing accuracy.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, and in particular to a business intelligent auxiliary processing method, apparatus, equipment and medium. Background Technology

[0002] With the deepening of digital transformation, the volume of business processing in various industries is experiencing explosive growth, and traditional manual review models can no longer meet the needs of efficient and accurate business processing. Existing intelligent auxiliary processing technologies for business operations are usually based on single-modal data analysis, such as relying solely on text or image information to identify user intent and extract business entities. However, in real-world business scenarios, user-submitted business data often contains multiple modalities, such as text, images, tables, and voice. Single-modal processing methods cannot fully utilize the complementary information between multimodal data, resulting in low accuracy in intent recognition and low completeness in entity extraction.

[0003] Furthermore, existing technologies typically rely on single characteristics or simple rules to make decisions when reviewing business data, lacking the ability to infer and verify the logical consistency within the business data. This means that when inconsistencies or logical conflicts exist within the business data, the system may still make incorrect review decisions, reducing the reliability and accuracy of business processing.

[0004] For example, in healthcare scenarios, the business data submitted by patients typically encompasses multiple modalities, such as electronic medical record texts, medical images, and voice consultation records. Existing technologies often rely solely on a single text modality for medical insurance reimbursement review, neglecting the supplementary diagnostic information contained in the image and voice modalities. This leads to inaccurate identification of reimbursement intentions for complex conditions. Furthermore, existing medical insurance review systems lack the ability to reason and verify the logical consistency between diagnostic conclusions, medication lists, and expense details. This can easily result in the automatic approval of unreasonable reimbursements and the inability to dynamically switch between automatic review and manual verification, posing a significant risk to the medical insurance fund.

[0005] For example, in fintech scenarios, existing loan preliminary review systems lack in-depth logical reasoning when reviewing the matching relationship between income and loan amount, and lack quantitative assessment of the confidence level of review decisions. This may result in high-risk transactions being automatically approved by error, or low-risk transactions being unnecessarily transferred to manual review, which seriously affects business processing efficiency and risk control security.

[0006] Therefore, how to implement a business intelligence-assisted processing method to improve the accuracy of business processing is a technical problem that urgently needs to be solved. Summary of the Invention

[0007] This invention provides a business intelligent auxiliary processing method, apparatus, equipment and medium, the main purpose of which is to solve the problem of low accuracy in business processing.

[0008] Firstly, to achieve the above objectives, the present invention provides a business intelligent auxiliary processing method, comprising: Obtain various modal business data used by the target user to process the target business, and determine the modal type to which each modal business data belongs; Initial modal features corresponding to each modal type are extracted from the service data of that modality, and multiple initial modal features are fused by self-attention using a preset self-attention mechanism to obtain target modal features; Based on the target modal features, the target user's operational intent is identified, and business entities are extracted from the modal business data; The business entity and the operation intention are structurally combined into structured business information, and logical reasoning is performed on the structured business information. Based on the result of the logical reasoning, it is determined whether there is a logical conflict in the structured business information. When there is no logical conflict in the structured business information, the target decision field in the structured business information is extracted, and the confidence level of the target decision field is calculated. Determine whether the confidence level is greater than or equal to a preset confidence threshold; When the confidence level is greater than or equal to a preset confidence threshold, the target business is automatically approved. When the confidence level is lower than the preset confidence threshold, the target business will be transferred to manual review.

[0009] Secondly, the present invention also provides a business intelligent auxiliary processing device, comprising: The modality type determination module is used to acquire multiple modal business data used by the target user to process the target business, and to determine the modality type to which each modal business data belongs. The modal feature extraction module is used to extract the initial modal features corresponding to the modal service data based on each modal type, and to perform self-attention fusion of multiple initial modal features using a preset self-attention mechanism to obtain the target modal features; An operation intent recognition module is used to recognize the operation intent of the target user based on the target modal features, and to extract business entities from the modal business data; The logical reasoning and judgment module is used to structurally combine the business entity and the operation intention into structured business information, perform logical reasoning on the structured business information, and determine whether there is a logical conflict in the structured business information based on the result of the logical reasoning. The confidence calculation module is used to extract the target decision field from the structured business information and calculate the confidence of the target decision field when there is no logical conflict in the structured business information. The confidence level determination module is used to determine whether the confidence level is greater than or equal to a preset confidence level threshold; The first business processing module is used to automatically review and approve the target business when the confidence level is greater than or equal to a preset confidence threshold. The second business processing module is used to transfer the target business to manual review when the confidence level is lower than a preset confidence threshold.

[0010] Thirdly, the present invention also provides an electronic device, the electronic device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, which enables the at least one processor to perform the business intelligence auxiliary processing method described above.

[0011] Fourthly, the present invention also provides a computer-readable storage medium storing at least one computer program, which is executed by a processor in an electronic device to implement the business intelligent auxiliary processing method described above.

[0012] In this embodiment of the invention, during the multimodal fusion stage, a self-attention mechanism is used to dynamically capture global dependencies between modalities such as text, speech, and images, effectively suppressing noise interference and generating more discriminative target modal features, significantly improving the accuracy and robustness of subsequent intent recognition and entity extraction. By structurally combining entities and intents and performing logical reasoning, contradictions or constraint conflicts in information can be automatically detected, preventing the system from performing business operations based on erroneous information, thus enhancing pre-verification capabilities and decision interpretability. Extracting target decision fields and calculating confidence levels under the premise of logical non-conflict achieves a quantitative closed loop from information integrity verification to field-level reliability assessment, providing uncertainty measurement basis for high-risk scenarios. An automatic / manual review and triage mechanism based on confidence thresholds maximizes the proportion of automated processing while ensuring security, improving the accuracy of business processing. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of an application environment for a business intelligent auxiliary processing method according to an embodiment of the present invention. Figure 2 A flowchart illustrating a business intelligent auxiliary processing method provided in an embodiment of the present invention; Figure 3 A functional block diagram of a business intelligent auxiliary processing device provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device for implementing a business intelligent auxiliary processing method according to an embodiment of the present invention; Figure 5 This is another structural schematic diagram of an electronic device for implementing a business intelligent auxiliary processing method according to an embodiment of the present invention.

[0015] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] It should be noted that in the technical solutions disclosed in this invention, the acquisition of user information (personal image data (e.g., facial videos or pictures, facial feature videos or pictures, etc.) and personal privacy information (e.g., name, ID number, occupation, address, etc.)) is all completed with the user's knowledge and consent, and the acquisition of the relevant user information is legal and compliant.

[0017] To enable those skilled in the art to better understand the technical solutions of this disclosure, and to fully understand and implement the process of how this disclosure applies technical means to solve technical problems and achieve corresponding technical effects, the technical solutions in the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, not all embodiments. The embodiments of this disclosure and the various features within them can be combined with each other without conflict, and the resulting technical solutions are all within the protection scope of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without creative effort should fall within the protection scope of this disclosure.

[0018] It should be noted that the terms "first," "second," etc., used in this disclosure and the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0019] This application provides a business intelligent auxiliary processing method. The execution subject of the business intelligent auxiliary processing method includes, but is not limited to, at least one of the electronic devices that can be configured to execute the device provided in this application, such as a server or a terminal. In other words, the business intelligent auxiliary processing method can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0020] The intelligent auxiliary processing method of this invention can be applied to, for example, Figure 1 In this application environment, the client communicates with the server via a network. The server can obtain multiple modal business data from the target user for processing target business through the client, determine the multiple modal types to which the multiple modal business data belong; extract the initial modal features of the corresponding modal business data based on the multiple modal types, and perform self-attention fusion on multiple initial modal features to obtain target modal features; identify the target user's operation intention based on the target modal features, and extract business entities from the modal business data; combine the business entities and operation intentions into structured business information, and perform logical reasoning on the structured business information, judging whether there is a logical conflict based on the result of the logical reasoning; when there is no logical conflict in the structured business information, extract the target decision field from the structured business information and calculate the confidence level of the target decision field; when the confidence level is greater than or equal to the preset confidence threshold, the target business is automatically approved; when the confidence level is lower than the preset confidence threshold, the target business is transferred to manual review and processing, and finally the business processing result is output and fed back to the client.

[0021] The client can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. The server can be implemented using a standalone server or a server cluster consisting of multiple servers. The invention will now be described in detail through specific embodiments.

[0022] Reference Figure 2 The diagram shown is a flowchart illustrating a business intelligence-assisted processing method according to an embodiment of the present invention. In this embodiment, the business intelligence-assisted processing method includes: S1. Obtain various modal business data used by the target user to process the target business, and determine the modal type to which each modal business data belongs.

[0023] In this embodiment of the invention, in a healthcare scenario, a patient uploads an electrocardiogram (ECG) waveform file and a text description of their chief complaint. The format identifier of the ECG file is extracted as a waveform data type, and after successful matching through a mapping table, it is determined to be a physiological signal modality. However, if the symptom description fails to match through the mapping table, a modality recognition model is invoked. After analyzing its semantic features, the model predicts it to be a natural language modality, thus ultimately determining two modality types. In a fintech scenario, when a user submits a loan application, they upload a screenshot of their income statement and a photo of their handwritten signature. The screenshot is identified as an image modality and successfully matched. The signature photo, after model analysis, is predicted to be a mixed image and text modality after it contains handwritten text and a seal image. Meanwhile, the fields in the application form belong to the structured text modality, ultimately determining three modality types: image, text, and structured text. This lays a multimodal data foundation for subsequent feature extraction and intent recognition.

[0024] In this embodiment of the invention, determining the modal type to which various modal service data belong includes: Obtain the data format identifier of each piece of business data from the various modal business data; The data format identifier is matched with a preset modality type mapping table; For successfully matched business data, the modality of the business data is determined according to the corresponding modality in the modality mapping table; For business data that fails to match, a preset modality recognition model is invoked to classify and predict the business data, thereby obtaining the predicted modality type and prediction confidence of the business data. When the prediction confidence level is greater than or equal to a preset prediction confidence level threshold, the prediction modality type is determined as the modality type of the business data.

[0025] In this embodiment of the invention, the format tag information pre-written in the file header or encapsulation structure of the modal service data is read. These tags are usually written by the data generating end according to specific specifications when generating modal service data, and include file extension, media type encoding or custom type mark, etc. By parsing the protocol header or file attribute field of the data packet, these identification information are extracted, so as to quickly obtain preliminary clues as to which modality the data may belong without fully parsing the data content, and provide a basis for subsequent classification judgment.

[0026] Specifically, the modality type mapping table is a set of key-value pairs that are pre-maintained manually or by a rule engine. Each known data format identifier is explicitly associated with a specific modality type. By using the obtained format identifier as an index to search and compare in the mapping table, and using efficient retrieval methods such as hash lookup or tree structure retrieval, it is determined whether the format identifier exists in the known key set of the mapping table, thereby determining whether there is a directly corresponding modality type relationship.

[0027] Since the relationships in the mapping table are strictly defined and verified in advance and have extremely high accuracy, there is no need to perform complex calculations or reasoning. The deterministic logic is directly used to assign the modality type pointed to by the format identifier as the final result to the business data. This method can quickly process a large amount of standard format data with extremely low computational cost, ensuring the efficiency of the processing flow and the reliability of the results.

[0028] Furthermore, the modality recognition model includes a multi-classifier trained on a convolutional neural network or Transformer architecture. It can deeply analyze the underlying features of the data, such as the texture and color of an image, the spectral features of audio, or the semantic vector of text. When business data is input into the model, after multiple nonlinear transformations and feature extraction, the model outputs a probability distribution containing all possible modality types. The system selects the type with the highest probability as the predicted modality type, and this probability value is used as the prediction confidence level to measure the model's confidence in this judgment.

[0029] Furthermore, the predicted confidence level obtained in the previous step is compared with a pre-set confidence threshold. Only when the predicted confidence level is greater than or equal to the threshold is the predicted modality type officially determined as the modality type of the business data. This process uses confidence gating decision technology, the purpose of which is to ensure the quality of the automatic identification results and avoid mistakenly treating uncertain guesses of the model as facts. By setting a high judgment threshold, it is ensured that the prediction results are only adopted when the model is very confident. For data with low confidence, it is usually intercepted and transferred to the manual review process or marked as an unknown type, thus achieving a good balance between automated processing and accuracy.

[0030] In this embodiment of the invention, multimodal data fusion can compensate for the deficiencies of incomplete information or noise interference in single modality, improve the robustness and accuracy of feature extraction, and thus enhance the performance of core tasks such as business classification, prediction or recommendation; clarifying the modality type provides key prior information for subsequent targeted preprocessing, modality alignment and construction of fusion models, and supports dynamic allocation of computing resources and fault tolerance mechanism in the case of missing modalities.

[0031] S2. Extract the initial modal features corresponding to the modal service data based on each modal type, and use a preset self-attention mechanism to perform self-attention fusion of multiple initial modal features to obtain the target modal features.

[0032] In this embodiment of the invention, the step of extracting the initial modal features corresponding to the modal service data based on each modal type includes: Based on the various modal types, extract image data and text data from the modal business data; Extract the grid features and object features from the image data, and embed and stitch the grid features and object features to obtain the image features of the image data; Calculate the word embedding feature sequence of the text data, and perform feature encoding on the word embedding feature sequence to obtain the text features of the text data; The image features and text features are combined to form the initial modal features of the modal business data.

[0033] In this embodiment of the invention, modality type is used as the classification label. All business data entries are traversed, and content with image type labels in the data load is routed to the image processing pipeline, while content with text type labels is routed to the natural language processing pipeline. This achieves the diversion of heterogeneous data, enabling the subsequent feature extraction process to adopt specialized processing strategies for the data characteristics of different modalities, thus avoiding the feature ambiguity problem caused by processing all data in a uniform way.

[0034] Specifically, for the extracted image data, grid features and object features are extracted separately. Grid feature extraction uses shallow feature extraction technology in convolutional neural networks, which uses multiple convolutional kernels to slide and scan the image, capturing local spatial structure information such as edges, textures, and color patches. Object feature extraction uses deep semantic feature extraction technology, which uses a deeper network structure to perceive a larger receptive field, thereby recognizing the complete outline, components, and semantic meaning of objects in the image. Feature embedding and concatenation technology is used to map these two types of features to vector spaces of the same dimension, and then they are concatenated end-to-end along the feature dimension, so that the detailed information at the grid level and the semantic information at the object level are integrated into a complete image feature vector, allowing subsequent models to perceive both local details and understand global semantics.

[0035] Specifically, for the extracted text data, a pre-trained word vector table is used to convert each word in the text into a dense vector of fixed dimensions. These vectors can reflect the semantic similarity between words, thereby transforming discrete text symbols into continuous numerical representations. Subsequently, feature encoding is performed on this word vector sequence. The word vector sequence of the entire sentence is gradually integrated through a recurrent neural network, so that the encoding result at each position can aggregate the semantic information of the context, and finally output a text feature vector that can represent the meaning of the entire text.

[0036] Furthermore, by employing methods such as direct splicing or weighted fusion, image feature vectors and text feature vectors are merged along the feature dimension, allowing information from different modalities to coexist in the same feature space. This enables the understanding of business data to take into account both visual content and linguistic description, forming an initial modal feature representation that is richer in information and more comprehensive in expression, thus providing high-quality feature input for subsequent in-depth analysis and decision-making.

[0037] In this embodiment of the invention, the step of using a preset self-attention mechanism to perform self-attention fusion of multiple initial modal features to obtain target modal features includes: Each initial modal feature is converted into a feature vector of the same dimension to obtain a sequence of feature vectors; Calculate the query vector, key vector, and value vector for each feature vector in the feature vector sequence; Calculate the attention weight between the two corresponding feature vectors based on the query vector and the key vector; The value vector and the attention weight are weighted and summed to obtain the weighted fusion vector corresponding to the feature vector. The weighted fusion vectors corresponding to all the feature vectors are combined to generate a preliminary fusion feature matrix. The preliminary fused feature matrix is ​​nonlinearly mapped to obtain the target modal features.

[0038] In this embodiment of the invention, since the features of different modalities may come from different network layers or encoders during extraction, their original dimensions are often inconsistent and cannot be directly used for subsequent computational interaction. Therefore, a dedicated fully connected mapping layer is provided for each modal feature. This mapping layer linearly transforms the input features of arbitrary dimensions into a unified preset target dimension space through the learned weight matrix, so that the feature vectors of all modalities are completely equal in length, thereby laying the foundation for consistent data format for subsequent parallel computing and interactive fusion.

[0039] Specifically, three independent weight matrices are used to linearly project the same input feature vector, thereby generating three new vectors with different functions. The query vector represents what information the current feature wants to obtain, the key vector represents what information the current feature contains that can be queried, and the value vector represents the specific content actually carried by the current feature. Through this decomposition method, different aspects of a feature vector are separated, making the interaction between features more flexible and precise, no longer limited to the simple superposition of the original features.

[0040] In detail, the query vector of the current feature is compared one by one with the key vectors of all features in the sequence. The strength of the association between them is measured by calculating the directional similarity between the two in the vector space. The higher the similarity, the more complementary or related the information of the two modalities is. Then, all the calculated similarity scores are normalized and transformed into a probability distribution form with a sum of one. This gives a set of attention weights, which clarifies how much attention should be allocated to each of the other features in the sequence when fusing the current feature.

[0041] Furthermore, attention weights are used as coefficients to weight and accumulate the value vectors. In this way, the output of each position has absorbed the most relevant information in the global scope. Finally, the weighted fusion vectors calculated from all positions are arranged in order and recombined into a complete matrix. Each row in this matrix represents a new feature enhanced by contextual information.

[0042] Furthermore, the matrix is ​​input into a module containing a multi-layer fully connected network, and non-linear activation functions are introduced between layers. Through this progressive non-linear transformation, the fused features can be abstracted and refined more deeply, eliminating redundant information and strengthening key semantics. The final output target modal features have a more powerful expressive ability and can more accurately reflect the deep correlation between multiple modal business data.

[0043] In this embodiment of the invention, the self-attention mechanism can dynamically learn the global dependencies and relative importance between features of different modalities. It can adaptively weight the contributions of each modality without relying on external annotations, thereby effectively highlighting key modalities and suppressing the interference of redundant or noisy modalities on the fusion results. Secondly, this fusion method breaks through the fixed combination pattern of traditional splicing or weighted averaging methods, and can capture deep nonlinear interactions and complementary information between modalities, improve the semantic richness and discriminative ability of feature representation, and significantly enhance the decision-making accuracy and generalization ability of computer systems in multimodal business scenarios.

[0044] S3. Identify the target user's operational intent based on the target modal features, and extract the business entities from the modal business data.

[0045] In this embodiment of the invention, identifying the target user's operational intent based on the target modal features includes: The target modal features are classified using a preset intent classification model to obtain the probability distribution of each intent category in the preset intent category set. Select the intent category with the highest probability value from the probability distribution as the initial intent to be identified; Obtain the target user's historical operation sequence, and extract the target user's historical intent tags within a preset time window from the historical operation sequence; Calculate the prior probability of intent transfer based on the historical intent tags; The preliminary identified intent and the prior probability of intent transfer are fused using Bayesian methods to obtain the posterior probability of each intent category; The intent category with the highest posterior probability is taken as the target recognition intent, and the recognition confidence corresponding to the target recognition intent is calculated. When the recognition confidence level is greater than or equal to the preset recognition confidence level threshold, the target recognition intent is taken as the target user's operation intent.

[0046] In this embodiment of the invention, the intent classification model is typically composed of a multi-layer fully connected network or an attention inference layer. The number of nodes output by the last layer corresponds one-to-one with the total number of intent categories. After normalization, each node represents the probability that the target modal feature belongs to that type of intent. The model comprehensively judges the multimodal deep semantics contained in the target modal features and finally outputs a complete probability distribution. This distribution reflects the model's initial judgment tendency on all possible intents at the current moment.

[0047] Specifically, it iterates through each probability value in the probability distribution, finds the one with the largest value by comparing them one by one, and directly locks the corresponding intent category as the most likely intent. This method is equivalent to making the most direct judgment based solely on the feature performance at the current moment without considering any additional contextual information. Therefore, it is only an intermediate result and requires subsequent steps for correction and enhancement.

[0048] In detail, all operation records of the user within a certain period of time are retrieved from the database or cache, arranged into an operation sequence in chronological order, and the records within a fixed time range closest to the current moment are extracted. The confirmed intent tags corresponding to each operation are extracted from these records. These historical intent tags constitute a true reflection of the user's recent behavioral habits and reflect the actual pattern of changes in the user's intent in continuous operations.

[0049] Furthermore, statistics are compiled on the frequency with which a user performs a different intention after performing a certain intention in the historical operation sequence. The proportion of this frequency to the total number of transitions is used to approximate the probability of transitioning from one intention to another. By statistically analyzing the transition frequencies between all intention pairs, a complete intention transition matrix is ​​constructed, which depicts the statistical patterns of the evolution of user intentions in continuous operations.

[0050] Furthermore, using the probability distribution corresponding to the initially identified intent as likelihood information and the prior probability of intent transfer as prior information, Bayes' theorem is used to fuse the two. The core idea is to use historical behavioral patterns to correct the judgment result at the current moment. If the initially identified intent matches the intent that the user has frequently transferred to in the past... Figure 1 If the intention is consistent with the user's behavior, the posterior probability will be further increased; conversely, if the intention is seriously inconsistent with the user's behavior, the posterior probability will be reduced. Through this fusion, the system obtains a set of posterior probabilities that simultaneously consider current feature performance and historical behavior, making the judgment result more consistent with the user's true behavioral logic.

[0051] Specifically, the item with the largest value is found again in the fused posterior probability distribution, and its corresponding intent category is determined as the final target recognition intent. At the same time, the largest posterior probability value is directly used as the recognition confidence. Only when the recognition confidence is greater than or equal to the threshold is the target recognition intent officially output as the target user's operation intent. When the confidence does not meet the threshold, the current judgment is usually retained and entered into a pending confirmation state or the user is asked to perform a second interaction for confirmation, so as to ensure that the final output operation intent has sufficient reliability and avoid business execution errors caused by misjudgment.

[0052] For example, in a healthcare scenario, the electrocardiogram waveform features and the patient's complaint text features are fused with self-attention and then input into an intent classification model to obtain the probability distribution of whether the patient is seeking medical advice, purchasing medication, or viewing reports. Combined with the patient's recent multiple searches for cold medicine historical intent tags, Bayesian fusion is performed to finally identify the intention to purchase medication and extract the drug name and symptom keywords from the business entity.

[0053] For example, in a fintech scenario, the image features of a bank statement screenshot and the text features of a loan application form are fused together through self-attention and then input into an intent classification model. This model is then combined with the user's recent browsing history tags for financial products to perform fusion inference, ultimately identifying the intent to seek financial advice and extracting business entity information such as amount, term, and risk level from the business data.

[0054] In this embodiment of the invention, the fused target modal features already contain complementary and related information across modalities. Based on this, intent recognition can break through the semantic limitations of a single modality and more accurately understand the user's true operational purpose. Especially in complex interaction scenarios where user expressions are ambiguous, implicit, or multi-channel parallel, it significantly improves the robustness and accuracy of recognition. The simultaneous extraction of business entities can extract key objects, attributes, or relationships from heterogeneous data, providing structured support for intent parsing.

[0055] S4. The business entity and the operation intention are structurally combined into structured business information, and logical reasoning is performed on the structured business information.

[0056] In this embodiment of the invention, the logical reasoning of the structured business information includes: The structured business information is parsed into a set of fact triples; Obtain the rule prerequisites for each business logic rule in the preset rule engine, perform pattern matching between each fact triple in the fact triple set and the rule prerequisites, and obtain a set of business logic rules that successfully match the rule prerequisites. Generate corresponding reasoning conclusion facts based on each business logic rule in the set of business logic rules, and add the reasoning conclusion facts to the set of fact triples to form an extended fact set; Detect whether there are contradictory fact pairs in the extended fact set; If there are contradictory fact pairs in the extended fact set, then the fact triples and business logic rules on which the fact pairs are based are traced back to generate them respectively. Generate a contradiction tracing path for the fact pair based on the fact triples and business logic rules; The severity of the logical conflict is calculated based on the aforementioned contradiction tracing path; If the severity of the conflict is greater than or equal to a preset conflict determination threshold, then the structured business information is determined to have a logical conflict. If the severity of the conflict is less than a preset conflict determination threshold, then the structured business information is determined to have no logical conflict.

[0057] S5. Based on the results of logical reasoning, determine whether there is a logical conflict in the structured business information.

[0058] In this embodiment of the invention, named entity recognition technology in natural language processing is used to locate entity objects from business text. At the same time, relation extraction technology is used to identify the semantic relationships between entities. These entities and relations are combined into a standardized expression of "entity-relationship-entity". Each triple represents an indivisible atomic fact. In this way, unstructured or semi-structured business descriptions are transformed into a logical symbol sequence that computers can understand, providing the most basic logical atomic unit for subsequent automated reasoning.

[0059] Specifically, the rule engine stores a large amount of business logic that "if certain conditions are met, certain operations will be performed". It takes the currently known fact triples as input and tries to fill the variable slots in the rule premises one by one. It uses string matching or semantic similarity comparison to determine whether the facts meet the triggering requirements of the rule. Once all the premises of a rule can find corresponding supporting evidence in the fact triple set, the rule is marked as a successful match, thereby filtering out the set of all executable business logic rules in the current state.

[0060] In this process, each activated rule is traversed, and its conclusion is used as a new fact triple. These newly generated facts represent implicit knowledge derived from known information. These new facts are then incorporated into the existing fact base, allowing subsequent reasoning to continue based on these newly discovered facts.

[0061] Furthermore, the system performs pairwise comparisons on all fact triples in the set to check whether there are completely mutually exclusive attribute descriptions for the same entity, or whether there is a logical direct negation relationship between two facts. For example, one fact claims a certain state is true while another fact claims that the state is false. By constructing a conflict detection matrix between facts, the system can quickly locate those fact combinations that cannot be logically true at the same time, thereby identifying potential logical conflict points.

[0062] Specifically, by utilizing the source markers inherent in each fact or by maintaining a parent-child dependency graph in the reasoning process, starting from the conflicting conclusion facts, we trace back along the reasoning chain, layer by layer, to find the upstream rules that led to the conclusion and the original input facts that support those rules. This tracing process constructs a complete chain of evidence from result to cause, clarifying whether the contradiction stems from the incorrect entry of the original data or from a logical fallacy in a certain reasoning rule.

[0063] Starting with the original facts, using the rules as intermediate nodes, and the conflicting facts as the endpoint, these are connected in the logical space to form a clear causal evolution path. This path not only shows how the contradiction is deduced step by step, but also indicates at which reasoning stage a logical fork occurs, thus concretizing the abstract logical conflict into a business logic chain that can be understood and examined by humans.

[0064] The process involves statistically analyzing the number and types of high-weight nodes along the tracing path to calculate a quantified severity score. This score reflects the potential disruptive power or scope of impact that a logical conflict could cause to the business if it were to be executed. If the calculated severity exceeds a preset safety threshold, it indicates that the logical contradiction is an intolerable hard error, and a formal judgment of logical conflict in the structured business information will be made, triggering an alarm or halting the process. If the severity is below the threshold, it indicates that the contradiction may only be a minor inconsistency caused by untimely information updates, or a soft conflict that can be automatically covered by business rules. In this case, it is determined that there is no substantial logical conflict, and the business process is allowed to continue. This ensures business rigor while avoiding frequent false alarms caused by oversensitivity.

[0065] For example, in a healthcare scenario, the extracted drug name entity is combined with the purchase intent to form structured business information, i.e., the patient intends to buy amoxicillin. This information is then parsed into fact triples and matched in the rule engine. It is found that the drug is a prescription drug and requires a prescription to purchase. However, there is no prescription information in the patient's history. A contradictory fact pair is deduced. After tracing the source, it is found that the conflict stems from the incompatibility between the intent and the drug attribute rules. After calculating the severity, it is determined that there is a logical conflict.

[0066] For example, in a fintech scenario, the financial entity, the intention to seek financial advice, and the risk level are combined into structured business information. It is found that the user's risk tolerance is conservative, but the risk level of the intended financial product is aggressive. The two constitute a contradictory fact pair. After tracing the source, it is confirmed that the conflict stems from the mismatch between the user's own attributes and the product attributes. It is calculated that the conflict involves fund security and has a high weight, and the severity exceeds the threshold, so it is determined that there is a logical conflict.

[0067] In this embodiment of the invention, by converting heterogeneous entities and intentions into a unified structured representation, the computer can escape the semantic noise of the original modality and perform efficient processing and reasoning in a standardized form, reducing the risk of information ambiguity and inconsistency. Secondly, the introduction of a logical reasoning mechanism can automatically detect contradictions or anomalies hidden in the structured information, preventing the system from executing business operations based on erroneous or contradictory information. This prevents invalid processing, data anomalies, or security risks caused by logical errors, thereby improving the accuracy of business processing.

[0068] S6. When there is a logical conflict in the structured business information, a semantic error correction prompt is generated for the structured business information. Based on the semantic error correction prompt, the structured business information is corrected to obtain structured business information without logical conflicts.

[0069] In this embodiment of the invention, the contradiction tracing path and the specific conflict facts detected in the previous step are used as input. A pre-trained natural language generation model is used to transform these structured logical conflict information into human-readable natural language descriptions. The model automatically organizes language templates according to the type of conflict, clearly indicating where the contradiction occurs and what the specific content of the contradiction is. For example, it may indicate that the precondition of a certain rule is incompatible with another fact, thereby generating a semantically clear and clearly targeted error correction prompt text, providing intuitive guidance for subsequent automatic or manual error correction.

[0070] Specifically, the system analyzes the conflict points indicated in the error correction prompts, locates the specific fact triples or business rules that cause logical contradictions, and tries various correction strategies, including deleting redundant facts that cause conflicts, modifying attribute values ​​that are incompatible with other facts, or adjusting the triggering conditions of rules. The system uses a backtracking search algorithm to verify the corrected information space, ensuring that each modification does not introduce new logical conflicts. After repeated iterations until all fact triples satisfy the logical consistency constraint, the system finally outputs a cleaned and corrected, logically self-consistent structured business information, thereby ensuring that downstream business processes can be executed securely based on correct and contradictory data.

[0071] In this embodiment of the invention, semantic error correction prompts transform abstract logical conflicts into natural language or structured feedback that users or the system can understand, lowering the threshold for conflict localization and understanding, and significantly improving the transparency and explainability of the error correction process. Secondly, prompt-based automatic or semi-automatic error correction can perform targeted corrections based on the conflict type, avoiding business process stagnation caused by direct discarding or error interruption, achieving smooth recovery and continuous execution in conflict scenarios, and improving overall processing efficiency and task completion success rate.

[0072] S7. When there is no logical conflict in the structured business information, extract the target decision field from the structured business information and calculate the confidence level of the target decision field.

[0073] In this embodiment of the invention, calculating the confidence level of the target decision field includes: Obtain the field type and field value of the target decision field; Based on the field type, the corresponding basic confidence calculation model is matched from the preset confidence calculation model library; Calculate the basic confidence score of the field value based on the basic confidence calculation model; Obtain the field hierarchy position of the target decision field in the preset decision tree, and calculate the hierarchy decay coefficient based on the field hierarchy position; The confidence score of the target decision field is obtained by multiplying the base confidence score by the hierarchical decay coefficient.

[0074] In this embodiment of the invention, the metadata information of the target field is read by accessing the structured definition of the decision result object. The field type indicates whether the field belongs to the data category such as text, numeric, enumeration, or boolean, while the field value is the specific content assigned to the field in the current decision result. These two pieces of information together constitute the basic input for subsequent confidence calculation. By identifying the field type, the calculation strategy to be adopted is determined. At the same time, by obtaining the field value, the quality and reliability of the value itself are judged, providing the necessary raw materials for subsequent graded confidence assessment.

[0075] Specifically, a set of models trained for different field types is maintained. Each type of field has a dedicated model adapted to its data characteristics. For example, numerical fields may use an evaluation model based on statistical distribution deviation, while text fields may use an evaluation model based on semantic integrity. The field type is used as an index key to search and match in the model library. The current field type is compared with the pre-registered type labels in the model library to accurately route to the basic confidence calculation model most suitable for processing the data of that type, ensuring that subsequent calculations can make full use of the inherent characteristics of the data of that type.

[0076] In detail, the field value is fed into the selected base model as input. The model evaluates the value based on its inherent attributes. For example, for text values, the model checks the character completeness and semantic fluency. For numerical values, the model checks whether they fall within a reasonable business range. After internal feature extraction and non-linear mapping, the model outputs a base score that reflects the reliability of the value itself.

[0077] Furthermore, the hierarchy depth of the field in the decision tree structure is located, which is the path length required from the root node to the field node. A preset decay function is used to map the hierarchy depth to a decay coefficient between zero and one. The design logic of this function is that the deeper the hierarchy of the field, the smaller its influence on the final decision result, so its confidence should be appropriately discounted. The shallower the hierarchy and the closer the field is to the root node, the greater its influence, and the closer the decay coefficient is to one. In this way, the system quantifies the corrective effect of the field's weight position in the decision system on its confidence.

[0078] Furthermore, by using the base confidence score as the original quality assessment and the hierarchical decay coefficient as the position weight modulation factor, the product of the two retains both the quality information of the value itself and incorporates the structural weight information of the field in the decision tree. This allows the final output confidence score to reflect a comprehensive consideration of both the quality of the value and the decision level. A field with high value quality but located deep in the decision tree will have its final confidence score appropriately lowered, while a core field with average value quality but located at the root of the decision tree will have its confidence score well preserved. This achieves a refined and differentiated assessment of the confidence score of decision fields.

[0079] In this embodiment of the invention, after passing the logical verification, key fields directly related to the final decision are further extracted. This effectively filters redundant information, reduces the computational burden on subsequent decision-making modules, and improves system processing efficiency. Secondly, the quantitative assessment of confidence level can reflect the credibility of the target decision field originating from the multimodal fusion and reasoning process, providing a reliable basis for uncertainty measurement in high-risk or sensitive business scenarios. Low-confidence fields can trigger manual review or supplementary data collection processes, while high-confidence fields can directly drive automated execution.

[0080] S8. Determine whether the confidence level is greater than or equal to the preset confidence threshold.

[0081] S9. When the confidence level is greater than or equal to the preset confidence threshold, the target business is automatically approved.

[0082] S10. When the confidence level is lower than the preset confidence threshold, the target business is transferred to manual review.

[0083] In this embodiment of the invention, a judgment standard as a safety red line is pre-set based on the business risk level and historical data statistics. During runtime, the confidence score calculated in real time is directly compared with this standard. When the confidence score is greater than or equal to the preset threshold, the target business will be automatically approved. That is, the approval instruction in the business process engine is automatically triggered, the status field of the business is updated to approved using workflow-driven technology, and subsequent execution instructions are automatically generated and sent to the downstream business module. The whole process does not require any manual operation, which greatly improves the processing speed and system throughput of high-confidence businesses.

[0084] When the system determines that the confidence level is lower than the preset threshold, it uses task routing and distribution technology to suspend the business and separate it from the automated pipeline. It then generates a manual review work order containing all relevant decision-making basis, confidence analysis report and potential risk points, and pushes it to the manual review terminal for professional personnel to conduct a second review, thereby effectively preventing business losses caused by model misjudgment.

[0085] For example, in a healthcare scenario, the confidence score for the decision field of a patient's medication purchase intention is calculated, provided that the drug name is clearly identified and the patient's historical medication purchase records match the current intention. Figure 1 If the basic confidence level is high and the medication purchase intention is located at the root node of the decision tree with a decay coefficient close to one, and the final confidence level exceeds the preset threshold, the application is automatically approved and an electronic prescription is directly generated and transferred to the pharmacy for dispensing. However, if the medication requested by the patient conflicts with their existing allergy history, the confidence level is significantly reduced to below the threshold due to hierarchical decay and factual contradictions. In this case, the medication purchase application is transferred to manual review, where a pharmacist verifies the allergy information before deciding whether to approve it.

[0086] For example, in a fintech scenario, if the risk level matching field corresponding to a user's financial advice intent has a reasonable value and is consistent with the user profile, and the calculated confidence level is higher than the threshold, the application is automatically approved and financial product recommendation information is pushed to the user. However, if the investment amount entered by the user far exceeds the range that their income statement can support, the decision field will have an abnormal value, causing the basic confidence level to drop sharply. Even if the intent is accurately identified, the final confidence level will still be lower than the threshold. Therefore, the financial application will be transferred to the manual review process, where a risk control specialist will conduct a second verification of the match between the income statement and the investment amount.

[0087] In this embodiment of the invention, the automated traffic triage mechanism based on confidence threshold can quantify and differentiate decision risk levels: high-confidence transactions are automatically approved by the system, which can significantly improve processing efficiency, reduce the cost of manual intervention, and achieve second-level response for large-scale transactions; while low-confidence transactions are promptly transferred to manual review, avoiding potential errors or risks caused by blind machine decision-making, enhancing the overall robustness and fault tolerance of the system, and improving the consistency of business processing and user satisfaction.

[0088] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0089] like Figure 3 The diagram shown is a functional block diagram of a business intelligent auxiliary processing device provided in an embodiment of the present invention.

[0090] In this embodiment of the disclosure, a business intelligent auxiliary processing device is provided, which corresponds one-to-one with the business intelligent auxiliary processing method of the above embodiments. For example... Figure 3 As shown, this intelligent business auxiliary processing device 100 can be installed in an electronic device. According to its functions, the intelligent business auxiliary processing device 100 includes a modality type determination module 101, a modality feature extraction module 102, an operation intent recognition module 103, a logical reasoning judgment module 104, a confidence calculation module 105, a first business processing module 106, and a second business processing module 107. Detailed descriptions of each functional module are as follows: The modality type determination module 101 is used to acquire multiple modality business data of the target user for processing the target business, and to determine the modality type to which each modality business data belongs; The modal feature extraction module 102 is used to extract the initial modal features corresponding to the modal service data based on each modal type, and to perform self-attention fusion of multiple initial modal features using a preset self-attention mechanism to obtain the target modal features; The operation intent recognition module 103 is used to recognize the operation intent of the target user based on the target modal features, and to extract business entities from the modal business data; The logical reasoning and judgment module 104 is used to structurally combine the business entity and the operation intention into structured business information, perform logical reasoning on the structured business information, and determine whether there is a logical conflict in the structured business information based on the result of the logical reasoning. The confidence calculation module 105 is used to extract the target decision field from the structured business information and calculate the confidence of the target decision field when there is no logical conflict in the structured business information. Confidence determination module 106 is used to determine whether the confidence level is greater than or equal to a preset confidence threshold. The first business processing module 107 is used to automatically review and approve the target business when the confidence level is greater than or equal to a preset confidence threshold. The second business processing module 108 is used to transfer the target business to manual review when the confidence level is lower than a preset confidence threshold.

[0091] In one embodiment, the modality type determination module 101, when determining the modality type to which various modality service data belongs, is used for: Obtain the data format identifier of each piece of business data from the various modal business data; The data format identifier is matched with a preset modality type mapping table; For successfully matched business data, the modality of the business data is determined according to the corresponding modality in the modality mapping table; For business data that fails to match, a preset modality recognition model is invoked to classify and predict the business data, thereby obtaining the predicted modality type and prediction confidence of the business data. When the prediction confidence level is greater than or equal to a preset prediction confidence level threshold, the prediction modality type is determined as the modality type of the business data.

[0092] In one embodiment, when the modal feature extraction module 102 performs the initial modal feature extraction based on each modal type to extract the corresponding modal service data, it is used to: Based on the various modal types, extract image data and text data from the modal business data; Extract the grid features and object features from the image data, and embed and stitch the grid features and object features to obtain the image features of the image data; Calculate the word embedding feature sequence of the text data, and perform feature encoding on the word embedding feature sequence to obtain the text features of the text data; The image features and text features are combined to form the initial modal features of the modal business data.

[0093] In one embodiment, when the modal feature extraction module 102 performs self-attention fusion of multiple initial modal features using a preset self-attention mechanism to obtain target modal features, it is used to: Each initial modal feature is converted into a feature vector of the same dimension to obtain a sequence of feature vectors; Calculate the query vector, key vector, and value vector for each feature vector in the feature vector sequence; Calculate the attention weight between the two corresponding feature vectors based on the query vector and the key vector; The value vector and the attention weight are weighted and summed to obtain the weighted fusion vector corresponding to the feature vector. The weighted fusion vectors corresponding to all the feature vectors are combined to generate a preliminary fusion feature matrix. The preliminary fused feature matrix is ​​nonlinearly mapped to obtain the target modal features.

[0094] In one embodiment, when the operation intent recognition module 103 performs the operation intent recognition based on the target modal features of the target user, it is used to: The target modal features are classified using a preset intent classification model to obtain the probability distribution of each intent category in the preset intent category set. Select the intent category with the highest probability value from the probability distribution as the initial intent to be identified; Obtain the target user's historical operation sequence, and extract the target user's historical intent tags within a preset time window from the historical operation sequence; Calculate the prior probability of intent transfer based on the historical intent tags; The preliminary identified intent and the prior probability of intent transfer are fused using Bayesian methods to obtain the posterior probability of each intent category; The intent category with the highest posterior probability is taken as the target recognition intent, and the recognition confidence corresponding to the target recognition intent is calculated. When the recognition confidence level is greater than or equal to the preset recognition confidence level threshold, the target recognition intent is taken as the target user's operation intent.

[0095] In one embodiment, when performing logical reasoning on the structured business information, the logical reasoning judgment module 104 is used to: The structured business information is parsed into a set of fact triples; Obtain the rule prerequisites for each business logic rule in the preset rule engine, perform pattern matching between each fact triple in the fact triple set and the rule prerequisites, and obtain a set of business logic rules that successfully match the rule prerequisites. Generate corresponding reasoning conclusion facts based on each business logic rule in the set of business logic rules, and add the reasoning conclusion facts to the set of fact triples to form an extended fact set; Detect whether there are contradictory fact pairs in the extended fact set; If there are contradictory fact pairs in the extended fact set, then the fact triples and business logic rules on which the fact pairs are based are traced back to generate them respectively. Generate a contradiction tracing path for the fact pair based on the fact triples and business logic rules; The severity of the logical conflict is calculated based on the aforementioned contradiction tracing path; If the severity of the conflict is greater than or equal to a preset conflict determination threshold, then the structured business information is determined to have a logical conflict. If the severity of the conflict is less than a preset conflict determination threshold, then the structured business information is determined to have no logical conflict.

[0096] In one embodiment, when calculating the confidence level of the target decision field, the confidence calculation module 105 is used to: Obtain the field type and field value of the target decision field; Based on the field type, the corresponding basic confidence calculation model is matched from the preset confidence calculation model library; Calculate the basic confidence score of the field value based on the basic confidence calculation model; Obtain the field hierarchy position of the target decision field in the preset decision tree, and calculate the hierarchy decay coefficient based on the field hierarchy position; The confidence score of the target decision field is obtained by multiplying the base confidence score by the hierarchical decay coefficient.

[0097] In this invention, the specific limitations of a business intelligent auxiliary processing device can be found in the above-described limitations of the business intelligent auxiliary processing method, and will not be repeated here. Each module in the aforementioned business intelligent auxiliary processing device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0098] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 4As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile and / or volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external clients via a network connection. When the computer program is executed by the processor, it implements the functions or steps of the business intelligence auxiliary processing method on the server side.

[0099] In one embodiment, a computer device is provided, which may be a client, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with an external server via a network connection. When the computer program is executed by the processor, it implements the functions or steps of the business intelligence auxiliary processing method on the client side.

[0100] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps: Obtain various modal business data used by the target user to process the target business, and determine the modal type to which each modal business data belongs; Initial modal features corresponding to each modal type are extracted from the service data of that modality, and multiple initial modal features are fused by self-attention using a preset self-attention mechanism to obtain target modal features; Based on the target modal features, the target user's operational intent is identified, and business entities are extracted from the modal business data; The business entity and the operation intention are structurally combined into structured business information, and logical reasoning is performed on the structured business information. Based on the result of the logical reasoning, it is determined whether there is a logical conflict in the structured business information. When there is no logical conflict in the structured business information, the target decision field in the structured business information is extracted, and the confidence level of the target decision field is calculated. Determine whether the confidence level is greater than or equal to a preset confidence threshold; When the confidence level is greater than or equal to a preset confidence threshold, the target business is automatically approved. When the confidence level is lower than the preset confidence threshold, the target business will be transferred to manual review.

[0101] In the several embodiments provided by this invention, it should be understood that the disclosed devices and apparatuses can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0102] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0103] Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0104] In some embodiments of this example, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the method described in the above embodiments.

[0105] The readable storage medium of the present invention stores a computer program, which, when executed by a processor of an electronic device, can perform the following: Obtain various modal business data used by the target user to process the target business, and determine the modal type to which each modal business data belongs; Initial modal features corresponding to each modal type are extracted from the service data of that modality, and multiple initial modal features are fused by self-attention using a preset self-attention mechanism to obtain target modal features; Based on the target modal features, the target user's operational intent is identified, and business entities are extracted from the modal business data; The business entity and the operation intention are structurally combined into structured business information, and logical reasoning is performed on the structured business information. Based on the result of the logical reasoning, it is determined whether there is a logical conflict in the structured business information. When there is no logical conflict in the structured business information, the target decision field in the structured business information is extracted, and the confidence level of the target decision field is calculated. Determine whether the confidence level is greater than or equal to a preset confidence threshold; When the confidence level is greater than or equal to a preset confidence threshold, the target business is automatically approved. When the confidence level is lower than the preset confidence threshold, the target business will be transferred to manual review.

[0106] It should be noted that the functions or steps that can be implemented by the computer-readable storage medium or computer device described above can be referred to the relevant descriptions on the server side and client side in the foregoing method embodiments. To avoid repetition, they will not be described one by one here.

[0107] Computer-readable storage media may also store at least one computer-executable program / instruction, such as computer-readable instructions. Computer-readable storage media include, but are not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and / or cache memory. Computer-readable storage media may include, for example, read-only memory (ROM), hard disk, flash memory, etc. For example, a non-transitory computer-readable storage medium may be connected to a computing device such as a computer, and then, when the computing device executes the computer-readable instructions stored on the computer-readable storage medium, the various methods described above can be performed.

[0108] In addition, the computer device may include (but is not limited to) a data bus, an input / output (I / O) bus, a display, and input / output devices (e.g., keyboard, mouse, speakers, etc.).

[0109] In one embodiment, the at least one computer-executable instruction may also be compiled into or comprise a software product / computer program product, wherein one or more computer-executable instructions are executed by a processor to perform the steps of the various functions and / or methods in the embodiments described herein.

[0110] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Furthermore, any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory.

[0111] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0112] In the embodiments provided in this disclosure, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, 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 marked in the blocks may occur in a different order than those marked in the drawings. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0113] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

[0114] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

Claims

1. A business intelligent auxiliary processing method, characterized in that, The method includes: Obtain various modal business data used by the target user to process the target business, and determine the modal type to which each modal business data belongs; Initial modal features corresponding to each modal type are extracted from the service data of that modality, and multiple initial modal features are fused by self-attention using a preset self-attention mechanism to obtain target modal features; Based on the target modal features, the target user's operational intent is identified, and business entities are extracted from the modal business data; The business entity and the operation intention are structurally combined into structured business information, and logical reasoning is performed on the structured business information. Based on the result of the logical reasoning, it is determined whether there is a logical conflict in the structured business information. When there is no logical conflict in the structured business information, the target decision field in the structured business information is extracted, and the confidence level of the target decision field is calculated. Determine whether the confidence level is greater than or equal to a preset confidence threshold; When the confidence level is greater than or equal to the preset confidence threshold, the target business is automatically approved. When the confidence level is lower than the preset confidence threshold, the target business will be transferred to manual review.

2. The business intelligent auxiliary processing method as described in claim 1, characterized in that, The process of determining the modality type to which various modal service data belong includes: Obtain the data format identifier of each piece of business data from the various modal business data; The data format identifier is matched with a preset modality type mapping table; For successfully matched business data, the modality of the business data is determined according to the corresponding modality in the modality mapping table; For business data that fails to match, a preset modality recognition model is invoked to classify and predict the business data, thereby obtaining the predicted modality type and prediction confidence of the business data. When the prediction confidence level is greater than or equal to a preset prediction confidence level threshold, the prediction modality type is determined as the modality type of the business data.

3. The business intelligent auxiliary processing method as described in claim 1, characterized in that, The step of extracting initial modal features corresponding to each modal service data based on each modal type includes: Extract image and text data from the modal business data based on the various modal types; Extract the grid features and object features from the image data, and embed and stitch the grid features and object features to obtain the image features of the image data; Calculate the word embedding feature sequence of the text data, and perform feature encoding on the word embedding feature sequence to obtain the text features of the text data; The image features and text features are combined to form the initial modal features of the modal business data.

4. The business intelligent auxiliary processing method as described in claim 1, characterized in that, The step of using a preset self-attention mechanism to perform self-attention fusion on multiple initial modal features to obtain target modal features includes: Each initial modal feature is converted into a feature vector of the same dimension to obtain a sequence of feature vectors; Calculate the query vector, key vector, and value vector for each feature vector in the feature vector sequence; Calculate the attention weight between the two corresponding feature vectors based on the query vector and the key vector; The value vector and the attention weight are weighted and summed to obtain the weighted fusion vector corresponding to the feature vector. The weighted fusion vectors corresponding to all the feature vectors are combined to generate a preliminary fusion feature matrix. The preliminary fused feature matrix is ​​nonlinearly mapped to obtain the target modal features.

5. The business intelligent auxiliary processing method as described in claim 1, characterized in that, The step of identifying the target user's operational intent based on the target modal features includes: The target modal features are classified using a preset intent classification model to obtain the probability distribution of each intent category in the preset intent category set. Select the intent category with the highest probability value from the probability distribution as the initial intent to be identified; Obtain the target user's historical operation sequence, and extract the target user's historical intent tags within a preset time window from the historical operation sequence; Calculate the prior probability of intent transfer based on the historical intent tags; The preliminary identified intent and the prior probability of intent transfer are fused using Bayesian methods to obtain the posterior probability of each intent category; The intent category with the highest posterior probability is taken as the target recognition intent, and the recognition confidence corresponding to the target recognition intent is calculated. When the recognition confidence level is greater than or equal to the preset recognition confidence level threshold, the target recognition intent is taken as the target user's operation intent.

6. The business intelligent auxiliary processing method as described in claim 1, characterized in that, The logical reasoning of the structured business information includes: The structured business information is parsed into a set of fact triples; Obtain the rule prerequisites for each business logic rule in the preset rule engine, perform pattern matching between each fact triple in the fact triple set and the rule prerequisites, and obtain a set of business logic rules that successfully match the rule prerequisites. Generate corresponding reasoning conclusion facts based on each business logic rule in the set of business logic rules, and add the reasoning conclusion facts to the set of fact triples to form an extended fact set; Detect whether there are contradictory fact pairs in the extended fact set; If there are contradictory fact pairs in the extended fact set, then the fact triples and business logic rules on which the fact pairs are based are traced back to generate them respectively. Generate a contradiction tracing path for the fact pair based on the fact triples and business logic rules; The severity of the logical conflict is calculated based on the aforementioned contradiction tracing path; If the severity of the conflict is greater than or equal to a preset conflict determination threshold, then the structured business information is determined to have a logical conflict. If the severity of the conflict is less than a preset conflict determination threshold, then the structured business information is determined to have no logical conflict.

7. The business intelligent auxiliary processing method as described in claim 1, characterized in that, The calculation of the confidence level of the target decision field includes: Obtain the field type and field value of the target decision field; Based on the field type, the corresponding basic confidence calculation model is matched from the preset confidence calculation model library; Calculate the basic confidence score of the field value based on the basic confidence calculation model; Obtain the field hierarchy position of the target decision field in the preset decision tree, and calculate the hierarchy decay coefficient based on the field hierarchy position; The confidence score of the target decision field is obtained by multiplying the base confidence score by the hierarchical decay coefficient.

8. A business intelligent auxiliary processing device, characterized in that, The device includes: The modality type determination module is used to acquire multiple modal business data used by the target user to process the target business, and to determine the modality type to which each modal business data belongs. The modal feature extraction module is used to extract the initial modal features corresponding to the modal service data based on each modal type, and to perform self-attention fusion of multiple initial modal features using a preset self-attention mechanism to obtain the target modal features; An operation intent recognition module is used to recognize the operation intent of the target user based on the target modal features, and to extract business entities from the modal business data; The logical reasoning and judgment module is used to structurally combine the business entity and the operation intention into structured business information, perform logical reasoning on the structured business information, and determine whether there is a logical conflict in the structured business information based on the result of the logical reasoning. The confidence calculation module is used to extract the target decision field from the structured business information and calculate the confidence of the target decision field when there is no logical conflict in the structured business information. The confidence level determination module is used to determine whether the confidence level is greater than or equal to a preset confidence level threshold; The first business processing module is used to automatically review and approve the target business when the confidence level is greater than or equal to a preset confidence threshold. The second business processing module is used to transfer the target business to manual review when the confidence level is lower than the preset confidence level threshold.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the business intelligence auxiliary processing method as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the business intelligent auxiliary processing method as described in any one of claims 1 to 7.

Images