Office process optimization method and device based on artificial intelligence, equipment and medium

Abstract

The invention relates to an office process optimization method and device based on artificial intelligence, equipment and a medium. The method comprises the steps of performing cross-modal feature alignment processing according to multi-modal office data to generate a structured data tuple; performing intention classification processing through a preset Transform model, generating and obtaining a node attribute set by querying a preset office process knowledge graph according to an intention classification label, and analyzing office process rule constraints in the set to generate a process node instruction; according to the instruction and the real-time service context, rule matching is carried out through a predefined rule, and when a matching result is that matching is passed, a normal task package is generated; and when the matching is not passed, generating and carrying out conflict analysis processing according to the exception description, and generating a manual intervention request. According to the method, through multi-modal data fusion, dynamic rule matching, systematized conflict analysis and the like, the multi-modal office data processing efficiency and the process decision accuracy are improved, and the adaptability of the office process to a real-time business scene is enhanced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computers, and particularly relates to an office process optimization method and device based on artificial intelligence, equipment and a medium. BACKGROUND

[0002] With the deepening of enterprise digital transformation, data in office scenarios presents significant multi-modal characteristics, covering not only office text data such as approval documents, task allocation emails, meeting minutes, but also office image data such as scanned expense receipts, form screenshots, and contract scans. However, traditional office processes highly rely on manual intervention, which not only leads to long data processing period and high labor cost, but also easily introduces information omission or misjudgment due to manual operation errors.

[0003] To improve the inefficiency of manual processing, although some studies attempt to introduce data processing tools and automation means, there are still multi-link technical defects. On the one hand, existing office process optimization technologies mostly use simple splicing or single modal independent processing and then summarizing, lack cross-modal feature alignment mechanisms optimized for office scenarios, cannot establish semantic association and feature matching between different types of data through cross-modal attention mechanisms, and are difficult to generate structured data tuples with unified structure and complete information, resulting in lack of reliable data basis for subsequent process decision-making. On the other hand, in the dynamic rule matching link, the technology often performs static checking based on fixed predefined rules, without flexible adjustment in combination with real-time business context such as current office time period, business priority, and dynamic configuration of department authority, leading to mismatch between matching results and actual business scenarios. SUMMARY

[0004] Therefore, it is necessary to provide an office process optimization method, device, equipment and medium based on artificial intelligence to improve the integration accuracy and processing efficiency of multi-modal office data, enhance the adaptability of dynamic rule matching to real-time business scenarios, and improve the accuracy of process decision-making.

[0005] In a first aspect, the application provides an office process optimization method based on artificial intelligence, comprising:

[0006] According to the multi-modal office data, cross-modal feature alignment processing is performed through a cross-modal attention mechanism to generate a structured data tuple, and the multi-modal office data includes office text data, office image data, and office voice data;

[0007] Based on the structured data tuple, an intent classification processing is performed through a preset Transformer model to generate an intent classification label, a node attribute set is obtained by querying a preset office process knowledge graph according to the intent classification label, and office process rule constraints in the node attribute set are parsed to generate a process node instruction;

[0008] According to the flow node instruction and the real-time business context, a dynamic rule matching process is performed through a predefined rule to obtain a matching result, and when the matching result is a matching pass, a normal task package is generated, wherein the normal task package is used to instruct an API gateway to call an RPA robot to perform an office automation operation and generate a flow log.

[0009] When the matching result is a matching fail, an artificial intervention request is generated and conflict resolution processing is performed through an OWL ontology inference machine according to an exception description, and the artificial intervention request includes an office problem description, a conflict type and an associated flow node ID, and the artificial intervention request is used to trigger a man-machine collaborative decision-making process.

[0010] In one of the embodiments, according to the multi-modal office data, cross-modal feature alignment processing is performed through a cross-modal attention mechanism to generate a structured data tuple, including:

[0011] According to the office text data, semantic analysis processing is performed through a BERT entity recognition model to extract office entities and entity relationships, and a text entity set is generated, and the office text data includes approval documents, task allocation emails and meeting minutes;

[0012] According to the office image data, office target recognition processing is performed through a YOLOv5 target detection model to locate office key target regions and generate a visual feature vector, and based on the coordinate information of the office key target region, character extraction processing is performed on the office key target region through an OCR recognition model to obtain recognized text, and the office image data includes scanned expense reimbursement receipts, form screenshots and contract scans;

[0013] According to the office voice data, voice waveform feature extraction and speech-to-text processing are performed through a Wav2Vec2 speech processing model to generate a speech transcription text, and the office voice data includes conference task recordings, voice approval instructions and customer service communication recordings;

[0014] Based on the text entity set, the visual feature vector, the recognized text and the speech transcription text, feature association and fusion processing are performed through a cross-modal attention mechanism, and a structured data tuple is generated through PCA dimensionality reduction processing according to a cross-modal similarity matrix.

[0015] In one of the embodiments, based on the structured data tuple, intent classification processing is performed through a preset Transformer model, and a node attribute set is obtained through a query to a preset office process knowledge graph according to an intent classification label, and office process rule constraints in the node attribute set are analyzed to generate a flow node instruction, including:

[0016] Based on structured data tuples, intent classification is performed using a pre-defined Transformer model to generate intent classification labels. The pre-defined Transformer model includes an encoding layer and a classification layer. The encoding layer encodes the intent of the structured data tuples to generate intent embedding vectors. The classification layer is a fully connected layer that performs classification mapping on the intent embedding vectors to obtain the probability distribution of each category and generate intent classification labels.

[0017] The knowledge graph of the preset office process is searched based on the intent classification tags to obtain the set of node attributes that match the intent classification tags. The set of node attributes includes node permission rules, task time constraints and data format requirements.

[0018] The Drools rule parser is used to parse the node attribute set using Boolean logic expressions to extract the office workflow rule constraints.

[0019] Based on the mandatory field requirements constrained by the office process rules, core parameters are extracted from structured data tuples. The core parameters include information about the task subject, execution time, and task target.

[0020] The system integrates office workflow rules and constraints with core parameters and encapsulates them into workflow node instructions in JSON format. The workflow node instructions include rule identifier fields, core parameter fields, and execution requirement fields.

[0021] In one embodiment, dynamic rule matching is performed based on process node instructions and real-time business context using predefined rules to obtain a matching result. When the matching result is successful, a normal task package is generated. This normal task package instructs the API gateway to call the RPA robot for office automation operations, generating process logs, including:

[0022] Obtain the real-time business context, which includes the current office hours, business priority, and departmental permission configuration.

[0023] Based on the operation instructions and constraints of the process node instructions, combined with the real-time business context, the logical consistency between the execution parameters in the process node instructions and the constraints of the real-time business context is verified through predefined rules. If the verification is successful, the matching result is a successful match; otherwise, the matching result is a failed match. The predefined rules include parameter threshold sub-rules, permission mapping sub-rules, and timeliness adaptation sub-rules.

[0024] When the matching result is a successful match, a normal task package is generated according to the execution parameters in the process node instruction. The execution parameters include the operation object and the execution steps.

[0025] Based on the normal task package, the RPA robot is called through the API gateway to perform office automation operations and obtain the execution results. The office automation operations include automatic form filling, cross-system data synchronization, and approval status push.

[0026] Based on the execution timestamps and results of normal task packages and office automation operations, process logs are generated. The process logs include a unique log ID, task association ID, operation details, result data, and status identifier.

[0027] In one embodiment, after generating the normal task package, the method further includes:

[0028] Based on the process log, extract the time series data of the office process sorted by the operation execution timestamp, and divide the time series data of the office process by a sliding window to obtain multiple sets of continuous log data windows;

[0029] The statistical distribution deviation between two adjacent log data windows is calculated using the KS test algorithm. The statistical distribution deviation is then compared with a significance threshold. If the statistical distribution deviation is greater than the significance threshold, it is determined that there is process drift; otherwise, it is determined that the process is stable. The comparison results are obtained.

[0030] For log data windows where the comparison results indicate process drift, abnormal indicators are extracted based on historical statistical data of normal office process operation to obtain abnormal log features. Abnormal log features include execution time deviation, parameter conflict frequency, approval node timeout number and data synchronization failure rate.

[0031] The abnormal log features are matched with the node attributes in the preset office process knowledge graph and mapped to the corresponding node identifier to generate a drift report. The drift report includes the drift node ID, drift type and offset. The node attributes include node function, operation parameter range and associated nodes. The node identifier is a unique code for each node in the preset office process knowledge graph.

[0032] According to the drift report, the office business process model was reconstructed using the alpha algorithm of the Celonis process mining engine to identify the node relationships to be adjusted. The office business process model is a visual process model built based on a preset office process knowledge graph.

[0033] Based on the node relationships to be adjusted, the edge weights and constraints of the corresponding node identifiers in the preset office process knowledge graph are updated using the gradient descent algorithm, generating the updated preset office process knowledge graph.

[0034] In one embodiment, after generating a request for human intervention, the method further includes:

[0035] Based on manual intervention requests, office risks are predicted through a risk prediction model, and an office process risk score is output. The risk prediction model is built based on historical risk data of office processes and a logistic regression algorithm.

[0036] Based on the risk score of office processes, the human-machine execution weight ratio is obtained by mapping through the Sigmoid function.

[0037] Based on the conflict type in the manual intervention request, query the node attributes corresponding to the conflict type in the preset office process knowledge graph, extract the corresponding solutions, and generate AI decision suggestions that include conflict resolution steps, implementing entities, and time requirements.

[0038] After sending a request for manual intervention to the management terminal, receive manual office decision-making instructions sent by the management terminal;

[0039] The AI ​​decision suggestions and human decision instructions are weighted and voted on according to the human-machine execution weight ratio. The decision support degree is calculated, the final decision of the office process is generated, and the final decision of the office process is sent to the management terminal.

[0040] Receive and generate binary feedback tags based on the confirmation results of the final decision on the office process from the management terminal;

[0041] The difference between the intent classification output probability distribution of the preset Transformer model and the binary feedback label is calculated using the cross-entropy loss function to generate the loss value.

[0042] Based on the loss value, the weight matrix of the preset Transformer model is updated using the gradient descent backpropagation algorithm, and the update magnitude of the weight matrix is ​​constrained by L2 regularization to generate an optimized Transformer model.

[0043] In one embodiment, the decision support is calculated using the following formula:

[0044] D final =γ·D AI ·C AI +(1-γ)·D human ·C human

[0045] Among them, D final For decision support, D AI The binary code for AI decision recommendations, where 1 indicates approval and 0 indicates rejection. AI The confidence level of AI decision-making is determined by pre-setting the maximum value of the output probability distribution of the Transformer model, D. human This is the binary encoding of the manual decision-making instruction, where 1 indicates approval and 0 indicates rejection. humanγ represents the credibility of human decision-making, and γ represents the weighting ratio of human execution.

[0046] Secondly, this application also provides an artificial intelligence-based office workflow optimization device, including:

[0047] The multimodal data fusion module is used to perform cross-modal feature alignment processing based on multimodal office data through a cross-modal attention mechanism to generate structured data tuples. Multimodal office data includes office text data, office image data, and office voice data.

[0048] The intent classification and rule parsing module is used to classify intents based on structured data tuples through a preset Transformer model, generate intent classification labels, obtain node attribute sets by querying a preset office process knowledge graph, parse the office process rule constraints in the node attribute sets, and generate process node instructions.

[0049] The dynamic rule matching module is used to perform dynamic rule matching based on process node instructions and real-time business context, and obtain matching results. When the matching result is successful, a normal task package is generated. The normal task package is used to instruct the RPA robot to perform office automation operations through the API gateway and generate process logs.

[0050] The conflict resolution and human intervention module is used to generate and process conflict resolution through the OWL ontology inference engine when the matching result is a match failure. The human intervention request includes a description of the office problem, the conflict type, and the ID of the associated process node. The human intervention request is used to trigger the human-machine collaborative decision-making process.

[0051] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the first aspect.

[0052] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the first aspect.

[0053] The aforementioned AI-based office workflow optimization methods, devices, equipment, and media align features of multimodal office data through a cross-modal attention mechanism, generating structured data tuples and establishing semantic associations and feature matching between different types of office data. This provides a data foundation for subsequent office intent recognition and workflow decision-making. Secondly, a pre-defined Transformer model is used for intent classification to generate intent classification labels. By querying a pre-defined office workflow knowledge graph to obtain node attribute sets and parsing office workflow rule constraints, workflow node instructions are generated. This accurately captures the core intent of office tasks, ensuring a high degree of alignment between workflow node instructions and actual office business needs. Combining workflow node instructions with real-time business context, dynamic rule matching is performed using predefined rules to obtain matching results. When a match is successful, a normal task package is generated, instructing the user to perform automated office operations, thus replacing repetitive manual labor. This not only reduces operational errors but also shortens the office workflow processing cycle. When a match fails, an anomaly description is generated, and conflict resolution is performed using the OWL ontology inference engine to generate a request for human intervention. This further systematically locates the root cause of the conflict and provides targeted solutions, triggering a human-machine collaborative decision-making process. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 A flowchart of an artificial intelligence-based office workflow optimization method provided as an exemplary embodiment of the present invention;

[0056] Figure 2 A flowchart of a method for generating structured data tuples is provided as an exemplary embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of an artificial intelligence-based office workflow optimization device provided as an exemplary embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0059] In one embodiment, such as Figure 1As shown, an AI-based office workflow optimization method is provided. This embodiment illustrates the method's application to a terminal, but it is understood that the method can also be applied to a server, or to a system including both a terminal and a server, and implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0060] S101: Based on multimodal office data, cross-modal feature alignment is performed through a cross-modal attention mechanism to generate structured data tuples. Multimodal office data includes office text data, office image data, and office voice data.

[0061] Specifically, cross-modal attention mechanisms can simulate the human logic of attention allocation to multi-source information. By calculating the semantic similarity between features of different modalities, feature-level alignment and fusion are achieved. Office text data can include unstructured text such as approval documents, task assignment emails, and meeting minutes; office image data can include visual data such as scanned expense receipts, screenshots of business forms, and scanned contracts; and office voice data can include audio data such as meeting recordings, voice approval instructions, and customer service communication recordings. Subsequently, text feature vectors, visual feature vectors, and voice feature vectors can be extracted from the above data respectively. By calculating the attention weights among the three types of feature vectors—for example, using the text feature vector as a baseline, calculating the semantic correlation between the image, voice, and text feature vectors using the Softmax function to obtain the attention weight matrix—and then weighting and summing the three types of feature vectors based on this matrix, cross-modal feature alignment can be achieved, generating structured data tuples with unified dimensions and complete semantics.

[0062] S102: Based on structured data tuples, intent classification is performed through a preset Transformer model to generate and classify intent labels. The set of node attributes is obtained by querying a preset office process knowledge graph, and the office process rule constraints in the set of node attributes are parsed to generate process node instructions.

[0063] Specifically, the pre-defined office workflow knowledge graph stores prior knowledge such as node attributes and rule constraints of enterprise office processes. The Transformer model can be pre-trained based on structured data containing historical office tasks and corresponding intent labels in the office workflow intent dataset, thus possessing the ability to accurately understand the semantics of office scenarios. By inputting the generated structured data tuples into the pre-defined Transformer model, an intent classification probability distribution can be output, and the category with the highest probability is taken as the intent classification label. Subsequently, using the intent classification label as the query condition, the set of nodes matching the intent classification label is obtained from the pre-defined office workflow knowledge graph, forming a node attribute set. By parsing the office workflow rule constraints in the node attribute set—that is, converting natural language rules into machine-recognizable logical expressions—and combining them with the task parameters in the intent classification label, the workflow node instructions can be obtained.

[0064] S103: Based on the process node instructions and real-time business context, perform dynamic rule matching processing through predefined rules to obtain the matching result. When the matching result is successful, generate a normal task package. The normal task package is used to instruct the RPA robot to perform office automation operations through the API gateway and generate process logs.

[0065] Specifically, process node instructions define the basic rules for process execution. However, in real-world office scenarios, dynamic changes in business status, such as current time period, task priority, and departmental permission changes, may cause fixed instructions to fail to meet actual needs. Therefore, dynamic rule matching can be performed based on process node instructions and real-time business context using predefined rules. These predefined rules can integrate logic such as parameter thresholds, permission mapping, and timeliness adaptation to construct a corresponding matching mechanism. For example, the execution parameters in the process node instructions can be compared with the real-time business context from multiple dimensions to verify their logical consistency. When the verification is successful, a normal task package is generated to instruct the process node instructions to be converted into standardized execution scripts that can be recognized by the RPA robot, enabling automated operations across systems without human intervention, such as automatically filling out expense forms, synchronizing customer data across multiple systems, and pushing approval status to relevant personnel. Furthermore, process logs can be generated in real time during the operation, recording information such as task execution time, operation content, and result status, providing a basis for subsequent process auditing.

[0066] S104: When the matching result is a failure, generate and, based on the exception description, perform conflict resolution through the OWL ontology inference engine to generate a manual intervention request. The manual intervention request includes a description of the office problem, the conflict type, and the associated process node ID. The manual intervention request is used to trigger the human-machine collaborative decision-making process.

[0067] Specifically, a failure to match dynamic rules indicates a conflict between the execution parameters of the process node instructions and the real-time business context. If this conflict were handled automatically by the system, it could lead to violations or process blockage. Therefore, a manual intervention request can be generated to trigger a human-machine collaborative decision-making process. The anomaly description can accurately record the specific content of the conflict. The OWL ontology inference engine can perform logical reasoning on the anomaly description based on a pre-defined ontology rule base, such as identifying the core elements of the conflict, matching the corresponding conflict type, and locating the specific process node where the conflict occurred. The subsequently generated manual intervention request integrates a detailed description of the office problem, the conflict type, and the associated process node ID, ensuring that the person in charge can quickly locate the problem and understand the nature of the conflict.

[0068] The aforementioned method comprehensively integrates office information from different modalities by collecting multimodal office data and utilizing cross-modal attention mechanisms for feature alignment. Secondly, it classifies intents using a pre-defined Transformer model, generating and labeling intents accordingly to accurately identify user intents. It can also generate corresponding process node instructions based on a pre-defined office workflow knowledge graph, providing clear guidance for subsequent automated operations. Furthermore, it performs dynamic rule matching based on process node instructions and real-time business context using predefined rules, ensuring that automated office workflows adapt to constantly changing business needs. When a match is successful, a normal task package is generated, instructing the API gateway to call the RPA robot for office automation, significantly improving efficiency and reducing repetitive work. Additionally, when a match fails, an exception is generated and, based on the exception description, a conflict resolution process using the OWL ontology inference engine generates a request for human intervention. This allows for timely human intervention when the automated workflow encounters unsolvable problems, ensuring the smooth operation of the workflow.

[0069] In one embodiment, such as Figure 2 As shown, based on multimodal office data, cross-modal feature alignment is performed using a cross-modal attention mechanism to generate structured data tuples, including:

[0070] S201: Based on office text data, semantic parsing is performed using the BERT entity recognition model to extract office entities and entity relationships, and generate a set of text entities. Office text data includes approval documents, task assignment emails, and meeting minutes.

[0071] S202: Based on office image data, office target recognition processing is performed using the YOLOv5 target detection model to locate key office target areas and generate visual feature vectors. Based on the coordinate information of key office target areas, character extraction processing is performed on key office target areas using an OCR recognition model to obtain recognized text. Office image data includes scanned expense receipts, form screenshots, and scanned contract copies.

[0072] S203: Based on office voice data, the Wav2Vec2 voice processing model is used to extract voice waveform features and perform speech-to-text processing to generate speech-transcribed text. Office voice data includes meeting task recordings, voice approval instructions, and customer service communication recordings.

[0073] S204: Based on text entity sets, visual feature vectors, recognized text and speech-transcribed text, feature association and fusion processing are performed through cross-modal attention mechanism, and structured data tuples are generated by calculating and performing dimensionality reduction processing through PCA based on the cross-modal similarity matrix.

[0074] Specifically, for office text data, the BERT entity recognition model is used for semantic analysis. This BERT entity recognition model can be obtained by fine-tuning a pre-trained BERT model using office domain text data. The fine-tuning dataset can contain multiple annotated office texts, labeled with core office entities such as task subject, execution time, and approval authority. Illustratively, the model can segment the text data into word units, capture contextual semantic information through the encoding layer, and then use a CRF (Conditional Random Field) classification layer to predict entity labels for each word unit. Finally, it extracts office entities and entity relationships in a uniformly annotated format, forming a text entity set.

[0075] For office image data, the YOLOv5 object detection model can be used first. The image is input into the backbone network to extract multi-scale features, and shallow positional features and deep semantic features are fused through the neck network. Then, the bounding box coordinates and target feature vector of the office target are output to the detection head. This target feature vector is used as the visual feature vector. Subsequently, based on the coordinate information of the key office target area obtained by localization, the corresponding area in the image can be cropped into sub-images through coordinate mapping. These sub-images are then input into an OCR recognition model such as the Textract model based on convolutional neural networks and recurrent neural networks. The convolutional neural network can extract the local texture features of the sub-images, and the recurrent neural network can capture the temporal dependencies of the character sequence. The CTC (Connectionist Temporal Classification) algorithm can be combined to solve the problem of the mismatch between the length of the input feature sequence and the output character sequence, and finally, the recognized text corresponding to the target area is obtained. For processing office speech data, the Wav2Vec2 speech processing model can be adopted, such as the pre-trained Wav2Vec2-Large model. First, the speech waveform data is processed through convolutional layers to extract Mel-spectral features. Then, a Transformer encoder captures the temporal dependencies of the speech sequence, and the model's self-supervised learning mechanism can be utilized. Subsequently, in the speech-to-text processing stage, the CTC decoding algorithm can be used to map the feature sequence output by the encoder into a text sequence, generating speech-transcribed text, while retaining the speech waveform features as the basis for subsequent speech feature fusion.

[0076] Finally, the text entity set can be converted into text feature vectors using a pre-trained Word2Vec model. Similarly, the recognized text and the speech-transcribed text can be converted into semantic feature vectors using the same Word2Vec model. The semantic vectors of the recognized text are used to enhance the semantic expression of the image modality, while the semantic vectors of the speech-transcribed text are used to convey the semantic information of the speech. Visual feature vectors can be mapped to maintain consistency with the feature dimensions of other modalities through fully connected layers. Furthermore, the speech waveform features output by Wav2Vec2 can be mapped to the same feature dimensions through fully connected layers as supplementary features for the speech modality. Subsequently, a cross-modal attention module is constructed based on a multi-head attention mechanism. Using text feature vectors as queries, and visual feature vectors, recognized text semantic vectors, speech-transcribed text semantic vectors, and speech waveform mapping features as keys and values, attention weights between each modality feature are calculated. Finally, the attention weights and values ​​are weighted and summed to obtain the cross-modal attention features. Based on this cross-modal attention feature, a cross-modal similarity matrix is ​​calculated. This matrix has a dimension of 4×4, with rows representing text, visual, image-text, and speech modalities, columns representing the core features of each modality, and matrix elements representing the cosine similarity between the features of two modalities. Based on this matrix, PCA dimensionality reduction is used to eliminate feature redundancy. Finally, the dimensionality-reduced cross-modal features are integrated with the core entity information of each modality to generate structured data tuples.

[0077] In one embodiment, based on structured data tuples, intent classification is performed using a preset Transformer model to generate and label intents. Then, a set of node attributes is obtained by querying a preset office workflow knowledge graph, and the office workflow rule constraints within the node attribute set are parsed to generate workflow node instructions, including:

[0078] Based on structured data tuples, intent classification is performed using a pre-defined Transformer model to generate intent classification labels. The pre-defined Transformer model includes an encoding layer and a classification layer. The encoding layer encodes the intent of the structured data tuples to generate intent embedding vectors. The classification layer is a fully connected layer that performs classification mapping on the intent embedding vectors to obtain the probability distribution of each category and generate intent classification labels.

[0079] The knowledge graph of the preset office process is searched based on the intent classification tags to obtain the set of node attributes that match the intent classification tags. The set of node attributes includes node permission rules, task time constraints and data format requirements.

[0080] The Drools rule parser is used to parse the node attribute set using Boolean logic expressions to extract the office workflow rule constraints.

[0081] Based on the mandatory field requirements constrained by the office process rules, core parameters are extracted from structured data tuples. The core parameters include information about the task subject, execution time, and task target.

[0082] The system integrates office workflow rules and constraints with core parameters and encapsulates them into workflow node instructions in JSON format. The workflow node instructions include rule identifier fields, core parameter fields, and execution requirement fields.

[0083] Specifically, the encoding layer can be composed of six stacked Transformer encoders. Each encoder layer includes a multi-head self-attention mechanism and a feedforward neural network. First, the cross-modal feature vectors from the structured data tuples are concatenated with the Word2Vec vectors of the core entities to form the input vector. When the input vector passes through the encoding layer, the multi-head self-attention mechanism calculates the association weights of each feature within the input vector to capture the key feature associations related to the intent. The feedforward neural network then performs a non-linear transformation on the attention output features to generate an intent embedding vector. This vector can accurately represent the business intent semantics corresponding to the structured data tuples. The classification layer is a two-layer fully connected network. The first fully connected layer introduces non-linear expressive power through the ReLU activation function, and the second fully connected layer converts the output into a probability distribution for each category through the Softmax function. The category corresponding to the maximum probability is taken as the intent classification label.

[0084] This is an illustrative example of a pre-defined office workflow knowledge graph, a graph-structured database built upon historical office workflow data and office rule documents. It can be stored using Neo4j graph database. In this graph structure, nodes represent key steps in the office workflow, each assigned a unique node ID, and edges represent the flow relationships between nodes. During retrieval, an intent-node mapping table is first constructed. Based on intent classification tags, a set of candidate nodes is determined from the mapping table. Then, a Cypher query is used to retrieve the complete attributes of the candidate nodes from the knowledge graph, obtaining the set of node attributes matching the intent classification tags. Subsequently, the Drools rule parser performs Boolean logic expression parsing on the node attribute set, converting the natural language-described node attributes into machine-recognizable structured rule constraints, providing a clear compliance basis for the generation of workflow node instructions. For example, natural language attributes can be converted into Boolean logic expressions conforming to Drools syntax. By inputting the preprocessed Boolean logic expressions into the Drools rule parser, a lexical analyzer can break the expressions down into multiple lexical units and verify their grammatical correctness based on pre-defined syntax rules. Finally, a rule constraint object that can be directly executed by the machine is generated, extracting the office workflow rule constraints. The required fields for office workflow rule constraints can be derived from the constraints in the rule constraint object, and these required fields can be further categorized into three core parameters: task subject, execution time, and task target information. For example, a mapping table between core parameters and structured data tuple fields can be constructed, and corresponding parameters can be extracted from the structured data tuples through field matching. Finally, office workflow rule constraints and core parameters can be integrated and encapsulated into workflow node instructions in JSON format.

[0085] In one embodiment, dynamic rule matching is performed based on process node instructions and real-time business context using predefined rules to obtain a matching result. When the matching result is successful, a normal task package is generated. This normal task package instructs the API gateway to call the RPA robot to perform office automation operations and generate process logs, including:

[0086] Obtain the real-time business context, which includes the current office hours, business priority, and departmental permission configuration.

[0087] Based on the operation instructions and constraints of the process node instructions, combined with the real-time business context, the logical consistency between the execution parameters in the process node instructions and the constraints of the real-time business context is verified through predefined rules. If the verification is successful, the matching result is a successful match; otherwise, the matching result is a failed match. The predefined rules include parameter threshold sub-rules, permission mapping sub-rules, and timeliness adaptation sub-rules.

[0088] When the matching result is a successful match, a normal task package is generated according to the execution parameters in the process node instruction. The execution parameters include the operation object and the execution steps.

[0089] Based on the normal task package, the RPA robot is called through the API gateway to perform office automation operations and obtain the execution results. The office automation operations include automatic form filling, cross-system data synchronization, and approval status push.

[0090] Based on the execution timestamps and results of normal task packages and office automation operations, process logs are generated. The process logs include a unique log ID, task association ID, operation details, result data, and status identifier.

[0091] Specifically, the real-time business context can serve as a dynamic constraint for the current office scenario, providing an environmental adaptation basis for the execution of process node instructions. Subsequently, dynamic rule matching can be performed based on the process node instructions and the real-time business context. For example, parameter threshold sub-rules are used to verify whether the value or format of the execution parameters conforms to the threshold range of the real-time business; permission mapping sub-rules are used to verify whether the department permissions of the operating subject match the operation object of the process node instruction; and timeliness adaptation sub-rules are used to determine whether the operation execution time conforms to the real-time office time constraints. The verification logic can use an AND operation to combine the results of the three sub-rules. Only when all sub-rules pass verification is the matching result considered a successful match; otherwise, it is considered a failed match.

[0092] When the matching result is successful, a normal task package is generated based on the execution parameters in the process node instructions. This transforms abstract instructions into a structured task description that the RPA robot can directly execute. For example, the task package uses XML format and contains three core nodes: task header, operation sequence, and parameter set. The task header stores the task ID and priority; the operation sequence stores the execution steps in an ordered list, with each step containing the operation type, target control path, and operation parameters; and the parameter set stores global variables such as the current user ID.

[0093] Specifically, to invoke RPA robots to perform office automation operations based on normal task packages, task distribution and operation coordination need to be achieved through an API gateway. These office automation operations can include three high-frequency operations in office scenarios: automatic form filling, cross-system data synchronization, and approval status push notifications. The API gateway, acting as an intermediary layer, receives normal task packages and routes tasks to idle RPA robot nodes using a load balancing algorithm. After parsing the task package, the RPA robot uses screen scraping technology to locate the UI controls of the target system and executes automated operations according to the operation sequence. For example, when automatically filling out forms, the robot uses optical character recognition to assist in locating fuzzy controls, simulates text input via the keyboard, and simulates clicking the submit button via the mouse. During the operation, the RPA robot provides real-time feedback on the execution status to the API gateway and finally returns structured data containing the operation results. Finally, based on the normal task package, execution timestamp, and execution results, a process log can be generated to provide a basis for subsequent process analysis and problem troubleshooting.

[0094] In one embodiment, after generating a normal task package, the method further includes:

[0095] Based on the process log, extract the time series data of the office process sorted by the operation execution timestamp, and divide the time series data of the office process by a sliding window to obtain multiple sets of continuous log data windows;

[0096] The statistical distribution deviation between two adjacent log data windows is calculated using the KS test algorithm. The statistical distribution deviation is then compared with a significance threshold. If the statistical distribution deviation is greater than the significance threshold, it is determined that there is process drift; otherwise, it is determined that the process is stable. The comparison results are obtained.

[0097] For log data windows where the comparison results indicate process drift, abnormal indicators are extracted based on historical statistical data of normal office process operation to obtain abnormal log features. Abnormal log features include execution time deviation, parameter conflict frequency, approval node timeout number and data synchronization failure rate.

[0098] The abnormal log features are matched with the node attributes in the preset office process knowledge graph and mapped to the corresponding node identifier to generate a drift report. The drift report includes the drift node ID, drift type and offset. The node attributes include node function, operation parameter range and associated nodes. The node identifier is a unique code for each node in the preset office process knowledge graph.

[0099] According to the drift report, the office business process model was reconstructed using the alpha algorithm of the Celonis process mining engine to identify the node relationships to be adjusted. The office business process model is a visual process model built based on a preset office process knowledge graph.

[0100] Based on the node relationships to be adjusted, the edge weights and constraints of the corresponding node identifiers in the preset office process knowledge graph are updated using the gradient descent algorithm, generating the updated preset office process knowledge graph.

[0101] Specifically, the process log contains a complete execution trajectory sorted by operation execution timestamps. This can be divided into multiple consecutive log data windows using a sliding window method to capture local features of the process at different time periods. For example, the length of the sliding window is set to a fixed value, and the window sliding step is 5 steps. Each window contains log fields such as operation details, execution duration, and result status within that time period. Secondly, the statistical distribution deviation between adjacent log data windows is calculated using the KS test algorithm to quantitatively assess the degree of variation of process features over time, determining whether significant drift exists. This statistical distribution deviation reflects the distributional differences of process features between adjacent windows. Furthermore, as a non-parametric test method, the KS test does not require assuming that the data follows a specific distribution; it can directly measure distributional differences by calculating the maximum difference between the cumulative distribution functions of two samples, making it suitable for diverse and non-standardized feature data in office processes. For example, for each pair of adjacent windows, key features such as execution duration and parameter conflict frequency are selected as test variables. The cumulative distribution function of these variables is calculated for each window, and the statistic is calculated using the KS test formula. A larger value indicates a more significant distributional difference between the two windows. The calculated statistic is compared with a preset significance threshold. If it is greater than the threshold, it can be determined that there is a significant difference in the statistical distribution between adjacent windows, that is, there is process drift.

[0102] Specifically, for log data windows identified as exhibiting process drift, abnormal indicators can be extracted based on historical statistical data to form abnormal log features, providing a characteristic basis for subsequent drift tracing. Historical statistical data on the normal operation of the office process can refer to log data from the stable period of the process over the past 6 months. By calculating the mean and standard deviation of each feature in this dataset, the normal fluctuation range can be determined. Feature values ​​exceeding this range can be identified as abnormal indicators, thus obtaining abnormal log features. Execution time deviation can be the difference between the current window's execution time and the historical mean; parameter conflict frequency is the number of times operation parameters within the window do not conform to rule constraints; and approval node timeout count is the number of records of approval nodes exceeding time limits within the window. Subsequently, a feature-attribute mapping algorithm can be used to semantically associate each indicator in the abnormal log features with node attributes, and the best-matching node identifier can be determined by calculating the Euclidean distance between the feature value and the attribute standard value. A drift report can be generated based on the matching results, where the drift node ID is the matched node identifier, the drift type can be determined based on the abnormal features, and the offset can be obtained by quantifying the degree of deviation.

[0103] Specifically, the office workflow model is a visualized directed graph built upon a pre-defined office workflow knowledge graph. Nodes represent workflow steps, and edges represent flow relationships. When workflow drift occurs, the actual flow paths between nodes may deviate from the model's preset path. The Celonis engine's alpha algorithm can analyze log data from the drift window, extract activity sequences, reconstruct the workflow model based on causal relationships between activities, and identify node relationships that need adjustment. For example, it can extract all activities (nodes) and direct follow relationships (edges) from the logs, identify starting and ending activities, and then construct a new workflow model based on the causal relationship matrix, outputting a list of node relationships including additions, deletions, or weight changes. Finally, based on the node relationships to be adjusted, the edge weights and constraints of the pre-defined office workflow knowledge graph are updated using a gradient descent algorithm, generating an updated pre-defined office workflow knowledge graph. This achieves dynamic optimization of the knowledge graph, ensuring its consistency with the actual workflow. The edge weights represent the probability of flow between nodes, and the constraints can include node attributes such as timeliness and parameters. For example, the node relationships in the reconstruction model can be used as the target value, and the current edge weights of the knowledge graph can be used as the initial values. The objective function can be minimized by the gradient descent algorithm, where the objective function is the squared difference between the current weights and the target weights. The edge weights can be updated iteratively, and the constraints can be adjusted according to the offset of the abnormal log features.

[0104] In one embodiment, after generating a request for human intervention, the method further includes:

[0105] Based on manual intervention requests, office risks are predicted through a risk prediction model, and an office process risk score is output. The risk prediction model is built based on historical risk data of office processes and a logistic regression algorithm.

[0106] Based on the risk score of office processes, the human-machine execution weight ratio is obtained by mapping through the Sigmoid function.

[0107] Based on the conflict type in the manual intervention request, query the node attributes corresponding to the conflict type in the preset office process knowledge graph, extract the corresponding solutions, and generate AI decision suggestions that include conflict resolution steps, implementing entities, and time requirements.

[0108] After sending a request for manual intervention to the management terminal, receive manual office decision-making instructions sent by the management terminal;

[0109] The AI ​​decision suggestions and human decision instructions are weighted and voted on according to the human-machine execution weight ratio. The decision support degree is calculated, the final decision of the office process is generated, and the final decision of the office process is sent to the management terminal.

[0110] Receive and generate binary feedback tags based on the confirmation results of the final decision on the office process from the management terminal;

[0111] The difference between the intent classification output probability distribution of the preset Transformer model and the binary feedback label is calculated using the cross-entropy loss function to generate the loss value.

[0112] Based on the loss value, the weight matrix of the preset Transformer model is updated using the gradient descent backpropagation algorithm, and the update magnitude of the weight matrix is ​​constrained by L2 regularization to generate an optimized Transformer model.

[0113] Specifically, the human intervention request includes key information such as the description of the office problem and the type of conflict. The risk prediction model can be based on a logistic regression algorithm. The training data uses historical risk data of the enterprise's office processes, including conflict types, scope of impact, and loss amounts, to obtain a risk prediction model with conflict features as input and risk probabilities as output. By inputting the conflict features from the human intervention request into this model, an office process risk score can be obtained. The office process risk score is mapped to the human-machine execution weight ratio using a Sigmoid function, which can dynamically allocate the decision weights of AI and human based on the risk level. For example, the risk score can be normalized by x = (risk score - 50) / 10, mapping the score to the (-5, 5) interval, enhancing the function's sensitivity to the intermediate score range. Then, the AI ​​weight is calculated by substituting it into the Sigmoid function. That is, when the risk score is ≤30, the score x is -2, and the AI ​​weight σ(x) ≈ 0.12, that is, the AI ​​weight is 12%, and the human weight is 88%, prioritizing rapid human decision-making. Then, the conflict-node mapping table of the knowledge graph can be retrieved by using conflict type keywords, the corresponding node can be located, the Cypher query statement can be called to extract the solution, and then it can be packaged into AI decision suggestions according to the preset template.

[0114] After sending the request for human intervention to the management terminal and receiving the human decision-making instruction, a weighted vote is conducted based on the human-machine execution weight ratio to generate the final decision. For illustration, in this weighted voting process, if both the AI ​​decision suggestion and the human decision instruction agree to grant special approval or reject the application, a consistent final decision can be directly generated. If the two opinions conflict, the decision support degree is calculated using the following formula:

[0115] D final =γ·D AI ·C AI +(1-γ)·D human ·C human

[0116] Among them, D final For decision support, D AI The binary code for AI decision recommendations, where 1 indicates approval and 0 indicates rejection. AIThe confidence level of AI decision-making is determined by pre-setting the maximum value of the output probability distribution of the Transformer model, D. human This is the binary encoding of the manual decision-making instruction, where 1 indicates approval and 0 indicates rejection. human γ represents the credibility of human decision-making, and γ represents the weighting ratio of human execution.

[0117] Indicatively, a final decision can be made if the support level is ≥0.5, otherwise it can be rejected. The final decision in the generated office workflow can include the decision result, weighting criteria, conflict resolution criteria such as referencing budget rules from a knowledge graph or special notes from manual annotations, and is pushed to the management terminal for confirmation. The management terminal confirms the final decision, which can be either approved or rejected, corresponding to binary feedback labels of 1 and 0. Furthermore, if there is a discrepancy between the initial intent classification label output by the preset Transformer model (e.g., regular expense reimbursement approval) and the actual intent corresponding to the final decision (e.g., special approval for over-budget reimbursement), this label can serve as a correction signal to guide model optimization, ensuring that the feedback data accurately corresponds to the model's intent classification. For example, the difference between the model output and the feedback label can be calculated using the cross-entropy loss function to generate a loss value. The intent classification output of the preset Transformer model is a probability distribution for each category.

[0118] Specifically, a stochastic gradient descent algorithm can be used, with the gradient of the calculated loss value with respect to the weight matrix as the direction, iteratively updating the weights at a learning rate of 0.001. Furthermore, L2 regularization can be applied, which involves adding a squared term to the weights in the loss function to constrain the absolute growth of the weight matrix and prevent the model from over-relying on local features. For example, when the model's classification probability of the intention to approve reimbursements exceeding the budget is low, gradient descent can increase the weight of the corresponding feature, such as the amount exceeding the budget, while L2 regularization prevents this weight from increasing indefinitely. After multiple rounds of iterative optimization, an optimized Transformer model can be generated.

[0119] Based on the same inventive concept, such as Figure 3 As shown, this application also provides an AI-based office workflow optimization device 300 for implementing the AI-based office workflow optimization method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more AI-based office workflow optimization device embodiments provided below can be found in the limitations of the method embodiments above, and will not be repeated here. The device includes:

[0120] The multimodal data fusion module 301 is used to perform cross-modal feature alignment processing based on multimodal office data through a cross-modal attention mechanism to generate structured data tuples. The multimodal office data includes office text data, office image data, and office voice data.

[0121] The intent classification and rule parsing module 302 is used to classify intents based on structured data tuples through a preset Transformer model, generate intent classification labels, obtain node attribute sets by querying a preset office process knowledge graph, and parse the office process rule constraints in the node attribute sets to generate process node instructions.

[0122] The dynamic rule matching module 303 is used to perform dynamic rule matching based on process node instructions and real-time business context through predefined rules to obtain matching results. When the matching result is successful, a normal task package is generated. The normal task package is used to instruct the RPA robot to perform office automation operations through the API gateway and generate process logs.

[0123] The conflict resolution and human intervention module 304 is used to generate and process conflict resolution through the OWL ontology inference engine when the matching result is a match failure. The human intervention request includes a description of the office problem, the conflict type, and the ID of the associated process node. The human intervention request is used to trigger the human-machine collaborative decision-making process.

[0124] In one exemplary embodiment, the present invention also provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the AI-based office workflow optimization method of this application. A multi-core processor is preferred to improve the system's parallel processing capabilities. The memory provides sufficient temporary storage space to support program execution and data processing. The memory capacity should be large enough to accommodate a large amount of information and computational tasks.

[0125] In one exemplary embodiment, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the artificial intelligence-based office workflow optimization method of this application. The computer-readable storage medium may include: read-only memory (ROM), random access memory (RAM), solid-state drives (SSDs), or optical discs, etc. The random access memory may include resistive random access memory (ReRAM) and dynamic random access memory (DRAM).

[0126] The above-described embodiments are merely illustrative of several implementation methods of the embodiments of this application, and their descriptions are relatively specific and detailed. However, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the embodiments of this application, and these modifications and improvements all fall within the protection scope of the embodiments of this application.

Claims

1. An artificial intelligence-based method for optimizing office processes, characterized in that, The method includes: Based on multimodal office data, cross-modal feature alignment is performed through a cross-modal attention mechanism to generate structured data tuples. The multimodal office data includes office text data, office image data, and office voice data. Based on the structured data tuples, intent classification is performed using a preset Transformer model to generate and classify intent labels. Then, a set of node attributes is obtained by querying a preset office process knowledge graph, and the office process rule constraints in the set of node attributes are parsed to generate process node instructions. Based on the process node instructions and real-time business context, dynamic rule matching is performed using predefined rules to obtain a matching result. When the matching result is a successful match, a normal task package is generated. The normal task package is used to instruct the RPA robot to perform office automation operations through the API gateway and generate process logs. When the matching result is a failure, a conflict resolution process is performed using the OWL ontology inference engine based on the exception description to generate a manual intervention request. The manual intervention request includes a description of the office problem, the conflict type, and the associated process node ID. The manual intervention request is used to trigger a human-machine collaborative decision-making process.

2. The method according to claim 1, characterized in that, The process of generating structured data tuples by performing cross-modal feature alignment processing based on multimodal office data through a cross-modal attention mechanism includes: Based on the office text data, semantic parsing is performed using the BERT entity recognition model to extract office entities and entity relationships, generating a text entity set. The office text data includes approval documents, task assignment emails, and meeting minutes. Based on the office image data, the YOLOv5 target detection model is used to perform office target recognition processing to locate key office target areas and generate visual feature vectors. Based on the coordinate information of the key office target areas, the OCR recognition model is used to extract characters from the key office target areas to obtain recognized text. The office image data includes scanned expense receipts, form screenshots, and scanned contract copies. Based on the office voice data, speech waveform features are extracted and speech-to-text is processed using the Wav2Vec2 speech processing model to generate speech-transcribed text. The office voice data includes meeting task recordings, voice approval instructions, and customer service communication recordings. Based on the text entity set, the visual feature vector, the recognized text, and the speech-transcribed text, feature association and fusion processing are performed through a cross-modal attention mechanism. The cross-modal similarity matrix is ​​calculated, and the structured data tuple is generated through PCA dimensionality reduction processing.

3. The method according to claim 1, characterized in that, Based on the structured data tuples, intent classification processing is performed using a preset Transformer model to generate and, according to intent classification labels, a set of node attributes is obtained by querying a preset office workflow knowledge graph. The office workflow rule constraints in the node attribute set are then parsed to generate workflow node instructions, including: Based on the structured data tuples, intent classification processing is performed through the preset Transformer model to generate intent classification labels. The preset Transformer model includes an encoding layer and a classification layer. The encoding layer is used to encode the intent of the structured data tuples to generate intent embedding vectors. The classification layer is a fully connected layer. The classification layer is used to perform classification mapping processing on the intent embedding vectors to obtain the probability distribution of each category and generate the intent classification labels. The preset office process knowledge graph is retrieved based on the intent classification tags to obtain a set of node attributes that match the intent classification tags. The set of node attributes includes node permission rules, task time constraints, and data format requirements. The Drools rule parser is used to parse the node attribute set using Boolean logic expressions to extract office workflow rule constraints. Based on the mandatory field requirements constrained by the office process rules, core parameters are extracted from the structured data tuples. The core parameters include information about the task subject, execution time, and task target. The office workflow rules and constraints are integrated with the core parameters and encapsulated into workflow node instructions in JSON format. The workflow node instructions include rule identifier fields, core parameter fields, and execution requirement fields.

4. The method according to claim 1, characterized in that, The process involves dynamic rule matching based on the process node instructions and real-time business context, using predefined rules to obtain a matching result. When the matching result is successful, a normal task package is generated. This normal task package instructs the API gateway to call the RPA robot for office automation operations, generating process logs, including: Obtain the real-time business context, which includes the current office hours, business priority, and departmental permission configuration; Based on the operation instructions and constraints of the process node instructions, and in conjunction with the real-time business context, the logical consistency between the execution parameters in the process node instructions and the constraints of the real-time business context is verified through the predefined rules. If the verification is successful, the matching result is considered successful; otherwise, the matching result is considered unsuccessful. The predefined rules include parameter threshold sub-rules, permission mapping sub-rules, and timeliness adaptation sub-rules. When the matching result is that the match is successful, the normal task package is generated according to the execution parameters in the process node instruction. The execution parameters include the operation object and the execution steps. Based on the normal task package, the RPA robot is invoked through the API gateway to perform office automation operations and obtain the execution results. The office automation operations include automatic form filling, cross-system data synchronization, and approval status push. Based on the normal task package, the execution timestamp of the office automation operation, and the execution result, the process log is generated. The process log includes a unique log ID, a task association ID, operation details, result data, and status identifier.

5. The method according to claim 1, characterized in that, After generating the normal task package, the method further includes: Based on the process log, extract the office process time series data sorted by operation execution timestamp, and divide the office process time series data by a sliding window to obtain multiple sets of continuous log data windows; The statistical distribution deviation between two adjacent log data windows is calculated using the KS test algorithm. The statistical distribution deviation is then compared with a significance threshold. If the statistical distribution deviation is greater than the significance threshold, it is determined that there is process drift; otherwise, it is determined that the process is stable, and the comparison results are obtained. For the log data window that is determined to have process drift based on the comparison results, abnormal indicators are extracted based on the historical statistical data of the normal operation of the office process to obtain abnormal log features. The abnormal log features include execution time deviation, parameter conflict frequency, approval node timeout number and data synchronization failure rate. The abnormal log features are matched with the node attributes in the preset office process knowledge graph and mapped to the corresponding node identifier to generate a drift report. The drift report includes the drift node ID, drift type and offset. The node attributes include node function, operation parameter range and associated nodes. The node identifier is a unique code for each node in the preset office process knowledge graph. Based on the drift report, the office business process model is reconstructed using the α algorithm of the Celonis process mining engine to identify the node relationships to be adjusted. The office business process model is a visualized process model built based on the preset office process knowledge graph. Based on the node associations to be adjusted, the edge weights and constraints corresponding to the node identifiers in the preset office process knowledge graph are updated using the gradient descent algorithm to generate the updated preset office process knowledge graph.

6. The method according to claim 1, characterized in that, After generating the manual intervention request, the method further includes: Based on the aforementioned request for human intervention, an office risk prediction model is used to predict office risks and output an office process risk score. The risk prediction model is constructed based on historical risk data of office processes and a logistic regression algorithm. Based on the aforementioned office process risk score, the human-machine execution weight ratio is obtained through Sigmoid function mapping. Based on the conflict type in the manual intervention request, query the node attributes corresponding to the conflict type in the preset office process knowledge graph, extract the corresponding solutions, and generate AI decision suggestions that include conflict resolution steps, implementing entities, and time requirements. After sending the request for manual intervention to the management terminal, the system receives manual office decision-making instructions from the management terminal. The AI ​​decision suggestions and the human decision instructions are weighted and voted on according to the human-machine execution weight ratio to calculate the decision support degree, generate the final decision of the office process, and send the final decision of the office process to the management terminal. Receive and generate a binary feedback tag based on the confirmation result of the final decision on the office process from the management terminal; The difference between the intent classification output probability distribution of the preset Transformer model and the binary feedback label is calculated using the cross-entropy loss function to generate a loss value; Based on the loss value, the weight matrix of the preset Transformer model is updated using the gradient descent backpropagation algorithm, and the update magnitude of the weight matrix is ​​constrained by L2 regularization to generate an optimized Transformer model.

7. The method according to claim 6, characterized in that, The decision support level is calculated using the following formula: D final =γ·D AI ·C AI +(1-c)·D human ·C human Among them, D final Let D be the decision support level. AI The binary encoding of the AI ​​decision suggestion, where 1 indicates approval and 0 indicates rejection, C AI The confidence level of the AI ​​decision is determined by the maximum value of the output probability distribution of the preset Transformer model, D. human The binary encoding of the manual decision-making instruction is 1 for approval and 0 for rejection. human γ represents the credibility of human decision-making, and γ represents the weight ratio of human execution.

8. An office workflow optimization device based on artificial intelligence, characterized in that, The device includes: The multimodal data fusion module is used to perform cross-modal feature alignment processing based on multimodal office data through a cross-modal attention mechanism to generate structured data tuples. The multimodal office data includes office text data, office image data, and office voice data. The intent classification and rule parsing module is used to classify intents based on the structured data tuples using a preset Transformer model, generate intent classification labels, obtain a set of node attributes by querying a preset office process knowledge graph, and parse the office process rule constraints in the set of node attributes to generate process node instructions. The dynamic rule matching module is used to perform dynamic rule matching processing based on the process node instructions and real-time business context through predefined rules to obtain the matching result. When the matching result is a successful match, a normal task package is generated. The normal task package is used to instruct the RPA robot to perform office automation operations through the API gateway and generate process logs. The conflict resolution and human intervention module is used to generate and, based on the anomaly description, perform conflict resolution processing through the OWL ontology inference engine when the matching result is a mismatch. The human intervention request includes a description of the office problem, the conflict type, and the associated process node ID. The human intervention request is used to trigger a human-machine collaborative decision-making process.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images