Data collaboration processing method and system of OA platform

Abstract

The application is OA platform data collaborative processing method and system, relates to office automation technology field, and includes: based on structured approval content data and preconfigured approval role list, respectively calling each approval role corresponding role exclusive examination model to carry out parallel examination processing on the structured approval content data.In the application, the collaborative decision arbitration layer carries out semantic alignment and consensus degree evaluation on the multi-role examination opinions, realizes the automatic detection on the examination opinion conflict; through the three-level conflict resolution strategy, the system can select the resolution strategy according to the conflict complexity, and reduces the dependence on artificial coordination.

Description

Technical Field

[0001] This invention relates to the field of office automation technology, and in particular to a data collaborative processing method and system for OA platforms. Background Technology

[0002] In the enterprise office automation (OA) approval process, approval materials typically include documents in multiple modalities such as text applications, image attachments, table data, and PDF contracts.

[0003] In existing technologies, there are methods that configure exclusive approval templates and intelligent review models for different approval roles to parse multimodal documents and generate structured review results, thereby assisting decision-making at individual approval nodes.

[0004] However, this type of solution focuses on intelligent assistance for a single approval node, with each node operating independently, which presents the following problems: In cross-departmental approval scenarios, there is a lack of coordination and integration among the review opinions of multiple approval roles such as legal, financial, and business. When the review conclusions of multiple roles conflict, the system cannot automatically identify and resolve the conflict, and manual coordination is still required. The review knowledge of the preceding approval nodes cannot be effectively transferred to subsequent nodes, resulting in a fragmentation of approval knowledge between nodes. Different nodes repeatedly review the same information, leading to insufficient overall approval consistency. Existing conflict resolution solutions either expose conflicts directly to approvers for manual adjudication or process them mechanically according to fixed rules, lacking the ability to resolve conflicts intelligently based on semantic understanding.

[0005] Therefore, a data collaborative processing method and system for an OA platform are proposed to address the aforementioned problems. Summary of the Invention

[0006] The purpose of this invention is to provide a data collaborative processing method and system for an OA platform in order to solve the above-mentioned problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: Data collaboration processing methods in OA platforms include: The system receives a set of multimodal files to be approved, parses and processes each file in the set, and generates structured approval content data. The set of multimodal files includes at least two of the following: text application documents, attachment documents, image files, and table files. Based on the structured approval content data and the pre-configured list of approval roles, the role-specific review models corresponding to each approval role are invoked to perform parallel review processing on the structured approval content data. Each role-specific review model outputs the structured review opinions for its corresponding role. The role-specific review model is a model obtained based on a large language model and configured with prompts through the approval template for the corresponding role. The structured review opinions include review conclusions, confidence scores, risk levels, and reasoning paths. The structured review opinions output by each role are input into the collaborative decision-making arbitration layer. The review opinions are semantically aligned to generate semantic vectors for each role. The consensus index among multiple roles is calculated based on the semantic similarity between the semantic vectors of each review opinion. The consensus index is compared with a preset consensus threshold, and the conflict detection result is output. Based on the conflict detection results, when the conflict detection result is no conflict, the structured review opinions of each role are comprehensively processed to synthesize the collaborative approval decision result; when the conflict detection result is that a conflict exists, a multi-level conflict resolution strategy is adopted to resolve the conflict in sequence, and the collaborative approval decision result is synthesized based on the resolution result; wherein, the multi-level conflict resolution strategy includes rule arbitration, semantic mediation and weighted voting in order of increasing complexity.

[0008] Preferably, after generating the structured approval content data, the process further includes: Based on the structured data of each file in the multimodal file collection, a cross-file association index is established. Specifically, the key information fields in the structured data of each file are semantically encoded to generate semantic vectors corresponding to each key information field. Cosine similarity is used to calculate the semantic correlation between key information fields in different files. Key information fields with cosine similarity higher than the association threshold are associated with each other to form a unified semantic representation of the approval content. The association threshold is a numerical parameter pre-configured according to the type of approval item.

[0009] Preferably, before the parallel review processing is performed using the individual role-specific review models, the method further includes: The key information fields in the structured approval content data are semantically encoded to generate query vectors. The top K historical approval cases with semantic similarity to the query vectors are retrieved from the vector database, where K is a preset positive integer parameter for the number of searches. The review opinions and reasoning paths of the retrieved K historical approval cases are used as reference context and concatenated into the prompts of the role-specific review model, so that the role-specific review model can refer to the judgment basis and decision logic of historical approvals during the reasoning process.

[0010] Preferably, the parallel review process further includes: Based on the current approval node's position in the approval process and the structured review opinions already output by the previous approval node, context-aware prompts are dynamically generated. Specifically, the review conclusions, risk levels, and reasoning paths of the previous approval nodes are extracted as context summaries, and these context summaries are embedded into the prompts of the role-specific review model corresponding to the current approval node. This allows the role-specific review model of the current node to perceive the current approval process stage and the background of previous decisions, avoiding duplicate reviews of content already reviewed by previous nodes.

[0011] Preferably, the consensus index is calculated as follows: the cosine similarity of the semantic vectors of the review opinions between each role pair is summed and then divided by the total number of role pairs to obtain the consensus index; wherein, the consensus threshold is a numerical parameter pre-configured according to the type of approval item, and the range of values ​​for both the consensus index and the consensus threshold is a closed interval from negative one to positive one. After determining that a conflict exists, the process also includes: splitting the structured review opinions of each role according to the review dimensions, calculating the semantic similarity of the role pairs on each review dimension, identifying the specific review dimensions whose semantic similarity is lower than the dimension conflict threshold, marking the review dimension as a conflict dimension, and transmitting the conflict dimension and the review opinions of the corresponding role to the conflict resolution unit.

[0012] Preferably, the rule arbitration includes: obtaining the priority weights of each conflicting role; sorting the review conclusions of the conflicting roles according to the priority weights; selecting the review conclusion of the role with the highest priority as the preliminary ruling result; calculating the semantic relevance between the preliminary ruling result and the review conclusions of the remaining conflicting roles using cosine similarity; if all cosine similarities are higher than the ruling acceptance threshold, the rule arbitration is deemed successful, and the preliminary ruling result is transmitted to the decision synthesis unit; if there is a case where the cosine similarity is lower than the ruling acceptance threshold, the semantic mediation is initiated. The semantic mediation process includes: taking the review opinions, reasoning paths, and conflict dimension information of each conflicting role as input, constructing mediation prompts, calling a large language model to analyze the root causes of the conflict and generating mediation suggestions that take into account the concerns of all parties; using cosine similarity to calculate the semantic relevance between the mediation suggestions and the review opinions of each conflicting role; if the cosine similarity between the mediation suggestions and the review opinions of each conflicting role is higher than the mediation acceptance threshold, the semantic mediation is deemed successful; if there is a case where the cosine similarity is lower than the mediation acceptance threshold, the process proceeds to the weighted voting.

[0013] Preferably, the weighted voting includes: Based on the historical review accuracy and current review opinion confidence score of each conflict role's role-specific review model, the comprehensive decision weight of each conflict role is calculated. The comprehensive decision weight is calculated as follows: the sum of the historical review accuracy of the i-th conflict role multiplied by a first weight adjustment coefficient and the confidence score of that conflict role multiplied by a second weight adjustment coefficient is divided by the sum of all conflict roles calculated in the same way. The comprehensive decision weight of the i-th conflict role is obtained by multiplying the historical review accuracy by a first weight adjustment coefficient and the second weight adjustment coefficient by one. The historical review accuracy and the confidence score both fall within a closed interval from zero to one. Based on the comprehensive decision-making weight of each conflicting role, the risk level values ​​corresponding to the review conclusions of each conflicting role are weighted and summed to obtain a weighted risk score. The weighted risk score is then compared with a preset risk level classification standard to determine the final comprehensive risk level and review conclusion, which serves as the weighted voting decision result.

[0014] Preferably, after receiving the set of multimodal documents to be approved, the process further includes: Based on the type identifier of the approval item and the key fields in the structured approval content data, the approval complexity score is calculated. The key fields include the approval amount, the number of departments involved, and the number of document modalities. The normalized values ​​of each of the three key fields are obtained by normalization based on the range. The approval complexity score is then obtained by weighted summation of the normalized values. The sum of the weight coefficients of each field is one. The complexity score of the approval is compared with a preset complexity grading threshold to determine the complexity level of the approval item; for approval items with a complexity level lower than the preset level, they are automatically processed according to the preset approval rules; for approval items with a complexity level not lower than the preset level, the collaborative approval process of the large language model is triggered.

[0015] Preferably, after outputting the collaborative approval decision result, the method further includes: The structured review opinions of each role in this approval process, the conflict resolution process records, and the collaborative approval decision results are semantically encoded to generate an approval case vector for this approval case. The approval case vector and the corresponding approval process data are stored in the vector database, and the approval item type identifier is stored synchronously. Receive feedback information from the approver regarding the collaborative approval decision result. The feedback information includes the approver's adoption indicator of the collaborative approval decision result, the modified content, and the reason for the modification. The feedback information is associated with and stored with the corresponding approval case vector. Based on the accumulated feedback information, the percentage of review suggestions from each role-specific review model that are adopted by the approver is statistically analyzed, and the historical review accuracy parameters of each role-specific review model are updated. When the accumulated feedback information reaches a preset data volume threshold, the modified content and reasons for modification in the feedback information are used as training corpus to perform supervised fine-tuning and updates on each role-specific review model.

[0016] The data collaboration processing methods and systems for OA platforms include: The multimodal file parsing module is used to receive a set of multimodal files to be approved, parse and process each file in the multimodal file set, and generate structured approval content data; The multi-role parallel review module is used to perform parallel review processing of the structured approval content data based on the structured approval content data and the pre-configured list of approval roles, by calling the role-specific review model corresponding to each approval role, and outputting the structured review opinions of each role. The conflict detection module is used to input the structured review opinions output by each role into the collaborative decision-making arbitration layer, perform semantic alignment processing on each review opinion, generate semantic vectors of review opinions corresponding to each role, calculate the consensus index among multiple roles based on the semantic similarity between the semantic vectors of each review opinion, compare the consensus index with the preset consensus threshold, and output the conflict detection result. The decision synthesis module is used to synthesize collaborative approval decision results based on the conflict detection results. When the conflict detection results are that there is no conflict, the module comprehensively processes the structured review opinions of each role and synthesizes collaborative approval decision results. When the conflict detection results are that there is a conflict, the module adopts a multi-level conflict resolution strategy to resolve the conflict in sequence and synthesizes collaborative approval decision results based on the resolution results.

[0017] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention achieves automatic detection of conflicting review opinions by performing semantic alignment and consensus assessment on multi-role review opinions through a collaborative decision-making arbitration layer; through a three-level conflict resolution strategy, the system can adaptively select a resolution strategy according to the complexity of the conflict, reducing the reliance on manual coordination.

[0018] 2. This invention embeds the review conclusions and reasoning paths of preceding approval nodes into the prompts of subsequent nodes, enabling subsequent approval nodes to perceive the background of preceding decisions and reducing repeated reviews of the same information; by jointly parsing multimodal files and establishing cross-file association indexes, the review models of each role can conduct reviews based on a unified semantic representation of the approval content. Attached Figure Description

[0019] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0020] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0021] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0022] Example 1

[0023] Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0024] In this embodiment, it includes: Step 1: Joint parsing and structured processing of multimodal approval documents The system receives a set of multimodal files awaiting approval, parses and processes each file in the set, and generates structured approval content data. The multimodal file set includes at least two of the following: text application documents, attachment documents, image files, and table files. The structured approval content data is represented in JSON format and contains the content fields of each file, modality type identifiers, and file association identifiers.

[0025] For text files in the multimodal file collection, paragraph structure and key information fields are extracted to generate corresponding structured text data. For table files in the multimodal file collection, row and column structure and data items are identified to generate corresponding structured table data. For image or scanned document files in the multimodal file collection, a visual-language joint model is used to recognize the file content, extract heading levels, paragraph structure, table data, and chart information, and generate corresponding structured data.

[0026] The aforementioned visual-language joint model is a pre-trained visual-language model (VLM). The input to the VLM is image data from image files or scanned documents, and the output is structured data containing document structure information and text content. The VLM uses pre-trained weights for inference and requires no additional training for this implementation.

[0027] To enable subsequent review processes to understand the inherent relationships between different modal documents, after generating the structured data for each document, the following steps are included: establishing a cross-document association index based on the structured data of each document in the multimodal document set. Specifically, key information fields in the structured data of each document are semantically encoded to generate semantic vectors corresponding to each key information field. Cosine similarity is used to calculate the semantic relevance between key information fields in different documents. Key information fields with cosine similarity higher than an association threshold are associated with each other, forming a unified semantic representation of the approval content. The association threshold is a pre-configured numerical parameter based on the type of approval item, with a value range of [insert range here]. This is used to control the strictness of cross-document association mappings. For example, establishing an association mapping between the amount clause in the contract text and the total price field in the attached quotation allows subsequent review processes to identify consistency between the two.

[0028] Step 2: Multi-role parallel review and review comment generation Based on structured approval content data and a pre-configured list of approval roles, the system invokes the role-specific review model corresponding to each approval role to perform parallel review of the structured approval content data. Each role-specific review model outputs its corresponding structured review opinion. The list of approval roles includes at least two from legal, financial, and business roles. Each role-specific review model is a model obtained based on a large language model and configured with prompts using the corresponding role's approval template. Each role-specific review model receives structured approval content data as input, performs step-by-step reasoning using a thought chain prompt strategy, and outputs the corresponding role's structured review opinion. The structured review opinion includes the review conclusion, confidence score, risk level, and reasoning path.

[0029] The aforementioned thought chain prompt strategy refers to pre-setting step-by-step reasoning instructions in the prompts of the role-specific review model. This allows the large language model to analyze each review dimension of the approval content sequentially according to predefined reasoning steps, gradually generating intermediate reasoning results, and finally summarizing them into structured review opinions. The reasoning path is the sequence of intermediate conclusions generated by the large language model in each reasoning step.

[0030] The confidence score mentioned above is a numerical value output by the role-specific review model based on the information completeness of the current approval content and the certainty of the reasoning process, with a value range of [missing value]. A higher confidence score indicates that the role-specific review model has a higher degree of certainty regarding the current review conclusion.

[0031] Each role-specific review model is based on a large language model. The input to each model is a concatenated sequence of structured approval content data and role-specific prompts. The output is a structured review opinion containing the review conclusion, confidence score, risk level, and reasoning path. Supervised fine-tuning of each model is employed, using review opinions from historical approval cases as the training target. The cross-entropy loss function is used, and the Adam optimization algorithm is employed for parameter updates.

[0032] To improve the reasoning quality of the review models specific to each role, the following steps are included before the step-by-step reasoning in each role-specific review model: semantically encoding the key information fields in the structured approval content data to generate query vectors, and retrieving the top-ranked query vectors from the vector database based on their semantic similarity. 10 historical approval cases, among which The preset number of searches is the parameter. It is a positive integer. The retrieved... The review opinions and reasoning paths of historical approval cases are used as reference context and pieced into the prompts of the role-specific review model, so that the role-specific review model can refer to the judgment basis and decision-making logic of historical approvals during the reasoning process.

[0033] Building upon step 2, to enable subsequent approval nodes to access the review knowledge of preceding approval nodes, the following steps are also included: dynamically generating context-aware prompts based on the current approval node's position in the approval process and the structured review opinions already output by preceding approval nodes. Specifically, the review conclusions, risk levels, and reasoning paths of preceding approval nodes are extracted as context summaries. These context summaries are then embedded into the prompts of the role-specific review model corresponding to the current approval node. This allows the role-specific review model of the current node to perceive the current approval process stage and the background of preceding decisions, avoiding redundant review of content already reviewed by preceding nodes.

[0034] Step 3: Semantic alignment and conflict detection of review comments The structured review opinions output by each role are input into the collaborative decision-making arbitration layer, where semantic alignment and conflict detection are performed on each review opinion, and the conflict detection results are output.

[0035] The collaborative decision-making arbitration layer comprises a semantic alignment unit, a consensus assessment unit, a conflict resolution unit, and a decision synthesis unit. The semantic alignment unit receives structured review opinions from each role, encodes these opinions into vector representations in a unified semantic vector space, and transmits these vector representations to the consensus assessment unit. The consensus assessment unit calculates a consensus index based on the vector representations, determines the existence of conflicts based on a comparison of the consensus index with a preset threshold, and transmits the conflict detection result to either the conflict resolution unit or the decision synthesis unit. When a conflict exists, the conflict resolution unit receives conflict-related review opinions, performs conflict resolution processing, and transmits the resolution result to the decision synthesis unit. The decision synthesis unit synthesizes a unified collaborative approval decision based on the structured review opinions from each role's specific review model and the conflict resolution result.

[0036] Step 3 includes the following sub-steps: Step 301: Using a semantic alignment unit, semantically encode the review conclusions in the structured review opinions of each role, generating semantic vectors for each role's review opinions. The semantic encoding utilizes the text embedding capability of a large language model, mapping the text content of the review conclusions to fixed-dimensional semantic vectors. Let the first... The semantic vector of the review comments for each role is The total number of characters is ,in For character indexing, from Take to , It is a positive integer.

[0037] Step 302: Calculate the semantic similarity between the semantic vectors of each review opinion through the consensus assessment unit, and calculate the consensus index among multiple roles based on the semantic similarity.

[0038] The semantic similarity between each pair of characters is calculated using cosine similarity, and the formula is as follows: ; in, and The first The first character and the first Semantic vector of review opinions for each role, The dot product of two vectors. and Let these be the magnitudes of the two vectors, and Index the characters.

[0039] The formula for calculating the consensus index among multiple roles is: ; in, As a consensus indicator, The total number of roles involved in the review. For the first The first role and the first Semantic similarity between characters and For the traversal index of the summation, from Traversal to , from Traversal to Consensus index The range of values ​​is , A higher value indicates greater consistency in the review opinions of all roles.

[0040] Step 303: Consensus metrics With the preset consensus threshold Comparison, among which .when When it is determined that there are no significant conflicts in the review opinions of each role, and the conflict detection result is no conflict, the structured review opinions of each role are directly transmitted to the decision synthesis unit. If significant conflicts are found between the review opinions of different roles, and the conflict detection result indicates that a conflict exists, the conflict resolution process is triggered.

[0041] The above consensus threshold Consensus threshold is a pre-configured numerical parameter based on the type of approval item. Value range and consensus index The range of values ​​is consistent, all being [value range]. Different types of approval items can be configured with different consensus thresholds. For example, approval items involving large contract amounts can be configured with higher consensus thresholds to ensure that the opinions of all roles are highly consistent.

[0042] Step 4: Conflict Resolution and Collaborative Approval Decision Synthesis Based on the conflict detection results, the structured review opinions of each role are processed for conflict resolution and decision synthesis, and the collaborative approval decision results are output.

[0043] When the conflict detection result is no conflict, the decision synthesis unit directly processes the structured review opinions of each role to synthesize the collaborative approval decision result. Specifically, the review conclusions, risk levels, and reasoning paths of each role are summarized, the highest risk level among the risk levels of each role is taken as the comprehensive risk level, the reasoning paths of each role are merged to generate a comprehensive reasoning report, and the collaborative approval decision result is output.

[0044] When the conflict detection result indicates that a conflict exists, the conflict resolution unit uses a multi-level conflict resolution strategy to process the conflict sequentially, specifically including the following sub-steps: Step 401: Perform Level 1 Rule Arbitration. Based on the preset role priority configuration and approval business rules, a preliminary ruling is made on the review opinions of conflicting roles. Specifically, the priority weight of each conflicting role is obtained, and the review conclusions of the conflicting roles are sorted according to the priority weight. The review conclusion of the role with the highest priority is selected as the preliminary ruling result. Cosine similarity is used to calculate the semantic relevance between the preliminary ruling result and the review conclusions of the remaining conflicting roles. If all cosine similarities are higher than the ruling acceptance threshold, the rule arbitration is deemed successful, and the preliminary ruling result is transmitted to the decision synthesis unit. If there is a case where the cosine similarity is lower than the ruling acceptance threshold, the rule arbitration is deemed to have failed to completely resolve the conflict, and the process proceeds to Level 2 Semantic Mediation. The ruling acceptance threshold is a pre-configured numerical parameter based on the type of approval matter. The range of the ruling acceptance threshold is consistent with the range of the cosine similarity, both being [value missing]. .

[0045] Step 402: Perform the second-level semantic mediation. Using the review opinions, reasoning paths, and conflict dimension information of each conflicting party as input, a large language model is invoked for semantic analysis to generate conflict cause analysis results and mediation suggestions. Specifically, mediation prompts are constructed, containing the review conclusions of both parties, their respective reasoning paths, and supporting evidence. These prompts instruct the large language model to analyze the root causes of the conflict (e.g., differences in judgment standards for the same risk or misunderstandings of approval materials), and generate mediation suggestions that take into account the concerns of all parties. Cosine similarity is used to calculate the semantic relevance between the mediation suggestions and the review opinions of each conflicting party. If the cosine similarity between the mediation suggestions and the review opinions of each conflicting party is higher than the mediation acceptance threshold, semantic mediation is considered successful, and the mediation suggestions are passed to the decision synthesis unit as the resolution result. If there are cases where the cosine similarity is lower than the mediation acceptance threshold, the third-level weighted voting is initiated. The mediation acceptance threshold is a pre-configured numerical parameter based on the type of approval item, and its range is consistent with the range of cosine similarity. .

[0046] Step 403: Perform the third-level weighted voting. Based on the historical review accuracy and current review opinion confidence scores of the role-specific review models for each conflicting role, calculate the comprehensive decision weight for each conflicting role, perform weighted voting on the review conclusions of each conflicting role, and generate a weighted voting decision result.

[0047] Among them, the accuracy rate of historical review For the first The role-specific review model for each conflicting role is based on the historical review accuracy rate obtained from historical approval data statistics, with a value range of [value missing]. Confidence score For the first The confidence score of the current review opinion for each conflicting role, with a value range of [value missing]. .because and All have been normalized to The intervals, having the same dimensions, can be directly weighted. Index for conflicting roles.

[0048] The formula for calculating the overall decision weight of each conflicting role is as follows: ; in, For the first The overall decision-making weight of each conflicting role. For the first Accuracy of historical review of conflicting roles, For the first The confidence score of the current review opinion for each conflicting role. The total number of conflicting roles participating in the weighted vote. and The preset weight adjustment coefficient, and satisfies , For the traversal index of the summation, from Traversal to .

[0049] The above weighting adjustment coefficient and This is used to balance the contribution of historical review accuracy and current confidence score to the overall decision weight. When At that time, the overall decision-making weight relies more on the long-term historical performance of each role; when In this context, the overall decision-making weighting relies more heavily on the certainty of each role's immediate assessment of the current approval content. Weighting adjustment coefficient. and The specific values ​​are pre-configured based on the type of approval item, and the range of values ​​is [missing information]. And satisfy .

[0050] Based on the comprehensive decision-making weights of each conflicting role, the risk level values ​​corresponding to the review conclusions of each conflicting role are weighted and summed to obtain a weighted risk score. The weighted risk score is compared with the preset risk level classification standard to determine the final comprehensive risk level and review conclusion, which is then transmitted to the decision synthesis unit as the weighted voting decision result.

[0051] The decision synthesis unit synthesizes the final collaborative approval decision based on the structured review opinions and conflict resolution results of each role. The collaborative approval decision includes a comprehensive review conclusion, a comprehensive risk level, a summary of key risk points, differentiated concerns and suggestions for each role, and a record of the conflict resolution process.

[0052] The aforementioned multi-level conflict resolution strategies are executed sequentially in ascending order of complexity. Level 1, rule arbitration, makes a rapid ruling based on pre-configured rules, suitable for conflict scenarios with clear differences in role priorities. Level 2, semantic mediation, invokes a large language model for deep semantic analysis, suitable for scenarios where the root cause of the conflict involves misunderstandings or differences in judgment standards. Level 3, weighted voting, makes a comprehensive decision based on quantitative indicators, suitable for complex conflict scenarios where the first two levels cannot completely resolve the conflict. During the conflict resolution process, once a strategy at a certain level successfully resolves the conflict, subsequent levels of strategy are not executed.

[0053] To automatically verify the compliance of collaborative approval decision results, after the decision synthesis unit outputs the collaborative approval decision results, the following steps are included: inputting the collaborative approval decision results and approval process configuration information into the large language model, performing compliance verification processing on the collaborative approval decision results, and outputting a compliance verification report. The compliance verification processing includes: verifying whether the approval authority matches the amount and type of the approval item, verifying whether necessary attachments are missing from the approval materials, and verifying whether the approval process complies with preset internal control requirements. When the compliance verification identifies potential compliance risks, the risk points are marked in the compliance verification report and an early warning message is generated, which is then appended to the collaborative approval decision results and output together.

[0054] Building upon step 4, to enable the system to employ differentiated processing strategies for approval items of varying complexity, the following steps are also included: After receiving the set of multimodal documents to be approved, an approval complexity score is calculated based on the type identifier of the approval item and key fields in the structured approval content data (including approval amount, number of departments involved, and number of document modalities). Since the approval amount, number of departments involved, and number of document modalities have different dimensions, before weighted summation, normalization based on the range is applied to each of the three fields to eliminate the dimensional differences, obtaining normalized values ​​for each field. Then, the normalized values ​​are weighted and summed using the following formula:

[0055] ; in, Scoring based on the complexity of the approval process. , and These are the normalized values ​​of the approved amount, the number of departments involved, and the number of document modalities, respectively, after normalization based on the range. , and The preset weight coefficients for each field, and satisfying the following conditions: .

[0056] The above weighting coefficients , and These values ​​respectively reflect the contribution of the approval amount, the number of departments involved, and the number of document modalities to the approval complexity, with a range of [value range missing]. And satisfy Approval complexity score The range of values ​​is , A higher value indicates a higher level of complexity for the approval process.

[0057] Score the complexity of the approval process The complexity level of the approval item is determined by comparing it with a preset complexity grading threshold. For approval items with a complexity level lower than the preset level, the OA approval process engine automatically processes them according to the preset approval rules. For approval items with a complexity level not lower than the preset level, the large language model collaborative approval processing flow described in steps 1 to 4 above is triggered, and the collaborative approval decision result with confidence level labels is output.

[0058] Step 5: Knowledge Accumulation and Cross-Node Migration Updates for Approval To enable the reuse of historical approval knowledge in subsequent approvals, after outputting the collaborative approval decision results in step 4, the following steps are also included: Step 501: Semantically encode the structured review opinions, conflict resolution process records, and collaborative approval decision results of each role in this approval process to generate an approval case vector for this approval case, and store the approval case vector and the corresponding approval process data into the vector database.

[0059] The semantic encoding process described above uses the text embedding capability of a large language model to map the text content of structured review opinions, conflict resolution process records, and collaborative approval decision results into fixed-dimensional vectors. Then, the vectors of each part are concatenated or weighted to obtain an approval case vector that can comprehensively represent the whole picture of this approval case. When stored in the vector database, the corresponding approval item type identifier is stored simultaneously so that subsequent retrieval and filtering can be performed by item type.

[0060] Step 502: Receive feedback from the approver regarding the collaborative approval decision. The feedback includes the approver's adoption flag for the collaborative approval decision, the content of any modifications, and the reasons for those modifications. The feedback is then associated with and stored in the corresponding approval case vector.

[0061] Building upon step 5, to continuously optimize the long-term performance of the role-specific review models, the following steps are also included: Based on accumulated feedback information, the adoption rate of review suggestions from each role-specific review model by approvers is calculated, and the historical review accuracy parameters of each role-specific review model are updated. When the accumulated feedback information reaches a preset data volume threshold, the modification content and reasons for modification in the feedback information are used as training corpus to perform supervised fine-tuning updates on each role-specific review model, thereby improving the inference accuracy of each role-specific review model in subsequent approvals.

[0062] Building upon step 5, to enable the structured storage and retrieval of judgment criteria and decision-making logic for similar approval matters across different roles, the following steps are also included: Based on historical approval case data accumulated in the vector database, the review focus dimensions, typical risk judgment criteria, and decision-making logic for each role in different types of approval matters are extracted, constructing a cross-role review knowledge graph. Nodes in the review knowledge graph include approval matter type nodes, review role nodes, review dimension nodes, and risk type nodes. Edges between nodes represent the focus relationships between roles and review dimensions, and the association relationships between review dimensions and risk types. When a new approval matter enters the approval process, a search is performed in the review knowledge graph based on the approval matter type information to extract typical review patterns and risk focus points for each role in similar approval matters. The search results are then embedded as supplementary context into prompts for each role's specific review model.

[0063] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0064] It should be understood that in the various embodiments of this application, the sequence number of each process 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 this application.

[0065] In addition, the functional units in the various embodiments of this application 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.

[0066] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A data collaborative processing method for an OA platform, characterized in that, include: The system receives a set of multimodal files to be approved, parses and processes each file in the set, and generates structured approval content data. The set of multimodal files includes at least two of the following: text application documents, attachment documents, image files, and table files. Based on the structured approval content data and the pre-configured list of approval roles, the role-specific review models corresponding to each approval role are invoked to perform parallel review processing on the structured approval content data. Each role-specific review model outputs the structured review opinions for its corresponding role. The role-specific review model is a model obtained based on a large language model and configured with prompts through the approval template for the corresponding role. The structured review opinions include review conclusions, confidence scores, risk levels, and reasoning paths. The structured review opinions output by each role are input into the collaborative decision-making arbitration layer. The review opinions are semantically aligned to generate semantic vectors for each role. The consensus index among multiple roles is calculated based on the semantic similarity between the semantic vectors of each review opinion. The consensus index is compared with a preset consensus threshold, and the conflict detection result is output. Based on the conflict detection results, when the conflict detection result is no conflict, the structured review opinions of each role are comprehensively processed to synthesize the collaborative approval decision result; when the conflict detection result is that a conflict exists, a multi-level conflict resolution strategy is adopted to resolve the conflict in sequence, and the collaborative approval decision result is synthesized based on the resolution result; wherein, the multi-level conflict resolution strategy includes rule arbitration, semantic mediation and weighted voting in order of increasing complexity.

2. The data collaborative processing method for the OA platform according to claim 1, characterized in that, After generating the structured approval content data, the process also includes: Based on the structured data of each file in the multimodal file collection, a cross-file association index is established. Specifically, the key information fields in the structured data of each file are semantically encoded to generate semantic vectors corresponding to each key information field. Cosine similarity is used to calculate the semantic correlation between key information fields in different files. Key information fields with cosine similarity higher than the association threshold are associated with each other to form a unified semantic representation of the approval content. The association threshold is a numerical parameter pre-configured according to the type of approval item.

3. The data collaborative processing method for the OA platform according to claim 1, characterized in that, Before the parallel review process is carried out using the individual role-specific review models, it also includes: The key information fields in the structured approval content data are semantically encoded to generate query vectors. The top K historical approval cases with semantic similarity to the query vectors are retrieved from the vector database, where K is a preset positive integer parameter for the number of searches. The review opinions and reasoning paths of the K historical approval cases obtained from the retrieval are used as reference context and concatenated into the prompts of the role-specific review model.

4. The data collaborative processing method for the OA platform according to claim 1, characterized in that, The parallel review process also includes: Based on the current approval node's position in the approval process and the structured review opinions already output by the previous approval node, context-aware prompts are dynamically generated. Specifically, the review conclusions, risk levels, and reasoning paths of the previous approval nodes are extracted as context summaries, and these context summaries are embedded into the prompts of the role-specific review model corresponding to the current approval node. This allows the role-specific review model of the current node to perceive the current approval process stage and the background of previous decisions, avoiding duplicate reviews of content already reviewed by previous nodes.

5. The data collaborative processing method for an OA platform according to claim 1, characterized in that, The consensus index is calculated as follows: the cosine similarity of the semantic vectors of the review opinions between each role pair is summed and then divided by the total number of role pairs to obtain the consensus index; wherein, the consensus threshold is a numerical parameter pre-configured according to the type of approval item, and the range of values ​​for both the consensus index and the consensus threshold is a closed interval from negative one to positive one. After determining that a conflict exists, the process also includes: splitting the structured review opinions of each role according to the review dimensions, calculating the semantic similarity of the role pairs on each review dimension, identifying the specific review dimensions whose semantic similarity is lower than the dimension conflict threshold, marking the review dimension as a conflict dimension, and transmitting the conflict dimension and the review opinions of the corresponding role to the conflict resolution unit.

6. The data collaborative processing method for an OA platform according to claim 1, characterized in that, Rule arbitration includes: obtaining the priority weights of each conflicting role; ranking the review conclusions of the conflicting roles according to the priority weights; selecting the review conclusion of the role with the highest priority as the preliminary ruling result; calculating the semantic relevance between the preliminary ruling result and the review conclusions of the remaining conflicting roles using cosine similarity; if all cosine similarities are higher than the ruling acceptance threshold, the rule arbitration is deemed successful, and the preliminary ruling result is transmitted to the decision synthesis unit; if there is a case where the cosine similarity is lower than the ruling acceptance threshold, the semantic mediation process is initiated. The semantic mediation process includes: taking the review opinions, reasoning paths, and conflict dimension information of each conflicting role as input, constructing mediation prompts, calling a large language model to analyze the root causes of the conflict and generating mediation suggestions that take into account the concerns of all parties; using cosine similarity to calculate the semantic relevance between the mediation suggestions and the review opinions of each conflicting role; if the cosine similarity between the mediation suggestions and the review opinions of each conflicting role is higher than the mediation acceptance threshold, the semantic mediation is deemed successful; if there is a case where the cosine similarity is lower than the mediation acceptance threshold, the process proceeds to the weighted voting.

7. The data collaborative processing method for the OA platform according to claim 1, characterized in that, Weighted voting includes: Based on the historical review accuracy and current review opinion confidence score of each conflict role's role-specific review model, the comprehensive decision weight of each conflict role is calculated. The comprehensive decision weight is calculated as follows: the sum of the historical review accuracy of the i-th conflict role multiplied by a first weight adjustment coefficient and the confidence score of that conflict role multiplied by a second weight adjustment coefficient is divided by the sum of all conflict roles calculated in the same way. The comprehensive decision weight of the i-th conflict role is obtained by multiplying the historical review accuracy by a first weight adjustment coefficient and the second weight adjustment coefficient by one. The historical review accuracy and the confidence score both fall within a closed interval from zero to one. Based on the comprehensive decision-making weight of each conflicting role, the risk level values ​​corresponding to the review conclusions of each conflicting role are weighted and summed to obtain a weighted risk score. The weighted risk score is then compared with a preset risk level classification standard to determine the final comprehensive risk level and review conclusion, which serves as the weighted voting decision result.

8. The data collaborative processing method for an OA platform according to claim 1, characterized in that, After receiving the set of multimodal documents awaiting approval, the process also includes: Based on the type identifier of the approval item and the key fields in the structured approval content data, the approval complexity score is calculated. The key fields include the approval amount, the number of departments involved, and the number of document modalities. The normalized values ​​of each of the three key fields are obtained by normalization based on the range. The approval complexity score is then obtained by weighted summation of the normalized values. The sum of the weight coefficients of each field is one. The complexity score of the approval is compared with a preset complexity grading threshold to determine the complexity level of the approval item.

9. The data collaborative processing method for an OA platform according to claim 1, characterized in that, After outputting the collaborative approval decision results, it also includes: The structured review opinions of each role in this approval process, the conflict resolution process records, and the collaborative approval decision results are semantically encoded to generate an approval case vector for this approval case. The approval case vector and the corresponding approval process data are stored in the vector database, and the approval item type identifier is stored synchronously. Receive feedback information from the approver regarding the collaborative approval decision result. The feedback information includes the approver's adoption indicator of the collaborative approval decision result, the modified content, and the reason for the modification. The feedback information is associated with and stored with the corresponding approval case vector. Based on the accumulated feedback information, the percentage of review suggestions from each role-specific review model that are adopted by the approver is statistically analyzed, and the historical review accuracy parameters of each role-specific review model are updated. When the accumulated feedback information reaches a preset data volume threshold, the modified content and reasons for modification in the feedback information are used as training corpus to perform supervised fine-tuning and updates on each role-specific review model.

10. A data collaboration processing method and system for an OA platform, wherein the data collaboration processing method for an OA platform according to any one of claims 1-9 is characterized in that, include: The multimodal file parsing module is used to receive a set of multimodal files to be approved, parse and process each file in the multimodal file set, and generate structured approval content data. The multi-role parallel review module is used to perform parallel review processing of the structured approval content data based on the structured approval content data and the pre-configured list of approval roles, by calling the role-specific review model corresponding to each approval role, and outputting the structured review opinions of each role. The conflict detection module is used to input the structured review opinions output by each role into the collaborative decision-making arbitration layer, perform semantic alignment processing on each review opinion, generate semantic vectors of review opinions corresponding to each role, calculate the consensus index among multiple roles based on the semantic similarity between the semantic vectors of each review opinion, compare the consensus index with the preset consensus threshold, and output the conflict detection result. The decision synthesis module is used to synthesize collaborative approval decision results based on the conflict detection results. When the conflict detection results are that there is no conflict, the module comprehensively processes the structured review opinions of each role and synthesizes collaborative approval decision results. When the conflict detection results are that there is a conflict, the module adopts a multi-level conflict resolution strategy to resolve the conflict in sequence and synthesizes collaborative approval decision results based on the resolution results.

Images