Business system building method and system

Abstract

The invention discloses a business system building method and system, and relates to the field of business process management, and the method comprises the steps: recognizing related entities such as roles, demands, strategic targets and the like through a bidirectional undifferentiated coding model, carrying out semantic coding and element classification through an attention polarity model, and generating a benefit related party graph in combination with a rule engine; calculating a demand weight by means of a graph neural network, embedding a semantic vector to generate a weighted strategic vector, constraining an AI model to extract flow nodes, and correlating to generate a flow element table; a final-stage flow chart is generated through state coding and a near-end strategy optimization learning network, a programming environment is determined based on requirements, a code segment is matched to generate a service system, full-process automation from strategic analysis to service system generation is achieved, the development efficiency of the service system is improved, and the technical threshold is reduced.

Description

Technical Field

[0001] The present invention relates to the field of business process management, and in particular to a business system construction method and system. Background Art

[0002] In the digital transformation of enterprises, business system development must accurately align with strategic objectives and efficiently implement them. Traditional approaches rely on manual analysis of strategic elements, design of business processes, and code writing, leading to problems such as distorted strategic communication, inefficient process design, and long system implementation cycles.

[0003] The existing Chinese patent application with publication number CN119886777A discloses a process link evaluation method based on graph neural network. The graph neural network analyzes the bottlenecks and redundancies of process nodes to achieve the rationality evaluation of process design.

[0004] Existing technologies only focus on the static evaluation of process links, do not involve the automated decomposition and semantic transmission of strategic goals, and do not solve the problem of intelligent mapping of process nodes to code segments. Summary of the Invention

[0005] In view of the deficiencies in the existing technology, the present invention proposes a business system construction method and system to achieve intelligent business system construction based on automated disassembly and semantic conduction analysis of strategic goals.

[0006] The technical solution to achieve the purpose of the present invention is:

[0007] The business system construction method includes the following steps:

[0008] A bidirectional, indiscriminate encoding model is used to identify relevant entities in documents. An attention polarity model is used to semantically encode strategic objectives and classify them into internal or external elements based on attention weight aggregation. Polarity analysis is combined to create a list of strategic elements. A rule engine is used to connect roles and requirements to strategic objectives to generate a stakeholder map. Relevant entities include roles, requirements, and strategic objectives.

[0009] The demand weight of each requirement in the stakeholder graph is calculated through a graph neural network. The demand semantic vector supporting each strategic goal is embedded into the strategic semantic vector of the strategic goal in the strategic element list according to the corresponding demand weight to generate a weighted strategic vector. Based on the stakeholder graph, the GPT4-Turbo model is constrained, the process nodes of the business document are extracted and associated to generate a process element table. The strategic semantic vector and demand semantic vector are generated by the preset language model of strategic goal and demand input;

[0010] Each process element in the process element table is state-encoded, and the proximal strategy is used to optimize the learning network. Action space filtering is set based on the stakeholder graph. The process elements are connected based on the strategic transmission reward guidance to generate the final-level process diagram. The state encoding is implemented based on the demand weight table and the weighted strategy vector set.

[0011] The programming language and compilation environment are determined based on customer needs. The process semantic vectors of the process nodes in the final-level flowchart are screened and matched with the code semantic vectors of the code segments in the code library according to the logical verification function. The matching code segments are filled in for each process node to generate a business system. The logical verification function is equal to the norm value of the intersection of the process semantic vector and the code semantic vector divided by the norm value of the code semantic vector.

[0012] Furthermore, the bidirectional non-discriminatory coding model identifies strategic objectives and related entities in the basic document, including the following steps:

[0013] Use a tokenizer to split each document into short sentences and further split into subwords, adding a start marker at the beginning of the document and a split marker at the end of the short sentence;

[0014] Insert filler subwords before the segmentation markers of all short sentences so that the number of subwords in each short sentence is the same;

[0015] Map subwords to numbered vectors based on the vocabulary, generate position vectors based on the order of the subwords in the document, and superimpose the subword number vectors to generate the subword embedding vectors;

[0016] The embedding vectors of each subword in the short sentence are concatenated and input into the Transformer encoder. After being processed through 12 layers of continuous multi-head self-attention mechanism, residual connection, layer normalization and feedforward neural network, a context representation vector is generated.

[0017] The context representation of each subword in the context representation vector is converted into a probability distribution for each subword through linear modulation and the Softmax function using a linear classifier. The BIN label with the maximum probability is assigned to the subword. The BIN label uses the prefix + type format. Entity A has two BIN labels, BA and IA, representing the start word and continuation word of entity A, respectively. Non-entities only have the label NN.

[0018] According to the principle of continuous merging, the subwords of the start word and the continuation word in the short sentence that are sequentially continuous and whose BIN tags correspond to the same entity are merged to identify related entities in the document.

[0019] Furthermore, the principle of continuous consolidation includes the following steps:

[0020] Perform a start check on the first subword in the short sentence. If the label of the first subword is NN, skip it and perform a new start check on the second subword. If the label of the first subword represents the start word of the entity, perform a continuation check on the second subword.

[0021] If the label of the second subword represents a continuation of an entity and is the same as the entity of the first subword, the two subwords are merged and a continuation check is performed on the third subword;

[0022] If the label of the second subword is NN, the first subword is considered as an entity and a new starting test is performed on the third subword;

[0023] If the label of the second subword represents the start word of an entity, the first subword is treated as an entity alone and the continuation test of the second subword is treated as a new start test, and a new continuation test is performed on the third subword;

[0024] Stop after checking all subwords in the sentence.

[0025] Furthermore, the attention polarity model builds a checklist of strategic elements, including the following steps:

[0026] Convert strategic goals into strategic semantic vectors through a preset language model;

[0027] The product of the strategic semantic vector, the internal query vector, and the external query vector is converted into internal weight and external weight through the self-attention mechanism;

[0028] The product of the strategic semantic vector, internal weight and external weight is concatenated and converted into the internal and external probability distribution of the strategic target through linear modulation and Softmax function;

[0029] The convolution kernels of three different window sizes are convolved with the strategic semantic vector and local features are enhanced using the PReLU function and maximum pooling. Semantic features of three different scales are generated and spliced ​​into the fully connected layer. The polarity probability distribution is output based on the Softmax function.

[0030] Combine the internal and external probability distributions with the polarity probability distribution to construct the SWOT factor probability distribution of the strategic goal, and establish an edge between the SWOT factor with the highest probability and the strategic goal. The value of the edge is equal to the highest probability in the SWOT factor probability distribution. Group the strategic goals according to each SWOT factor and reconstruct the strategic factor list.

[0031] Furthermore, a rule engine is used to match the roles and requirements of related entities with strategic goals to generate a stakeholder map, which includes the following steps:

[0032] Establish entity chain rules, establish the affiliation relationship of the role demand chain and the supporting relationship of the demand strategic goal chain, and establish a two-way corresponding rule, that is, the roles and demands, and demands and strategic goals are all in a many-to-many form;

[0033] Each relevant entity is converted into a bag-of-words feature vector and a corresponding semantic vector through the TF-IDF converter and the preset language model, and then concatenated into an entity description vector;

[0034] Arrange and combine each role and requirement, as well as requirement and strategic goal, build role-requirement pairs and requirement-strategy goal pairs and calculate the cosine similarity respectively. Establish attribution edges and support edges between roles and requirements, as well as requirements and strategic goals whose cosine similarity is greater than or equal to the similarity threshold, and generate a stakeholder map. The attribution edge value and support edge value are equal to the corresponding cosine similarity.

[0035] Specifically, the graph neural network counts the number of roles with belonging edges for each requirement in the stakeholder graph as the out-degree, divides the out-degree of each requirement by the out-degree sum of all requirements to generate the requirement weight of each requirement, encodes each strategic goal in the strategic element list into a strategic semantic vector through a preset language model, retrieves the requirements that can support each strategic goal from the stakeholder graph and converts them into requirement semantic vectors through a preset language model, performs weighted summation based on the corresponding requirement weights and superimposes them with the strategic semantic vector of each strategic goal to generate a weighted strategic vector for each strategic goal.

[0036] Specifically, process node role matching rules are pre-established in the GPT4-Turbo model, all relevant role requirement chains are introduced in each matching pair of process node and role and integrated to build a template rule library, the natural language processing capability of the GPT4-Turbo model is used to identify process nodes in business documents and match the template rule library, and obtain the role requirement chain matching the process node, and the text description of the process node and the requirements in each role requirement chain are converted into process semantic vectors and requirement semantic vectors through the preset language model and the cosine similarity is calculated, the role requirement chain and the process node corresponding to the requirement semantic vector with the maximum cosine similarity are organized into process elements, and the process elements of all process nodes are combined to build a process element table.

[0037] Furthermore, each process element in the process element table is converted into a corresponding process state vector through state coding, including the following steps:

[0038] The text description of the process node and the role in each process element are converted into a process semantic vector and a role semantic vector through a preset language model;

[0039] Based on the stakeholder map and demand weight table, the requirements and corresponding demand weights related to the roles in the process elements are obtained. The role-related requirements are generated into a demand semantic vector through a preset language model. Based on the corresponding demand weights, a weighted sum is performed and the role semantic vectors of the roles are superimposed to generate a weighted role vector.

[0040] Search the stakeholder map for strategic goals that can be supported by the requirements of the process elements and obtain the corresponding weighted strategic vectors. Based on the weighted strategic vectors obtained by weighted aggregation of the supporting edge values ​​between the corresponding requirements and the strategic goals, generate the strategic association requirement vector for each process element.

[0041] The process semantic vector, weighted role vector and strategic association requirement vector of each process element are concatenated to generate the process state vector of each process element.

[0042] Furthermore, the proximal strategy is used to optimize the learning network to generate a final-level flow chart that meets the requirements of the business document, including the following steps:

[0043] The business strategy goals in the business document are converted into business strategy semantic vectors using a preset language model. The strategic conduction reward is designed to be equal to the cosine similarity between the current process state vector and the business strategy semantic vector multiplied by the conduction coefficient. The conduction coefficient is equal to the sum of the edge values ​​of the attribution and / or supporting edges that the role corresponding to the current process state vector passes through in the stakeholder graph to reach the business strategy goals.

[0044] Based on the proximal policy optimization algorithm, action space filtering is set up based on the graph structure of the stakeholder graph to generate a policy network. The policy network is used to select the next process element. Action space filtering means that when selecting the next process element, the process elements with roles or requirements that do not have direct or indirect edges connected to the roles and requirements in the current process element in the stakeholder graph are eliminated;

[0045] Select the process state vector of a process element as the starting point. The policy network selects the next process element to connect based on the proximal policy optimization algorithm and strategic conduction reward. Repeat this step until it stops.

[0046] If the stop is due to having traversed all process elements, a final-level flow chart is generated;

[0047] If the stop is due to spatial filtering and there are no selectable process elements, the process flow diagram being constructed is discarded and a new process element is reselected as the starting point;

[0048] Obtain all final-level flowcharts that traverse all process elements, and select the final-level flowchart with the largest cumulative strategic conduction reward for output.

[0049] A business system construction system for executing the business system construction method, comprising a graph construction module, a process extraction module, a process generation module, and an integrated construction module;

[0050] The graph construction module uses a bidirectional non-discriminatory encoding model to identify relevant entities in documents, an attention polarity model to semantically encode strategic goals, and classifies them into internal or external elements based on attention weight aggregation. Polarity analysis is combined to create a list of strategic elements, and a rule engine is used to connect roles and requirements to strategic goals to generate a stakeholder graph. The relevant entities include roles, requirements, and strategic goals.

[0051] The process extraction module calculates the demand weight of each requirement in the stakeholder graph through a graph neural network. It then embeds the demand semantic vector of the requirements supporting the strategic goal into the strategic semantic vector of each strategic goal in the strategic element list according to the corresponding demand weight, generating a weighted strategic vector. Based on the GPT4-Turbo model constrained by the stakeholder graph, it extracts the process nodes of the business document and associates them to generate a process element table. The strategic semantic vector and demand semantic vector are generated by the preset language model of the strategic goal and demand input.

[0052] The process generation module encodes the state of each process element in the process element table, uses a proximal strategy to optimize the learning network, sets action space filtering based on the stakeholder map, and connects process elements based on strategic transmission reward guidance to generate the final-level process diagram. The state encoding is implemented based on the demand weight table and the weighted strategy vector set.

[0053] The integrated construction module determines the programming language and compilation environment based on customer needs, screens and matches the process semantic vectors of the process nodes in the final-level flowchart with the code semantic vectors of the code segments in the code library according to the logical verification function, fills in the matching code segments for each process node, and generates a business system, where the logical verification function is equal to the norm value of the intersection of the process semantic vector and the code semantic vector divided by the norm value of the code semantic vector.

[0054] Compared with the existing technology, the present invention automatically identifies entities through a bidirectional non-discriminatory coding model, uses an attention polarity model and graph neural network to achieve semantic encoding of strategic goals, demand weight calculation and factor aggregation, and generates a list of strategic factors and a stakeholder map; uses the GPT4-Turbo model to constrain the extraction of process nodes, and combines the proximal policy optimization learning network to generate the final level flowchart; based on the logical verification function, it realizes the automatic matching of process semantic vectors and code semantic vectors, forming an end-to-end closed loop of "strategy-process-code". This solution constructs a cross-modal semantic mapping mechanism through the integration of multiple technologies, breaks through the bottleneck of manual intervention, and systematically solves the technical bottlenecks of cross-level factor transmission and cross-modal data mapping, realizing the automation of the entire process from strategic analysis to business system generation, significantly improving the efficiency of strategic implementation and the accuracy of system development. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Build method flow charts for business systems;

[0056] Figure 2 Flowchart of the processing steps for the bidirectional non-discriminatory coding model;

[0057] Figure 3 Schematic diagram of the attention polarity model;

[0058] Figure 4 Build a system diagram for the business system. DETAILED DESCRIPTION

[0059] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0060] Example 1

[0061] like Figure 1 As shown, the present invention discloses a business system construction method, comprising the following steps:

[0062] A bidirectional, indiscriminate encoding model is used to identify the different relevant entities in each document. An attention polarity model is used to semantically encode each strategic goal and calculate the internal and external attention weights using a learnable query vector. Internal or external factors are aggregated and categorized, and a list of strategic factors is created using polarity analysis. A rule engine is used to match the roles and requirements in the relevant entities to the strategic goals to generate a stakeholder map. The relevant entities include specific roles, requirements, and strategic goals, such as "suppliers," "order tracking," and "increasing market share."

[0063] The demand weight of each requirement in the stakeholder graph is calculated through a graph neural network, and a demand weight table is constructed. The demand semantic vector supporting each strategic goal is weighted and embedded according to the corresponding demand weight in the strategic semantic vector of each strategic goal in the strategic element list to construct a weighted strategic vector set. The entity chain rule constraints of the stakeholder graph are introduced into the GPT4-Turbo model, and the process nodes in the business documents are extracted and associated with the relevant role demand chain to generate a process element table. The strategic semantic vector and demand semantic vector are generated by inputting the strategic goal and demand into the preset language model, which includes the BERT model and the RoBERTa model.

[0064] Each process element in the process element table is converted into a corresponding process state vector through state coding. The proximal strategy is used to optimize the learning network. Based on the graph structure of the stakeholder graph, action space filtering is set to prohibit cross-chain jumps. Based on the strategic transmission reward, state transfer actions are learned to connect different process state vectors to generate a final-level process diagram that meets the requirements of the business document. Among them, the state coding is implemented based on the demand weight table and the weighted strategy vector set.

[0065] Determine the programming language and compilation environment based on customer needs, obtain the final-level flowchart, and convert the text description of each process node and the functional label of each code segment in the code library into a process semantic vector and a code semantic vector respectively based on the preset language model. Calculate the matching degree between the process node and the code segment based on the logical verification function, select the code segment with the highest matching degree and fill it into the process node to generate the business system. The logical verification function is equal to the norm value of the intersection of the process semantic vector and the code semantic vector divided by the norm value of the code semantic vector.

[0066] like Figure 2 As shown, further, the bidirectional non-discriminatory coding model identifies strategic goals and related entities in the basic document, including the following steps:

[0067] Use the WordPiece tokenizer to split each document into short sentences, and further split each short sentence into subwords, adding a start marker at the beginning of the document and a split marker at the end of each short sentence;

[0068] Get the maximum number of subwords in a short sentence, insert filler subwords between the end of each short sentence with a subword number less than the maximum number of subwords and the corresponding segmentation marker so that the number of subwords in each short sentence is equal to the maximum number of subwords;

[0069] Map each subword to a numbered vector based on the vocabulary, where the numbered vector of the padded subword is an all-zero vector. Sine-cosine positional encoding is used to generate a position vector based on the order of each subword in the document. This vector is then superimposed on the subword's numbered vector to fuse semantic information and position information, generating an embedding vector for each subword.

[0070] The embedding vectors of the subwords in each short sentence are concatenated and fed into a 12-layer Transformer encoder for multi-scale analysis. Each layer includes a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism concatenates the results of the 12 self-attention mechanisms and projects them back to the original dimension before the self-attention mechanism is input through linear weights. After residual connections and layer normalization, the results are fed into a feedforward neural network for feedforward processing. After passing through the 12-layer Transformer encoder, the context representation vector is output.

[0071] The context representation of each subword in the context representation vector is linearly modulated using a linear classifier and converted into a probability distribution for each subword using the Softmax function. The BIN label with the maximum probability is then assigned to the subword. The probability distribution records the probability of a subword belonging to each BIN label. The BIN label uses a prefix + type format. For any specific entity A, there are two BIN labels BA and IA, representing the start word and continuation word of entity A, respectively. For non-entities, there is only the label NN. Entity A is a strategic goal, a specific role, a specific need, or a specific strategic goal. The BIN format makes entity boundaries clear and separable, reducing boundary ambiguity.

[0072] According to the principle of continuous merging, the subwords of the start word and the continuation word in the short sentence that are sequentially continuous and whose BIN labels correspond to the same entity are merged to identify the relevant entities of each short sentence in the document.

[0073] Furthermore, the principle of continuous consolidation includes the following steps:

[0074] Perform a start check on the first subword in each short sentence. If the label of the first subword is NN, skip it and perform a new start check on the second subword. If the label of the first subword represents the start word of the entity, perform a continuation check on the second subword.

[0075] If the label of the second subword represents the continuation of the entity and is the same as the entity of the first subword, the two subwords are merged and the continuation test is performed on the third subword. If the label of the second subword is NN, the first subword is regarded as an entity alone and a new start test is performed on the third subword. If the label of the second subword represents the start word of the entity, the first subword is regarded as an entity alone and the continuation test of the second subword is regarded as a new start test, and a new continuation test is performed on the third subword.

[0076] Stop after checking all subwords in the sentence.

[0077] like Figure 3 As shown, further, the attention polarity model establishes a list of strategic elements, including the following steps:

[0078] Semantically encode strategic goals through a preset language model and convert them into strategic semantic vectors;

[0079] Through the self-attention mechanism, the product of the strategic semantic vector and the internal query vector and the product of the strategic semantic vector and the external query vector are converted into internal weights and external weights respectively. The internal query vector and the external query vector are learnable parameters in the attention polarity model. The internal weight and the external weight reflect the semantic relevance of the strategic semantic vector with internal factors and external factors respectively. The internal factors include strengths S and weaknesses W in SWOT factors, and the external factors include opportunities O and challenges T in SWOT factors.

[0080] The product of the strategic semantic vector, internal weight and external weight is concatenated and converted into the internal and external probability distribution of the strategic goal through linear modulation and Softmax function. The internal and external probability distribution records the probability of the strategic goal and the internal and external factors.

[0081] Three convolution kernels with different window sizes are slid on the strategic semantic vector to capture local semantic features of different scales. The PReLU function is used to enhance the nonlinear expression ability of the three convolution results and the dimension reduction is performed based on the maximum pooling to enhance the most significant local features. The semantic features of three different scales are generated, where the window sizes are 1, 3, and 5 respectively.

[0082] The semantic features of three different scales are spliced ​​and fused through a fully connected layer for dimensionality reduction. The polarity probability distribution is output based on the Softmax function. The polarity probability distribution records the probability of the strategic goal belonging to positive polarity or negative polarity. Among them, advantages S and opportunities O are positive factors, while disadvantages W and challenges T are negative factors.

[0083] Based on the knowledge of probability statistics, the probability distribution of the SWOT factors of the strategic goal is constructed by combining the internal and external probability distributions with the polarity probability distribution. An edge is established between the SWOT factor with the highest probability and the strategic goal. The value of the edge is equal to the highest probability in the probability distribution of the SWOT factor.

[0084] Filter and group strategic goals based on strengths S, weaknesses W, opportunities O, and challenges T, reorganize the grouping results in a structured table format, and generate a list of strategic elements.

[0085] Furthermore, a rule engine is used to match the roles and requirements of related entities with strategic goals to generate a stakeholder map, which includes the following steps:

[0086] Establish entity chain rules and the attribution relationship of the role demand chain, that is, the demand is proposed by the role, and establish the support relationship of the demand strategic goal chain, that is, satisfying the demand can partially or fully achieve the strategic goal;

[0087] Establish a two-way correspondence rule, with roles and requirements, and requirements and strategic goals in a many-to-many relationship. This rule means that one role can raise multiple requirements simultaneously, one requirement can be raised by different roles, one requirement can support multiple strategic goals, and one strategic goal can be supported by multiple requirements simultaneously.

[0088] Each relevant entity is converted into a bag-of-words feature vector using a TF-IDF converter. Each role, requirement, and strategic goal is converted into a role semantic vector, a requirement semantic vector, and a strategy semantic vector using a preset language model. The bag-of-words feature vector and the corresponding semantic vector of each relevant entity are concatenated into an entity description vector.

[0089] Arrange and combine each role and requirement to construct a role-requirement pair, calculate the cosine similarity of each role-requirement pair, establish attribution edges between roles and requirements whose cosine similarity is greater than or equal to the similarity threshold, arrange and combine each requirement and strategic goal to construct a requirement-goal pair and calculate the cosine similarity, establish support edges between requirements and strategic goals whose cosine similarity is greater than or equal to the similarity threshold, and the attribution edge value and support edge value are equal to the corresponding cosine similarity;

[0090] When the attribution edge and the supporting edge are fully established, a stakeholder map is generated.

[0091] Specifically, the graph neural network starts from each demand in the stakeholder graph, retrieves the number of roles with which each demand has an attributed edge as the out-degree of each demand, divides the out-degree of each demand by the out-degree sum of all demands to generate the demand weight of each demand. The demand weight reflects the importance of each demand, and the demand weight of each demand is combined to construct a demand weight table. Each strategic goal in the strategic element list is encoded into a strategic semantic vector through a preset language model. The needs that can support each strategic goal are retrieved from the stakeholder graph and converted into demand semantic vectors through a preset language model. The corresponding demand weights in the demand weight table are weighted summed and superimposed on the strategic semantic vector of each strategic goal to generate a weighted strategic vector for each strategic goal, and the weighted strategic vector set is constructed in combination.

[0092] Specifically, process node role matching rules are pre-established in the GPT4-Turbo model, and all relevant role requirement chains in the stakeholder graph are introduced into each matching pair of process node and role. The template rule library is integrated in JSON format, and the natural language processing capability of the GPT4-Turbo model is used to identify the process nodes existing in the business documents and match them in the template rule library to obtain the role requirement chain matching the process nodes. The text description of the process node and the requirements in each role requirement chain are converted into process semantic vectors and requirement semantic vectors respectively through the preset language model, and the cosine similarity between the process semantic vector and each relevant requirement semantic vector is calculated. The role requirement chain and the process node corresponding to the requirement semantic vector with the maximum cosine similarity are organized into process elements according to JSON format, and the process elements of all process nodes are combined to construct a process element table.

[0093] Furthermore, each process element in the process element table is converted into a corresponding process state vector through state coding, including the following steps:

[0094] The text description of the process node and the role in each process element are converted into a process semantic vector and a role semantic vector respectively through a preset language model;

[0095] Based on the stakeholder graph, the requirements related to the roles in the process elements are obtained and the corresponding requirement weights are obtained from the requirement weight table. Based on the graph neural network, the requirement semantic vectors generated by the preset language model for the role-related requirements are weighted and summed based on the corresponding requirement weights and superimposed on the role semantic vector of the role to generate a weighted role vector;

[0096] Search the stakeholder map for strategic goals that can be supported by the requirements of the process elements and obtain the corresponding weighted strategic vectors. Perform secondary weighted aggregation on all the obtained weighted strategic vectors based on the supporting edge values ​​between the corresponding requirements and the strategic goals to generate the strategic associated requirement vector for each process element.

[0097] The process semantic vector, weighted role vector and strategic association requirement vector of each process element are spliced ​​together to generate the process state vector of each process element. The process state vector indirectly reflects the state distribution of the strategic goals that can be achieved when the roles in each process element select different requirements at the process node.

[0098] Furthermore, the proximal strategy is used to optimize the learning network to generate a final-level flow chart that meets the requirements of the business document, including the following steps:

[0099] The business strategy objectives in the business document are converted into business strategy semantic vectors using a preset language model. The strategic conduction reward is designed to be equal to the cosine similarity between the current process state vector and the business strategy semantic vector multiplied by the conduction coefficient. The conduction coefficient is equal to the shortest path length of the role corresponding to the current process state vector to the business strategy objective in the stakeholder graph. The path length is equal to the sum of the edge values ​​of the attribution edges and / or supporting edges that the role passes through to reach the business strategy objective.

[0100] Define the action as selecting the next process element and connecting it to the current process element;

[0101] Based on the traditional proximal policy optimization algorithm, action space filtering is set up based on the graph structure of the stakeholder graph to generate a policy network. The policy network is used to guide the adaptive execution of the optimal action. Action space filtering means that when selecting the next process element, roles or requirements that do not have direct or indirect edges connected to the roles and requirements in the current process element in the stakeholder graph are eliminated;

[0102] A process state vector of a process element is selected as the starting point. The policy network generates the best action to reach the next process element based on the proximal policy optimization algorithm and the strategic transmission reward as the expert guidance. This step is repeated until all process elements are traversed or the next process element cannot be selected based on spatial filtering. If all process elements are traversed, a final-level flow chart is generated. If the next process element cannot be selected, the flow chart under construction is discarded and a new unselected process element is selected as the starting point.

[0103] Obtain all final-level flowcharts that traverse all process elements, and select the final-level flowchart with the largest cumulative strategic conduction reward for output. This final-level flowchart is the final-level flowchart that best meets business requirements.

[0104] Example 2

[0105] like Figure 4 As shown, the present invention also discloses a business system construction system for executing the business system construction method, including a graph construction module, a process extraction module, a process generation module and an integrated construction module;

[0106] The graph construction module uses a bidirectional non-discriminatory encoding model to identify different relevant entities in each document. It uses an attention polarity model to semantically encode each strategic goal and calculates internal and external attention weights using a learnable query vector. These are then aggregated and classified as internal or external elements. Polarity analysis and classification are then combined to create a list of strategic elements. A rule engine is then used to match the roles and requirements of relevant entities to the strategic goals to generate a stakeholder graph.

[0107] The process extraction module uses a graph neural network to calculate the demand weight of each requirement in the stakeholder graph, constructs a demand weight table, and embeds the demand semantic vectors supporting each strategic goal in the strategic semantic vector of each strategic goal in the strategic element list according to the corresponding demand weight. This constructs a weighted strategic vector set. The GPT4-Turbo model introduces the entity chain rule constraints of the stakeholder graph, extracts process nodes from business documents, and associates them with the relevant role demand chain to generate a process element table.

[0108] The process generation module converts each process element in the process element table into a corresponding process state vector through state encoding. It uses a proximal strategy to optimize the learning network. Based on the graph structure of the stakeholder graph, it sets action space filtering to prohibit cross-chain jumps. Based on the strategic transmission reward, it learns state transfer actions to connect different process state vectors. Based on the proximal strategy optimization algorithm, it updates the network parameters and generates a final-level process diagram that meets the requirements of the business document.

[0109] The integrated construction module determines the programming language and compilation environment based on customer needs, obtains the final-level flowchart, and converts the text description of each process node and the functional label of each code segment in the code library into a process semantic vector and a code semantic vector respectively based on the preset language model. The matching degree between the process node and the code segment is calculated based on the logical verification function, and the code segment with the highest matching degree is selected to fill in the process node to generate the business system.

[0110] The present invention discloses a method and system for building a business system. It automatically identifies entities through a bidirectional non-discriminatory coding model, uses an attention polarity model and a graph neural network to achieve semantic coding of strategic goals, demand weight calculation, and factor aggregation, and generates a list of strategic factors and a stakeholder map. It extracts process nodes through constraints using the GPT4-Turbo model, and generates a final-level flowchart in combination with a proximal strategy optimization learning network. It realizes automatic matching of process semantic vectors and code semantic vectors based on a logical verification function, forming an end-to-end closed loop of "strategy-process-code". This solution constructs a cross-modal semantic mapping mechanism through the integration of multiple technologies, breaks through the bottleneck of manual intervention, and systematically solves the technical bottlenecks of cross-level factor transmission and cross-modal data mapping, achieving full-process automation from strategic analysis to business system generation, significantly improving the efficiency of strategic implementation and the accuracy of system development.

[0111] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A business system construction method, characterized in that: The following steps are involved: A bidirectional, non-discriminatory encoding model is used to identify roles, requirements, and strategic goals in documents. An attention polarity model is used to semantically encode strategic goals and aggregate them into internal or external elements. Polarity analysis is combined to create a list of strategic elements. A rule engine is used to connect roles and requirements to strategic goals to generate a stakeholder map. Generate corresponding semantic vectors for roles, requirements, and strategic goals based on a preset language model. Calculate the demand weight for each requirement in the stakeholder map. Embed the demand semantic vectors supporting each strategic goal into the strategic semantic vectors of the strategic goals in the strategic element list based on the corresponding demand weights to generate a weighted strategic vector. Constrain the AI ​​model based on the stakeholder map, extract the process nodes of the business documents, and associate them to generate a process element table. Based on the demand weight table and the weighted strategic vector of each strategic goal, the process elements in the process element table are state-encoded. The proximal strategy is used to optimize the learning network. Action space filtering is set based on the stakeholder map. The process elements are connected based on the strategic transmission reward guidance to generate the final level flowchart. The business system is generated by matching the process nodes in the final-level flowchart with the code segments of the code base according to the logic verification function, wherein the logic verification function is equal to the norm value of the intersection of the process semantic vector of the process node and the code semantic vector of the code segment divided by the norm value of the code semantic vector.

2. The business system construction method according to claim 1, characterized in that: The method of using a bidirectional non-discriminatory coding model to identify roles, requirements, and strategic goals in a document includes the following steps: Use a tokenizer to split the document into short sentences and further split them into subwords, add a segmentation marker at the end of the short sentence, and insert filler subwords before the segmentation marker of the short sentence so that each short sentence has the same number of subwords; Map subwords to numbered vectors based on the vocabulary, generate position vectors based on the order of the subwords in the document, and superimpose the subword number vectors to generate the subword embedding vectors; The embedding vectors of each subword in the short sentence are concatenated and input into the Transformer encoder. After processing through a multi-head self-attention mechanism, residual connections, layer normalization, and a feedforward neural network, a context representation vector is generated. The context representation of each subword in the context representation vector is converted into a probability distribution for each subword through linear modulation and the Softmax function using a linear classifier. The BIN label with the maximum probability is assigned to the subword. The BIN labels of entity A include BA and IA, which represent the start and continuation words of entity A respectively. Non-entities only have the label NN. According to the principle of continuous merging, the subwords of the start word and the continuation word in the short sentence that are sequentially continuous and whose BIN tags correspond to the same entity are merged. The entities include roles, requirements and strategic goals.

3. The business system construction method according to claim 2, characterized in that: The principle of continuous consolidation includes the following steps: Based on the BIN tag test result of the first subword being a non-entity or the start word of an entity, a new test or a continued test is directly performed on the BIN tag of the second subword; If the BIN tag of the second subword is non-entity, then the BIN tag of the third subword is directly re-tested; If the BIN tag of the second subword is a continuation of the entity and is the same as the entity of the first subword, the two subwords are merged and the continuation check is continued on the BIN tag of the third subword; If the BIN tag of the second subword is the start word of the entity, the continuation test of the second subword is regarded as a new test, and a new continuation test is performed on the third subword; Stop after checking all subwords in the sentence.

4. The business system construction method according to claim 1, characterized in that: Using the attention polarity model to build a list of strategic elements involves the following steps: Convert strategic goals into strategic semantic vectors through a preset language model; The product of the strategic semantic vector, the internal query vector, and the external query vector is converted into internal weight and external weight through the self-attention mechanism; The product of the strategic semantic vector, internal weight and external weight is concatenated and converted into the internal and external probability distribution of the strategic target through linear modulation and Softmax function; Convolution kernels of different window sizes are used to convolve with the strategic semantic vector and sequentially pass through the PReLU function and maximum pooling to generate semantic features of different scales. After splicing, the fully connected layer and the Softmax function are used to output the polarity probability distribution. Combine the internal and external probability distributions with the polarity probability distribution to construct the SWOT factor probability distribution of the strategic goal, and establish an edge between the SWOT factor with the highest probability and the strategic goal. The value of the edge is equal to the highest probability in the SWOT factor probability distribution. Group the strategic goals according to each SWOT factor and generate a list of strategic factors.

5. The business system construction method according to claim 1, characterized in that: The method of using a rule engine to connect roles and requirements with strategic goals to generate a stakeholder map includes the following steps: Establish the affiliation relationship of the role demand chain and the supporting relationship of the demand strategic goal chain, and determine that the roles and demands, and demands and strategic goals are in a many-to-many form; Each relevant entity is converted into a bag-of-words feature vector and a corresponding semantic vector through the TF-IDF converter and the preset language model, and then concatenated into an entity description vector; Arrange and combine each role and requirement, as well as requirement and strategic goal, build role-requirement pairs and requirement-strategy goal pairs and calculate the cosine similarity respectively. Establish attribution edges and support edges between roles and requirements, as well as requirements and strategic goals whose cosine similarity is greater than or equal to the similarity threshold, and generate a stakeholder map. The attribution edge value and support edge value are equal to the corresponding cosine similarity.

6. The business system construction method according to claim 1, characterized in that: The proximal strategy is used to optimize the learning network to generate the final level flowchart, which includes the following steps: The business strategy goals in the business document are converted into business strategy semantic vectors using a preset language model. The strategic conduction reward is designed to be equal to the cosine similarity between the current process state vector and the business strategy semantic vector multiplied by the conduction coefficient. The conduction coefficient is equal to the sum of the edge values ​​of the attribution and / or supporting edges that the role corresponding to the current process state vector passes through in the stakeholder graph to reach the business strategy goals. Based on the proximal strategy optimization algorithm, spatial filtering is performed based on the stakeholder graph. When selecting the next process element, process elements with roles or requirements that do not have direct or indirect edges connected to the roles and requirements in the current process element in the stakeholder graph are eliminated. Select the process state vector of a process element as the starting point, select the next process element to connect based on the proximal policy optimization algorithm and strategic conduction reward, and repeat this step until it stops; If the stop is due to traversing all process elements, a final-level flow chart is generated. If the stop is due to spatial filtering resulting in no process elements, the flow chart being constructed is discarded and a new process element is reselected as the starting point. From the generated final-level flowchart, select the output with the largest cumulative strategic conduction reward.

7. The business system construction method according to claim 1, characterized in that: The state encoding of the process elements in the process element table based on the demand weight table and the weighted strategic vector of each strategic goal includes the following steps: The text description of the process node and the role in each process element are converted into a process semantic vector and a role semantic vector through a preset language model; Based on the stakeholder map and demand weight table, the requirements and corresponding demand weights related to the roles in the process elements are obtained. The role-related requirements are generated into a demand semantic vector through a preset language model. Based on the corresponding demand weights, a weighted sum is performed and the role semantic vectors of the roles are superimposed to generate a weighted role vector. Search the stakeholder map for the strategic goals supported by the requirements of the process elements and obtain the corresponding weighted strategic vectors. Generate the strategic associated demand vector for each process element based on the weighted strategic vectors obtained by weighted aggregation of the supporting edge values ​​between the corresponding requirements and the strategic goals. The process semantic vector, weighted role vector and strategic association requirement vector of each process element are concatenated to generate the process state vector of each process element.

8. The business system construction method according to claim 1, wherein: In the stakeholder graph, the number of roles with belonging edges to each requirement is counted as the out-degree, and the out-degree of each requirement is divided by the out-degree sum of all requirements to generate the requirement weight of each requirement. Each strategic goal in the strategic element list is encoded into a strategic semantic vector through a preset language model. The requirements supporting each strategic goal are retrieved from the stakeholder graph and converted into requirement semantic vectors through a preset language model. Based on the corresponding requirement weights, a weighted sum is performed and superimposed with the strategic semantic vector of each strategic goal to generate a weighted strategic vector for each strategic goal.

9. The business system construction method according to claim 1, characterized in that: Pre-establish process node role matching rules in the AI ​​model, introduce all relevant role requirement chains in each matching pair of process node and role and integrate to build a template rule library, use the natural language processing capability of the AI ​​model to identify process nodes in business documents and match the template rule library, obtain the role requirement chain matching the process node, and use the preset language model to convert the text description of the process node and the requirements in each role requirement chain into process semantic vectors and requirement semantic vectors and calculate the cosine similarity, organize the role requirement chains and process nodes corresponding to the requirement semantic vector with the maximum cosine similarity into process elements, and combine the process elements of all process nodes to build a process element table. The AI ​​model uses the GPT4-Turbo model.

10. Business system building system, characterized by, It includes graph construction module, process extraction module, process generation module and integration construction module; The graph construction module uses a bidirectional non-discriminatory encoding model to identify roles, requirements, and strategic goals in the document, uses an attention polarity model to semantically encode strategic goals and aggregate and classify them into internal elements or external elements, combines polarity analysis to establish a list of strategic elements, and uses a rule engine to connect roles and requirements with strategic goals to generate a stakeholder graph; The process extraction module generates corresponding semantic vectors for roles, requirements, and strategic supports based on a preset language model, calculates the demand weight of each requirement in the stakeholder map, embeds the demand semantic vector of the requirements supporting each strategic goal into the strategic semantic vector of the strategic goal in the strategic element list according to the corresponding demand weight, generates a weighted strategic vector, constrains the AI ​​model based on the stakeholder map, extracts the process nodes of the business document, and associates them to generate a process element table; The process generation module encodes the state of the process elements in the process element table based on the demand weight table and the weighted strategic vector of each strategic goal, optimizes the learning network using a proximal strategy, sets action space filtering based on the stakeholder map, and connects the process elements based on the strategic transmission reward guidance to generate a final-level process diagram; The integrated construction module generates a business system by matching the process nodes in the final-level flowchart with the code segments of the code library according to a logic verification function, wherein the logic verification function is equal to the norm value of the intersection of the process semantic vector of the process node and the code semantic vector of the code segment divided by the norm value of the code semantic vector.

Images