A production-study-research cooperation promotion method

Abstract

The application discloses a kind of production-study-research collaborative promotion method, the present application relates to artificial intelligence and production-study-research collaborative achievement transformation technical field, this method carries out structured information extraction to university scientific and technological achievement text and enterprise demand text, constructs function-effect-constraint causality graph, carries out node embedding coding and cross-domain similarity matching using graph neural network, finds the implicit application value between achievement and demand, models collaborative promotion network as Markov decision process, generates and adaptively reconfigures promotion relay workflow using strategy network dynamically, calculates technical scheme evolution cost based on graph edit distance, quantifies the knowledge contribution weight of each collaborator and drives benefit distribution, the present application effectively alleviates the supply and demand semantic gap, workflow rigidification and the problem that contribution degree is difficult to quantify, improves cross-domain matching accuracy and promotion conversion efficiency.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and the transformation of research results from industry-academia-research collaboration, specifically a method for promoting industry-academia-research collaboration. Background Technology

[0002] Industry-academia-research collaborative promotion refers to the process by which universities and research institutes transform basic scientific research results into actual enterprise productivity through a series of steps, including concept verification, pilot-scale amplification, and investment and financing matchmaking. During this process, the technological maturity of the results gradually increases from a lower level to the industrial application stage, requiring multiple entities to work together to complete engineering amplification, capital injection, and market connection. The closest existing technical solution to this invention is an industry-academia-research matchmaking platform based on keyword retrieval. The system structure of this platform includes a results release module, a demand release module, a search and matching server, and a user interaction terminal. University users upload scientific and technological results to the results database in the form of text abstracts through the results release module, while enterprise users upload technical pain points to the demand database in the form of natural language text through the demand release module. The search and matching server receives the keywords entered by the enterprises, uses a text similarity algorithm to search the results database, and returns the matching degree ranking list to the user interaction terminal for secondary screening by humans. The preliminary matching results determined by the screening enter the offline negotiation stage, where humans drive the subsequent concept verification, pilot-scale amplification, and investment and financing introduction. There is no information-based collaborative link between the various stages, and the evolution of intellectual property rights during the technology transfer process is not recorded or tracked. Summary of the Invention

[0003] The technical problem that this invention aims to solve is that existing industry-academia-research collaborative promotion systems only perform surface-level keyword or phrase similarity calculations between the result text and the demand text, without conducting in-depth causal deconstruction of the underlying physicochemical effects of the result and the engineering constraints of the demand, resulting in the inability to effectively discover and match results with cross-disciplinary application value.

[0004] The existing system uses a static shelf-style display for the promotion workflow, which lacks the ability to dynamically and adaptively orchestrate the complex collaborative network consisting of heterogeneous nodes such as universities, proof-of-concept centers, pilot production bases, venture capital institutions, and leading enterprises. Once a pilot production node fails to verify or is blocked by resources, the entire promotion chain comes to a standstill.

[0005] Furthermore, existing technologies do not record or quantify the evolution of technical features of achievements in each stage of collaborative promotion, making it impossible for collaborating parties to objectively calculate the contribution ratio of original intellectual property rights in derivative technical solutions, which can easily lead to disputes over the distribution of benefits.

[0006] To address the aforementioned technical problems, this invention provides a method for collaborative promotion between industry, academia, and research, comprising: extracting structured information from texts of scientific and technological achievements from universities and texts of enterprise needs, obtaining sets of functional-effect nodes on the achievement side and sets of constraint nodes on the demand side, and establishing directed edge relationships according to three types of attributes: function, effect, and constraint, to form a causal graph of function-effect-constraint.

[0007] A graph neural network is used to aggregate and embed the neighborhood information of the nodes in the causal graph to obtain the embedding vector of each node.

[0008] Calculate the similarity between the embedding vector of nodes on the result side and the embedding vector of nodes on the demand side, and determine the node pairs with similarity higher than a preset threshold as matching result-demand pairs.

[0009] Using this pairing as the initial state, a state vector is constructed that includes the technology maturity level, remaining available conversion funds, consumed conversion time, and the resource idleness of candidate collaborative nodes. The strategy network is used to output the next conversion target node type and generate a promotion relay workflow sequence. When an abnormal event is detected in the current conversion target node, the state corresponding to the abnormal event is re-input into the strategy network, and the alternative target node type is output and the subsequent workflow sequence is reconstructed.

[0010] The technical documents corresponding to the original technical solution and each modified technical solution are structured and parsed to construct a technical feature tree. The graph edit distance between adjacent versions of the technical feature tree and the graph edit distance of the original technical feature tree relative to the empty tree are calculated. Based on this, the contribution weight of the original technical solution and each collaborating party in the derived technical solution is calculated, and the benefit distribution is performed according to the contribution weight.

[0011] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0012] 1. This invention overcomes the limitations of traditional keyword surface matching methods in failing to reveal underlying causal relationships across disciplines by constructing a function-effect-constraint causal graph and combining it with graph neural networks for cross-domain topological similarity matching, thereby improving the coverage and matching accuracy of cross-domain results discovery.

[0013] 2. This invention models the collaborative promotion network as a Markov decision process and uses a policy network to perceive node status in real time, automatically generate and reconstruct promotion paths, thereby avoiding the problem of the entire conversion chain being stalled due to the abnormality of a single pilot node, and improving the continuity of the promotion process and conversion efficiency.

[0014] 3. This invention achieves objective, transparent, and traceable accounting of the contribution ratio between original academic achievements and derived technical solutions through an evolutionary contribution quantification method based on graph edit distance. This provides a quantifiable basis for the distribution of benefits among collaborating parties and reduces disputes over the distribution of benefits among multiple parties. Attached Figure Description

[0015] Figure 1 This is a general flowchart of the method in the embodiments of the present invention;

[0016] Figure 2 This is a flowchart of the function causal graph construction and cross-domain matching in an embodiment of the present invention;

[0017] Figure 3 This is a schematic diagram illustrating the dynamic workflow orchestration principle of reinforcement learning in an embodiment of the present invention. Detailed Implementation

[0018] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0019] The technical solution of the present invention will be further described below with reference to specific embodiments. The industry-university-research collaborative promotion method provided by the present invention includes three parts: construction and cross-domain matching of function-effect-constraint causal graph, dynamic workflow orchestration based on reinforcement learning, and knowledge contribution quantification based on evolution distance. The three parts respectively undertake the functions of deep matching of results and needs, dynamic generation and adaptive reconstruction of promotion path, and quantitative tracking of knowledge contribution during the collaboration process, and together constitute a complete methodological closed loop from result discovery to benefit distribution.

[0020] Regarding the construction and cross-domain matching of function-effect-constraint causal graphs, the structured information of patent texts and academic papers of university scientific and technological achievements is first extracted. Zero-sample triple extraction is then performed using a large language model to extract three types of information: physical entities, underlying physicochemical effects, and applicable conditions. This forms a set of function-effect nodes on the achievement side, denoted as... ;

[0021] Simultaneously, structured information is extracted from the enterprise's requirement text. A large language model is used to extract engineering constraints, performance requirements, and pain point descriptions from the enterprise's actual production scenarios, forming a set of constraint nodes on the requirement side, denoted as... Then, directed edge relationships are established for the two types of node sets according to the three attributes of function, effect, and constraint, forming a unified function-effect-constraint causal graph. This graph is then stored in a graph database, and each node in the graph carries its original text description and attribute label.

[0022] Furthermore, the GraphSAGE graph neural network algorithm is used to aggregate and embed the neighborhood information of each node in the causal graph, mapping the nodes to a unified low-dimensional embedding space to obtain the embedding vector representation of each node. The node update formula is as follows:

[0023] ;

[0024] in, Represents a node In the Embedded vectors after layer aggregation;

[0025] Represents a node In the The embedding vector of the layer, when hour This represents the initial feature vector of the node;

[0026] Represents a node The set of first-order neighbor nodes in the graph;

[0027] This represents the aggregation function performed on the embedding vectors of neighboring nodes; in this embodiment, the mean aggregation method is used.

[0028] This represents a vector concatenation operation;

[0029] Indicates the first The learnable weight matrix of the layer;

[0030] This represents a non-linear activation function.

[0031] Furthermore, the cosine similarity between the embedding vectors of the result-side nodes and the embedding vectors of the demand-side nodes is calculated as a quantitative indicator of their cross-domain matching degree. The calculation formula is as follows:

[0032] ;

[0033] in, This represents a function-effect node on the outcome side. This represents a constraint node on the demand side.

[0034] , These are the embedding vectors of the two vectors after the above embedding encoding;

[0035] , Let represent the Euclidean norm of the vectors respectively;

[0036] The value ranges from negative one to positive one. The larger the value, the higher the matching degree between the two at the underlying causal level. The result-demand node pairs with similarity higher than the preset threshold are sorted in descending order of similarity value, and a cross-domain recommendation list is output to complete the accurate promotion and matching of basic results to actual application scenarios in unrelated industries. If the matching result of a certain round does not meet the preset threshold condition, the edge relationship of the causal graph is supplemented and expanded and the similarity is recalculated to form a closed-loop judgment logic.

[0037] Regarding dynamic workflow orchestration based on reinforcement learning, the collaborative generalization network is modeled as a Markov decision process, with state vectors in the state space. At any moment Technology readiness level including current achievements Remaining available conversion funds Conversion time already consumed and the resource idleness vector of candidate cooperating nodes Actions in action space This indicates the type of the next conversion target node selected in the current state. Candidate collaborative node types include proof-of-concept centers, pilot production bases, venture capital institutions, and leading enterprises. The reward function is used to evaluate the contribution of the selected action to the promotion efficiency, and the calculation formula is as follows:

[0038] ;

[0039] in, Indicates the execution of an action The extent to which the post-implementation technology maturity level has improved compared to the pre-implementation level;

[0040] This indicates the time cost of performing the action;

[0041] This indicates the financial cost incurred in performing the action;

[0042] , , The weighting coefficients for the three indicators—technology maturity improvement, time cost, and capital cost—are derived from regression calibration of historical transaction data, and in this embodiment, the values ​​are specific and determined values.

[0043] An offline training sample set was constructed based on massive historical industry-academia-research transaction logs. The policy network was pre-trained using a near-end policy optimization algorithm. The optimization objective of the policy network was to maximize the pruned objective function.

[0044] ;

[0045] in, Represents the trainable parameters of the policy network;

[0046] Indicates the state of the new and old strategies. Select action The probability ratio;

[0047] Indicates time The advantage function estimate is used to measure the quality of the selected action relative to the average level.

[0048] This means cropping the ratio to a range. The clipping function;

[0049] These are the preset clipping hyperparameters;

[0050] This represents the expectation operation for the time step.

[0051] Deploy the pre-trained policy network to the online promotion system and receive its current status in real time. Output the target node type to which the next step should be transformed. Based on this output, the system automatically generates a promotion relay workflow sequence. During the workflow execution, the system continuously monitors the verification progress and resource usage status of each pilot node. If an abnormal event such as verification failure, funding shortage, or resource congestion is detected at a certain node, the status corresponding to the abnormal event is updated to the new current status, re-inputted into the policy network, triggering path replanning, and outputting the alternative node and the reconstructed subsequent workflow path, thereby realizing adaptive breakpoint continuation of the promotion process.

[0052] Regarding the quantification of knowledge contribution based on evolution distance, at the initial stage of collaborative promotion, the technical documents, code, and drawings corresponding to the original academic achievements provided by universities were structurally analyzed. Using abstract syntax trees and natural language dependency parsing methods, an original technical feature tree was constructed, denoted as... During subsequent pilot-scale testing and secondary development, whenever a collaborating party makes a substantial modification to the technical solution and submits a new version, the corresponding technical documents, code, and drawings are re-analyzed in a structured manner to construct the technical feature tree for that modification, denoted as follows: For the technical feature trees of two adjacent versions, the graph edit distance algorithm is used to calculate the cost of node addition, deletion, and replacement operations required to transform the feature tree of the previous version into the feature tree of the next version, denoted as . ,in ;

[0053] Simultaneously calculate the original technical feature tree itself relative to an empty tree Construction cost This value represents the amount of initial characteristic information carried by the original academic achievement itself.

[0054] Based on the evolution distance of each stage mentioned above, the contribution weight of the original academic achievements in the derived technical solutions is calculated using the following formula:

[0055] ;

[0056] in, This indicates the contribution weight of the original academic findings in the derived technological solution;

[0057] This represents the initial construction cost of the original feature tree;

[0058] This indicates that all collaborating parties will subsequently... The sum of the cumulative evolution costs generated in the next iteration modification, for participants in the ... The contribution weight of each collaborating party in this modification is calculated based on the proportion of its corresponding evolution cost to the total cost. The calculation formula is as follows:

[0059] ;

[0060] in, Indicates the first The contribution weight of the collaborating party corresponding to this modification has the same meaning in the denominator as described above. As can be seen from the above definition, With each The sum of these values ​​is always one, thus ensuring the numerical closure integrity of the contribution weight allocation. The contribution weights of each party calculated above are synchronized to the smart contract interface in real time. The smart contract automatically executes the benefit allocation according to the preset allocation rules, and automatically re-triggers the above calculation process after each feature tree update, realizing dynamic, transparent, and traceable updates of contribution.

[0061] In another implementation, the cross-domain matching part can directly utilize a pre-trained large language model to perform overall semantic vector encoding on the result text and the demand text, store the encoding results in a vector database, and calculate cross-domain semantic similarity using an approximate nearest neighbor retrieval algorithm. This replaces the graph neural network embedding and graph topology matching steps, and can still achieve the discovery of implicit cross-domain application value. In yet another implementation, the dynamic workflow orchestration part can pre-construct a rule table containing technology maturity thresholds, funding gap thresholds, and node congestion trigger conditions. The system reads the current state in real time and matches according to the priority conditions in the preset rule table, outputting the next recommended node and path reconstruction scheme. This replaces the Markov decision process modeling and policy network training steps, and can still achieve dynamic orchestration and adaptive reconstruction of the promotion workflow. The above alternative implementations can be combined with the knowledge contribution quantification part of this invention to form different technical solution combinations, all of which can achieve the invention's objective.

[0062] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A collaborative promotion method involving industry, academia, and research, characterized in that, include: The steps for constructing a function-effect-constraint causal graph are as follows: Structured information extraction is performed on the text of university scientific and technological achievements to obtain a set of function-effect nodes on the achievement side; structured information extraction is performed on the text of enterprise needs to obtain a set of constraint nodes on the demand side; directed edge relationships are established between the set of function-effect nodes and the set of constraint nodes according to the three attributes of function, effect, and constraint to form a function-effect-constraint causal graph. Node embedding encoding steps: A graph neural network is used to aggregate neighborhood information and embed the nodes in the function-effect-constraint causal graph to obtain the embedding vector of each node; Cross-domain matching steps: Calculate the similarity between the embedding vector of the result-side node and the embedding vector of the demand-side node, and determine the result-side node and demand-side node with similarity higher than a preset threshold as matching result-demand pairs; Dynamic workflow orchestration steps: Taking the result-demand pairing as the initial state, construct a state vector containing the technology maturity level, remaining available conversion funds, consumed conversion time, and resource idleness of candidate collaborative nodes. Use a strategy network to output the next conversion target node type based on the current state vector. Generate a promotion relay workflow sequence based on the output target node type. During the generation of the promotion relay workflow sequence, if an abnormal event is detected in the current conversion target node, update the state corresponding to the abnormal event to a new state vector, re-input it into the strategy network, output a replacement target node type, and reconstruct the subsequent promotion relay workflow sequence. Knowledge contribution quantification steps: Perform structured analysis on the technical documents, code, and drawings corresponding to the original technical solutions provided by universities, and construct the original technical feature tree; The technical documents, code and drawings corresponding to the modified versions of the technical solutions submitted by each collaborating party during the collaborative promotion process are sequentially structured and parsed to construct the technical feature tree after each modification; Calculate the graph edit distance between two adjacent versions of the technical feature tree, and the graph edit distance of the original technical feature tree relative to the empty tree; calculate the contribution weight of the original technical solution and the corresponding collaborator in each modification in the derived technical solution based on the graph edit distance, and perform benefit allocation based on the contribution weight.

2. The method according to claim 1, characterized in that, In the step of constructing the function-effect-constraint causal graph, a large language model is used to extract zero-sample triplets from the patent text and academic paper text of scientific and technological achievements of universities, and extract three types of information, namely physical entities, underlying physicochemical effects and applicable conditions, to form a set of function-effect nodes. By using a large language model, structured information is extracted from enterprise requirement texts, and engineering constraints, performance requirements, and pain point descriptions are extracted to form a set of constraint nodes.

3. The method according to claim 1, characterized in that, The graph neural network used in the node embedding encoding step is the GraphSAGE network. The embedding vectors of the first-order neighbor nodes of each node are aggregated by mean. The aggregation result is concatenated with the embedding vector of the node itself in the previous layer. After weight matrix transformation and nonlinear activation function processing, the embedding vector of the node in the current layer is obtained.

4. The method according to claim 1, characterized in that, In the cross-domain matching step, cosine similarity is used to calculate the similarity between the embedding vector of the result-side node and the embedding vector of the demand-side node. Sort result-side nodes and demand-side node pairs with similarity scores higher than a preset threshold in descending order of similarity scores, and output a cross-domain recommendation list.

5. The method according to claim 1, characterized in that, In the dynamic workflow orchestration step, the policy network is pre-trained using a near-end policy optimization algorithm; During pre-training, a reward value is calculated based on the improvement in the technology maturity level of the result after performing the transformation action of the target node, the time cost of performing the action, and the financial cost of performing the action. The parameters of the policy network are then updated based on the reward value.

6. The method according to claim 1, characterized in that, The candidate collaborative nodes include proof-of-concept centers, pilot production bases, venture capital institutions, and leading enterprises; the abnormal events include node verification failures, funding shortages, and resource congestion.

7. The method according to claim 1, characterized in that, In the knowledge contribution quantification step, abstract syntax trees and natural language dependency parsing methods are used to perform structured parsing of technical documents, code and drawings, and to construct a technical feature tree carrying the functional dependency relationship between feature nodes.

8. The method according to claim 1, characterized in that, The calculation of the contribution weights of the original technical solution and the corresponding collaborators in each modification of the derived technical solution based on the edit distance of each graph includes: The graph editing distance of the original technical feature tree relative to the empty tree is determined as the evolution cost of the original technical solution; The graph editing distance between adjacent versions of the technical feature tree is determined as the evolution cost of the corresponding modification collaborator; The contribution weight of the original technical solution is obtained by dividing the evolution cost of the original technical solution by the sum of the evolution cost of the original technical solution and the evolution cost of each modification. The evolution cost of each modification is divided by the sum of the evolution cost of the original technical solution and the evolution cost of each modification to obtain the contribution weight of the collaborating party for the corresponding modification. The sum of all contribution weights is one.

9. The method according to claim 1, characterized in that, The step of distributing benefits based on the contribution weight includes: synchronizing the calculated contribution weight of the original technical solution with the contribution weight of each modification collaborator to the smart contract interface, and having the smart contract distribute benefits according to a preset distribution rule; whenever the technical feature tree is updated, the calculation of the contribution weight and the distribution of benefits are retried.

Images