ODRL strategy generation method based on large language model multi-agent architecture

Abstract

The invention relates to an ODRL strategy generation method based on a large language model multi-agent architecture, which is applied to a system of a multi-agent architecture in an orchestrator-worker mode, and comprises an orchestrator agent and a group of multi-category worker agents, the method comprises the following steps: receiving natural language strategy input of a user through an orchestrator agent, identifying a structure type of language logic of the natural language strategy, and selecting at least one worker agent to form a processing chain according to the structure type; and calling a worker agent in the processing chain, processing the natural language strategy, generating an ODRL strategy, and outputting the ODRL strategy.

Description

Technical Field

[0001] This invention relates to the fields of large language models and ODRL (Open Digital Rights Language), and in particular to an ODRL policy generation method based on a multi-agent architecture of a large language model. Background Technology

[0002] The Open Digital Rights Language (ODRL) is a key standard recommended by the W3C for describing the use strategies of data assets in a machine-readable manner, playing a central role in safeguarding data sovereignty and interoperability in the data space.

[0003] In various scenarios, there is a need to automatically translate complex rules described in Natural Language (NL) (such as legal clauses and contracts) into ODRL policies. Based on this, existing technologies have proposed various intelligent models to achieve this task. For example, early models were based on "ontology-guided generation" methods, or later optimization schemes introduced "self-correcting rules."

[0004] However, the inventors of this invention application discovered in their research that existing models have the following significant shortcomings and defects: (1) Insufficient ability to handle complex logical structures, making it difficult to guarantee the accuracy and completeness of the generation strategy; (2) Generation quality is difficult to guarantee: Due to the extreme scarcity of high-quality parallel training corpora for “natural language to ODRL”, existing models are unable to learn a sufficiently rich set of patterns, resulting in defects in both grammatical correctness and semantic fidelity of the generated ODRL strategies. (3) Lack of systematic solutions: Most existing work treats the conversion from NL to ODRL as a single generative task, while ignoring its inherent complex cognitive process. Summary of the Invention

[0005] To address the aforementioned issues, the present invention aims to provide an ODRL policy generation method based on a large language model multi-agent architecture. This method can systematically and reliably convert complex natural language policies into high-quality ODRL policies, thereby lowering the application threshold of ODRL and promoting its widespread deployment in the data space.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, this application provides an ODRL policy generation method based on a large language model multi-agent architecture, applied to a multi-agent architecture system with an orchestrator-worker pattern, including an orchestrator agent and a group of multi-class worker agents; the method flow includes: The orchestrator agent receives the user's natural language policy input, identifies the structural type of the language logic of the natural language policy, and selects at least one worker agent to form a processing chain based on the structural type. The worker agent in the processing chain is invoked to process the natural language policy, generate the ODRL policy, and output it.

[0007] In one implementation, the structure type of the language logic of the natural language strategy includes: simple structure, parallel structure, and recursive structure.

[0008] In one implementation, the categories of worker agents include: rewriter agents, splitter agents, and ODRL generators; The rewriter agent is used to resolve the dependencies between policy units and eliminate the reference structure by rewriting, thus laying the groundwork for subsequent policy splitting. The splitter agent is used to split the content of parallel relations into multiple independent policy units for the natural language policy of the parallel structure. The ODRL generator is used to transform simple policy units, or multiple independent policy units generated by the rewriter agent and the splitter agent, into ODRL policies with preset quality standards.

[0009] In one implementation, the preset quality standard of the ODRL generator is monitored by integrating two optimization strategies: (1) Validator-based syntax optimization strategy; (2) Semantic optimization strategy based on LoRA enhancement.

[0010] In one implementation, the quality of the generated ODRL strategy is guaranteed by an iterative cycle of generation-verification-correction in the optimization strategy of (1).

[0011] In one implementation, in the optimization strategy of (2), key semantic elements are extracted from the input of the original natural language strategy based on an expert model, and after generating the ODRL strategy, the key semantic elements are checked during alignment verification.

[0012] In one implementation, the processing chain includes the following types: (1) ODRL generator; (2) splitter agent and ODRL generator; (3) rewriter agent, splitter agent and ODRL generator.

[0013] Secondly, an ODRL policy generation system based on a large language model multi-agent architecture is provided, including an orchestrator agent and a group of multi-class worker agents: The categories of worker agents include: rewriter agents, splitter agents, and ODRL generators; The rewriter agent is used to integrate content with recursive relationships into multiple independent policy units for natural language policies with recursive structures. The splitter agent is used to split the content of parallel relations into multiple independent policy units for the natural language policy of the parallel structure. The ODRL generator is used to convert simple policy units, or multiple independent policy units generated by the rewriter agent and the splitter agent, into ODRL policies with preset quality standards. The orchestrator agent receives the user's natural language policy input, identifies the structure type of the language logic of the natural language policy, selects at least one worker agent to form a processing chain based on the structure type, calls the worker agents in the processing chain to process the natural language policy, generates an ODRL policy, and outputs it.

[0014] The present invention has the following advantages due to the adoption of the above technical solutions: 1. Significantly Improves the Processing Capability and Accuracy of Complex Policies: The multi-agent architecture of this invention overcomes the bottleneck of single-agent models in handling complex logic by fundamentally overcoming the limitations of single-agent models through task decomposition and the use of specialized rewriter and splitter agents to preprocess recursive and parallel structures. Experimental results (as shown in Table 1 of the paper) demonstrate that, compared to the existing state-of-the-art (SOTA) methods, this invention improves the average semantic score across all use cases by 14.52%, with minimal performance degradation when handling complex use cases, exhibiting excellent stability.

[0015] 2. Ensures high quality of the generated strategy at the semantic and syntactic levels: Through a closed-loop syntactic correction strategy of "generation-verification-correction," this invention ensures that the syntactic compliance of the output ODRL strategy reaches a near-perfect level (over 99%). Simultaneously, the innovative LoRA-enhanced semantic reflection mechanism, by comparing with "semantic checkpoints" extracted by expert models, effectively guarantees a high degree of consistency between the generated content and the original intent, solving the semantic drift problem.

[0016] 3. Intelligent and efficient processing flow is achieved: The orchestrator agent can dynamically analyze input complexity and select the optimal processing path, avoiding the waste of resources by using the most complex process for all use cases. Experimental data (as shown in Table 2 of the paper) demonstrates that this intelligent orchestration workflow achieves near-theoretical optimal performance while consuming fewer computational resources (e.g., lower token consumption) than a fixed complex workflow, achieving an effective balance between performance and efficiency.

[0017] 4. Improved interpretability and maintainability: The modular agent design ensures that each component has a single and clearly defined function, and the entire transformation process is broken down into clear steps, which greatly enhances the interpretability and maintainability of the system. When optimization is needed for a specific type of transformation problem, only the corresponding "worker" agent needs to be upgraded, without modifying the entire system. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the network architecture of the system model and its input / output data flow in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0020] To address the problems arising from the single-architecture model processing of ODRL policy generation in existing technologies, an agent architecture is adaptively proposed. Specifically, this invention provides an ODRL policy generation method based on a large language model multi-agent architecture, applied to a multi-agent architecture system with an orchestrator-worker pattern, including an orchestrator agent and a group of multi-class worker agents; the method flow includes: The orchestrator agent receives the user's natural language policy input, identifies the structural type of the language logic of the natural language policy, and selects at least one worker agent to form a processing chain based on the structural type. The worker agent in the processing chain is invoked to process the natural language policy, generate the ODRL policy, and output it.

[0021] The following is based on the appendix of this application. Figure 1 Further details will be provided.

[0022] This invention proposes a multi-agent ODRL policy generation system based on a large language model, named AgentODRL. The system adopts a modular "Orchestrator-Workers" architecture, decomposing complex transformation tasks into processes completed collaboratively by multiple specialized agents.

[0023] System Architecture: Orchestrator-Worker Pattern This system consists of a central Orchestrator Agent and a group of specialized Worker Agents.

[0024] The orchestrator agent acts as the cognitive center of the system, responsible for receiving the user's natural language policy input. Figure 1 (This example uses only policy use cases). Its primary function is to quickly analyze and classify the complexity of the input, and then, based on the classification results, dynamically determine which worker agents to invoke and orchestrate an optimal processing path.

[0025] like Figure 1 As illustrated, the input natural language strategy can be classified by the orchestrator into complex structures (such as complex use cases of policy use cases) and non-complex use cases (such as simple use cases of policy use cases) based on the complexity of its language logic.

[0026] Complex use cases involve multiple policy units within a surface-level policy strategy that exhibit relatively complex relationships, such as parallel listings or recursive relationships with cross-references. For example, in a policy use case with a parallel structure, there are several use case units. In a recursive policy use case, at least two use case units have inter-clause reference relationships.

[0027] From the perspective of linguistic logic complexity, a simple policy input can be considered as having only one policy unit. To illustrate this with a policy use case, the input policy use case contains only one use case unit.

[0028] The worker agent consists of three core specialized agents, each designed to handle a specific subtask: The Rewriter Agent is specifically designed to handle use cases with recursive structures, eliminating referential dependencies between policies, flattening policies, and preparing for the splitting of subsequent policy units. It can identify and resolve cross-references in policies (such as "according to Article X..."), and through "structure-preserving inlining" techniques, embed the referenced clause content into the current clause, thereby eliminating semantic dependencies and transforming it into multiple independent policy units.

[0029] The Splitter Agent is specifically designed to handle use cases with parallel structures. It logically decomposes the input text based on changes in core semantic elements (such as changes in core assets, stakeholder relationships, or the fundamental purpose of the strategy), breaking down a multifaceted use case into multiple independent, self-contained strategy units in preparation for subsequent generation.

[0030] ODRL Generator Agent: This is the core execution unit of the entire workflow, responsible for converting well-structured rule text (preprocessed by a rewriter or splitter) into high-quality ODRL policies.

[0031] Based on the aforementioned agents and strategies, this system can dynamically combine different workflows according to the input: Simple path: Input → Orchestrator → Generator → Output Parallel path: Input → Orchestrator → Splitter → [Multiple policy units] → Generator → [Multiple ODRL policies] → Output Recursive path: Input → Orchestrator → Rewriter → Splitter → [Multiple policy units] → Generator → [Multiple ODRL policies] → Output In this way, the present invention decouples a complex, end-to-end conversion task into a series of well-defined subtasks that are efficiently processed by specialized intelligent agents, and achieves optimal overall performance and quality through intelligent orchestration. Furthermore, to ensure the quality of the ODRL policy output by the generator, this invention integrates two key optimization strategies: (1) Validator-based syntax optimization strategy: Introduce an iterative loop of “generate-validate-correct”.

[0032] Generation: The generator first produces an initial ODRL policy.

[0033] Validation: The strategy is submitted to a built-in validator implemented using the PYSHACL library, which performs rigorous syntax validation based on the ODRL official information model and vocabulary.

[0034] Correction: If validation fails, a detailed error report will be sent to the generator's LLM, prompting it to reflect and correct the error. This cycle will continue until the strategy passes validation or reaches the preset maximum number of attempts, thus ensuring that the final output syntax is near perfect.

[0035] (2) LoRA-enhanced semantic optimization strategy: An innovative reflection-based semantic optimization process is adopted.

[0036] Expert Model: Using the Low Rank Adaptation (LoRA) technique, a lightweight LLM (such as Qwen3-4B-Instruct) is fine-tuned into a high-precision semantic extraction "expert verification model" on a dataset containing 2,380 synthetic samples.

[0037] Semantic checkpoints: This expert model is responsible for extracting key semantic elements (such as parties, assets, behaviors, etc.) from the original NL input to form a structured "semantic checkpoint list".

[0038] Alignment verification: After generating ODRL, the main generator agent must check its output against this checkpoint list to ensure that each semantic point is accurately encoded, thereby guaranteeing semantic fidelity.

[0039] The effectiveness of the system and method in this application is illustrated below based on actual calculation examples.

[0040] Calculation example: 1. Dataset Construction Since there is a lack of standard benchmark datasets for ODRL generation in existing technologies, this invention constructs a novel dataset oriented towards data space scenarios, as follows: 1) First stage: Seed case construction.

[0041] We first manually constructed 70 seed use cases. These use cases were divided into three categories based on their logical structure: simple structure (40), parallel structure (20), and sequential dependency (i.e., recursive) structure (10).

[0042] The use cases are sourced from diverse sources, including academic literature, draft industry standards, Creative Commons license agreements, and logical templates from complex legal texts such as GDPR and CCPA, ensuring the representativeness and authenticity of the use cases.

[0043] 2) Second phase: LLM-based data augmentation.

[0044] We augmented 70 seed use cases using a large language model (Gemini 2.5 Pro) to expand the dataset size.

[0045] The enhancement process follows two main principles: core logic preservation (ensuring that the fundamental structure of the strategy, such as rule types and behavior-goal binding, remains unchanged) and context element transformation (diversifying specific elements such as stakeholders, data assets, and use cases to simulate real-world business changes).

[0046] After enhanced and rigorous human review, a high-quality dataset containing 770 test cases was finally constructed. This includes 440 simple test cases, 220 parallel relationship test cases, and 110 sequence dependency test cases. This dataset provides a solid foundation for the evaluation of this invention and future related research.

[0047] 2. Experimental Setup Evaluation indicators: 1) Grammar Score: This metric measures whether the generated ODRL policy strictly adheres to the specifications of the W3CODRL information model and vocabulary. We evaluate it using an automated validation workflow based on SHACL (Shapes Constraint Language). First, we build a comprehensive SHACL constraint rule set based on the official ODRL standard; then, we validate each generated ODRL policy using this rule set and count the number of constraint violations (N_errors). The formula for calculating the grammar score is as follows: Score_grammar(%) = (1 - N_errors / N_constraints) × 100% Where N_constraints is the total number of constraints in the SHACL rule set, and N_errors is the total number of errors reported by the validator.

[0048] 2) Semantic Score: This metric quantifies the extent to which the generated ODRL policy faithfully reflects the complete intent of the original natural language rules. Evaluation is performed through an automated workflow using an LLM-based "jury." First, a dedicated "semantic point identifier" LLM extracts a "semantic checkpoint list" as baseline facts from the original natural language use cases. Then, a "jury" consisting of two independent LLMs (K=2) compares the generated ODRL policy against this checkpoint list. Each reviewer scores the accuracy of the reflection for each semantic unit (S_ij). The final semantic score is the average of the scores from all reviewers across all semantic units. The calculation formula is as follows: Score_semantic(%) = (1 / N_units) * Σ_{i=1}^{N_units} [ (1 / K) * Σ_{j=1}^{K} S_ij ] × 100% Where N_units is the total number of semantic units, K is the number of reviewers, and S_ij is the score given by the j-th reviewer for the i-th semantic unit.

[0049] Experiment 1: Evaluation of ODRL generation strategies.

[0050] Objective: To progressively evaluate the effectiveness of the various enhancement strategies proposed in this invention compared to baseline methods.

[0051] The strategies are compared below, with the first two serving as baseline strategies: 1) Ontology-Guided Strategy (OGS): Baseline approach, using ODRL v2.2 ontology-guided LLM generation.

[0052] 2) SCR strategy: The state-of-the-art (SOTA) method in the existing technology, which uses predefined rules to modify the LLM output based on OGS.

[0053] 3) Semantic-Enhanced Strategy: Some of the solutions proposed in this invention only integrate the LoRA fine-tuning model to enhance semantic fidelity.

[0054] 4) Complete AgentODRL Process (AOFP): The complete integrated system of this invention combines semantic enhancement strategies and validator-based syntax correction strategies.

[0055] Test Model: Experiments were conducted on the GPT-4.1 series models (GPT-4.1, GPT-4.1-mini, GPT-4.1-nano). To ensure the robustness of the results, each experimental configuration was run three times and the average value was taken. The results of Experiment 1 are shown in Table 1. The best results are indicated in bold, and the second best results are indicated in italics and underlines.

[0056]

[0057] Table 1 Therefore, it can be seen that: 1) The complete process of this invention (AOFP) significantly outperforms existing technologies across all dimensions. Compared to the strongest baseline method (SCR), AOFP improves the average grammar score by 5.39% and the average semantic score by 14.52% across all models. The grammar correction strategy of this invention is extremely effective, improving the grammar score of all models to over 99% and essentially eliminating grammatical errors.

[0058] 2) The enhancement strategy of this invention has universality and can significantly improve the performance lower limit of different capability models. For example, it improves the semantic score of GPT-4.1-nano on recursive use cases from 34.87 to 61.53, an increase of 76.46%.

[0059] Experiment 2: Overall performance evaluation of the orchestrator-worker workflow.

[0060] Objective: To verify the overall efficiency and intelligence level of the adaptive "arranger-worker" workflow proposed in this invention.

[0061] Experimental procedure: 1) Establish performance benchmarks: By forcing all use cases to follow a fixed workflow path (such as "generator only", "splitter → generator", "rewriter → splitter → generator"), determine the "theoretical performance ceiling" for use cases of different complexities.

[0062] 2) Evaluate automated workflows: Test fully automated "orchestrator-worker" workflows to see if they can approach the theoretical ceiling in terms of performance through dynamic analysis and intelligent routing without human intervention.

[0063] Test Model: To more clearly observe the performance gains, this experiment selected GPT-4.1-nano, which has significant room for performance improvement, as the executing agent of the "generator." The upstream agents (orchestrator, splitter, and rewriter), which require higher analytical capabilities, used the more powerful GPT-4.1 model to maximize the module's capabilities. To ensure the robustness of the results, each experimental configuration was run three times and the average value was taken. The results of Experiment 2 are shown in Table 2, where the best-performing path values ​​for each type of use case are indicated in bold.

[0064]

[0065] Table 2 Therefore, it can be seen that: 1) Specialized worker agents are of immense value. Experiments have shown that "splitters" and "rewriters" designed for specific complexities are key to performance improvements. For example, when processing parallel structures, adding a "splitter" can improve the semantic score from 69.44 to 84.88. This validates the core principle of precisely aligning processing paths with task complexity.

[0066] 2) The orchestrator agent achieves an intelligent balance between performance and efficiency. The automated "orchestrator-worker" workflow achieved an average semantic score of 80.22, far exceeding the 72.35 score of the "generator-only" workflow. More importantly, while achieving near-theoretical optimal performance, it consumes fewer computational resources (e.g., lower token consumption) than a fixed workflow that forces all tasks to follow the most complex paths. This demonstrates that the orchestrator of this invention successfully achieves a dynamic balance between near-optimal performance and significantly reduced computational costs, making it a practical and efficient real-world solution. In the embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection between devices or units through some interfaces, and may be electrical, mechanical, or other forms.

[0067] The integrated units implemented as software functional units described above can be stored in a computer-readable storage medium. These software functional units, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0068] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating ODRL policies based on a large language model multi-agent architecture, characterized in that, A system with a multi-agent architecture applied to the orchestrator-worker pattern, comprising an orchestrator agent and a set of multi-class worker agents; The method flow includes: The orchestrator agent receives the user's natural language policy input, identifies the structural type of the language logic of the natural language policy, and selects at least one worker agent to form a processing chain based on the structural type. The worker agent in the processing chain is invoked to process the natural language policy, generate the ODRL policy, and output it.

2. The ODRL policy generation method based on a large language model multi-agent architecture according to claim 1, characterized in that, The structural types of the language logic of the natural language strategy include: simple structure, parallel structure, and recursive structure.

3. The ODRL policy generation method based on a large language model multi-agent architecture according to claim 2, characterized in that, The categories of worker agents include: rewriter agents, splitter agents, and ODRL generators; The rewriter agent is used to parse the dependencies between policy units of a recursive natural language policy and eliminate the reference structure therein through rewriting. The splitter agent is used to split the content of parallel relations into multiple independent policy units for the natural language policy of the parallel structure. The ODRL generator is used to transform simple policy units, or multiple independent policy units generated by the rewriter agent and the splitter agent, into ODRL policies with preset quality standards.

4. The ODRL policy generation method based on a large language model multi-agent architecture according to claim 3, characterized in that, The preset quality standard of the ODRL generator is monitored by integrating two optimization strategies: (1) Validator-based syntax optimization strategy; (2) Semantic optimization strategy based on LoRA enhancement.

5. The ODRL policy generation method based on a large language model multi-agent architecture according to claim 4, characterized in that, In the optimization strategy described in (1), the quality of the generated ODRL strategy is guaranteed by an iterative cycle of generation-verification-correction.

6. The ODRL policy generation method based on a large language model multi-agent architecture according to claim 4, characterized in that, In the optimization strategy described in (2), key semantic elements are extracted from the input of the original natural language strategy based on the expert model. After generating the ODRL strategy, the key semantic elements are checked during the alignment verification.

7. The ODRL policy generation method based on a large language model multi-agent architecture according to claim 3, characterized in that, The types of processing chains include: (1) ODRL generator; (2) splitter agent and ODRL generator; (3) rewriter agent, splitter agent and ODRL generator.

8. An ODRL policy generation system based on a large language model multi-agent architecture, characterized in that, It includes a scheduler agent and a set of multi-class worker agents: The categories of worker agents include: rewriter agents, splitter agents, and ODRL generators; The rewriter agent is used to integrate content with recursive relationships into multiple independent policy units for natural language policies with recursive structures. The splitter agent is used to split the content of parallel relations into multiple independent policy units for the natural language policy of the parallel structure. The ODRL generator is used to convert simple policy units, or multiple independent policy units generated by the rewriter agent and the splitter agent, into ODRL policies with preset quality standards. The orchestrator agent receives the user's natural language policy input, identifies the structure type of the language logic of the natural language policy, selects at least one worker agent to form a processing chain based on the structure type, calls the worker agents in the processing chain to process the natural language policy, generates an ODRL policy, and outputs it.

Images