Data sharing system and methods for training artificial intelligence

Abstract

The invention relates to a data sharing system (1) for transferring data to a third party in accordance with regulations concerning such transfer, comprising means for processing the data to be transferred by means of artificial intelligence. The data sharing system according to the invention is characterized by - a receiving interface (2) configured to receive a release case container comprising a data record, a data record description and a data record context; - a case database (3), comprising historical release case containers; - an evaluation module (4) configured to process the release case container and at least one historical release case container using a large language model (LLM), wherein the large language model (LLM) is trained to generate a compliance assessment; and - an output interface (5) designed to at least forward the data set to the third party taking into account the compliance assessment if the release case container meets the requirements at least up to a specified compliance threshold.

Description

[0001] The invention relates to a data release system for forwarding data to a third party according to the type defined in more detail in the preamble of claim 1 and to a method for training a large language model according to the type defined in more detail in the preamble of claim 8 and to a method for training a reinforcement learning model according to the type defined in more detail in the preamble of claim 9.

[0002] A company's business may require the exchange of data with third parties. For example, a vehicle manufacturer may collect operating data from its vehicles in the field and pass it on to a customer for analysis in either raw or processed form. For example, such data may describe the operating behavior of an internal combustion engine over its lifetime, which an engine control unit developer can use to optimize an engine management program.

[0003] The transfer of such data to third parties must be carried out with caution. This includes, for example, personal data or information relating to new technical and, in particular, unpublished ideas. Taking into account different domains, such as different markets, vehicle types, the type of data to be exchanged, the sender, the recipient, applicable law, and the like, the scope and manner of data exchange must be adapted. Rules governing data exchange are formulated in the form of regulations that must be adhered to by all parties involved. Such regulations can be based not only on laws but also on internal company policies.

[0004] Before such a data exchange can be approved, the corresponding approval case must be manually reviewed by a specialist to determine whether all regulations are being adhered to. Due to the increasing volume of data, the potentially complex legal framework that must be observed, and the context that varies greatly from case to case, this entails a very high level of manual effort. Granting approval therefore takes a long time and is associated with high costs. Furthermore, the complexity of such data exchange processes makes it increasingly impossible to understand all the relevant decision-making logic behind the decision to approve or reject a corresponding data exchange. If the appropriate review of such a data exchange case is not carried out with sufficient care, there is a risk that sensitive data will be leaked and / or applicable law will be violated. This must be avoided.

[0005] There is therefore a need to provide procedures and means that reduce the effort required for a data exchange process and increase the reliability that all relevant regulations are complied with.

[0006] US 2023 / 0315981 A1 discloses a method and system for assessing the quality of a document during a risk and compliance assessment. The system evaluates the quality of a document based on differences in content compared to a template. For this purpose, the document content is first examined using computational linguistics methods, also known as natural language processing (NLP). This allows a document-type-specific quality assessment model to be selected from a database and executed. The respective quality assessment model can be based on the use of artificial intelligence (AI). Documents to be assessed can include, for example, contracts, legal texts, guidelines, letters, operating instructions, or the like. These documents can be in Word or PDF format. They are exclusively text documents.

[0007] The present invention is based on the object of providing a data release system which makes it possible to automate data release processes while increasing the reliability that applicable regulations are complied with.

[0008] According to the invention, this object is achieved by a data release system having the features of claim 1. Advantageous embodiments and further developments as well as a method for training a large language model and a method for training a reinforcement learning model emerge from the dependent claims.

[0009] A generic data release system for the transfer of data to a third party in accordance with the respective regulations concerning the transfer, comprising means for processing the data to be transferred using artificial intelligence, is further developed according to the invention by - a receiving interface configured to receive a release case container comprising a data set, a data set description, and a data set context, wherein the data set description comprises a description of the data set and the data set context comprises a reason for forwarding the data set to a third party, each in computer-processable form; - a case database comprising historical release case containers, wherein the receiving interface is configured to store a received release case container in the case database; - an evaluation module configured to process the release case container and at least one historical release case container by means of a large language model, wherein the large language model is trained to generate a compliance assessment by comparing the release case container with the at least one historical release case container, comprising a description of the extent to which the release case container complies with the regulations relating to the transfer; and - an output interface configured to at least forward the data set to the third party taking into account the compliance assessment if the release case container complies with the regulations at least up to a specified compliance threshold.

[0010] The data release system according to the invention thus enables the automation of the data release and forwarding process. This allows existing work processes to be made more efficient and thus reduce costs. Since relevant regulations are automatically checked for compliance, it is reliably possible to identify and prevent corresponding violations of regulations or laws at an early stage.

[0011] The data release system is formed by one or more computing units. These can be all common types of computing units, such as in particular a desktop computer, a laptop, a server or server network, a system-on-a-chip, or the like. Due to the high computing requirements of large language models, high-performance computing clusters are particularly suitable for this purpose. The individual logical modules of the data release system according to the invention can also be distributed across multiple physical computing units. The data release system has corresponding interfaces for receiving and forwarding data. The data release system has read access to a computer-readable storage medium on which at least one computer program product is stored, the execution of which by a processor of the data release system causes the processor to execute corresponding method steps.Communication within the data release system and / or preferably via the receiving interface and / or output interface is preferably secured through the use of proven cryptographic encryption techniques.

[0012] The receive interface and output interface can be formed by one and the same physical interface. All common data transmission technologies and communication protocols can be used. In particular, data exchange is automated using proven application programming interfaces (APIs). The data release system receives a release case container via the receive interface. The release case container can be compiled manually by a user or automatically by a computer system. The release case container contains the data set to be exchanged with the third party, as well as a data set description and a data set context for each data set. The data set description and the data set context can be two separate computer-readable files.These files can be in any conceivable format and encoded in all common coding schemes. In particular, the dataset description and the dataset context contain the respective description or the reason for sharing in the form of a verbally formulated text. The information constituting the dataset can also take any conceivable file format. For example, it can be text files, a table such as an XML document, sensor data, photos, videos, CAD data, simulation data such as the results of a CFD simulation or an FEM simulation, the source code of a program, or the like.

[0013] The dataset description describes in particular what type of data is involved, what volume it has, what format it is in, how much storage space the respective dataset requires, and the like. The metadata of a computer-readable file can therefore be used to create or supplement the dataset description. In addition to the reason for the transfer, the dataset context can also describe the target market or domain within which the data exchange takes place. Additional information can, for example, describe the sender sending the dataset, the project within which the data exchange takes place, the recipient the dataset is intended for, and the like. In summary, the dataset description and the dataset context serve to describe all attributes relevant for a corresponding data exchange.

[0014] Historical release case containers are stored in the case database. Each time a release case container is received via the receive interface, it can be added to the case database and thus considered a new historical release case container.

[0015] The evaluation module uses artificial intelligence in the form of a generative pretrained transformer (GPT) with a large language model (LLM) to compare release case containers with historical release case containers. The large language model processes the data set, the data set description, and the data set context of the release case containers to be compared. It is also conceivable that the large language model supplements or corrects the data set description and / or the data set context depending on information detected in the data set.The large language model is capable of recognizing evaluation patterns in the historical release case containers, which enable the large language model to classify, based on the contents of the release case container, whether the release case container may be released for data exchange or must be rejected. The large language model does not need to know the relevant regulations itself. The result of the evaluation performed by the large language model is output in the form of the regulation compliance assessment. The regulation compliance assessment is also computer-processable information, in particular in the form of a separate computer-readable file. This file can also be available in any suitable file format.In particular, the compliance assessment includes a summary of the release case container's assessment by the large language model in the form of human-readable text. Based on the compliance assessment, a risk assessment can also be performed and an audit log can be generated, which summarizes the release case assessment process by the evaluation module. A compliance assessment can be added to a release case container.

[0016] The data release system is capable of automatically forwarding release case containers evaluated by the evaluation module to relevant third parties based on the extent to which the applicable regulations are complied with. To do so, the evaluation module can define the extent, for example, as a percentage or a number within a specified range. The data release system then compares this extent with the specified compliance threshold. If the extent exceeds the compliance threshold, the release case container, or at least the data set contained in the release case container, can be forwarded to the third party. For example, the relevant data can be transmitted over the internet. The compliance threshold can be individually adjusted for different release cases. This can also be done automatically by a computer system or manually by a user.

[0017] An advantageous further development of the data release system according to the invention is characterized by: - a review interface configured to output the release case container and the compliance assessment to a user and to receive a compliance statement from the user, wherein the compliance statement comprises a compliance assessment confirmed, rejected, or corrected by the user; and - a learning module configured to include the compliance statement in the release case container and, through the use of a reinforcement learning model, to formulate an input request for the large language model depending on the content of the compliance statement and to transmit this to the evaluation module for the next evaluation of a release case container.

[0018] It is to be assumed that the large language model, especially when it first starts working, will make errors when evaluating corresponding release case containers. By providing the verification interface and the learning module, it is possible to initiate a learning process to increase the reliability of the large language model in evaluating the respective release case containers.

[0019] The verification interface can be designed analogously to the receiving interface and / or output interface. The user is, in particular, a person entrusted with the applicable regulations or an expert, such as a lawyer. The user is able to review the assessment of the release case container using the large language model and thus confirm, reject, or correct the compliance assessment accordingly. As a result, the user provides the compliance statement. In the simplest case, this is the compliance assessment, which is supplemented by corresponding analysis information provided by the user. This analysis information can be understood as a marker that describes whether the compliance assessment was confirmed, rejected, or corrected. Corrections can be highlighted.Most preferably, the compliance statement also includes an explanation of why and how the user has confirmed, rejected or corrected the compliance assessment accordingly.

[0020] The learning module processes the compliance statement and, thanks to a suitably trained reinforcement learning model, is able to adapt appropriate input prompts for the large language model depending on the content of the compliance statement. An input prompt is also called a prompt. Such a prompt can be used to specifically influence the operation of a large language model.

[0021] Preferably, the evaluation module is configured to generate the compliance assessment for a release case container, taking into account a prompt formulated by the learning module and / or a compliance statement included in the release case container or a historical release case container. The large language model can thereby increase the reliability of a correct assessment of the release case container currently being classified. This is an iterative process, with each iteration increasing the classification capability of the large language model. The large language model itself is capable of processing the user-generated compliance statement and thus identifying patterns that lead to deviations between the compliance assessment and the compliance statement.This enables the large language model to classify release case containers like the user. From iteration to iteration, the number of evaluated compliance statements increases, so the assessment quality of the large language model improves over time.

[0022] A further advantageous embodiment of the data release system according to the invention further provides that the evaluation module is configured to generate, by means of the large language model during the generation of the regulatory compliance assessment, reference information comprising at least one reference to a historical release case container, based on whose extent of regulatory compliance the large language model generates the regulatory compliance assessment for the release case container. In other words, the evaluation module can thus present a user with an indication of relevant previous data release cases, on whose assessment the classification of the currently considered release case container is based. Correspondingly referenced historical release case containers thus exhibit similarities in the data set, the data set description, and / or the data set context.

[0023] According to a further advantageous embodiment of the data release system according to the invention, it has a regulations database comprising regulations relating to the transfer of data, wherein the regulations database is configured to include new regulations and / or to adapt an already stored regulation with an updated regulation, and wherein the evaluation module is configured to generate the regulation compliance assessment for a release case container taking into account the regulations stored in the regulations database. Thus, it is not only possible to classify the assessment of a respective release case container solely based on the patterns identified by comparing the release case container with historical release case containers, but the large language model can also directly read and consider the respective relevant regulations.The regulations can be formulated in natural language. Thanks to its ability to process and "understand" text formulated in language, the large language model can thus more reliably check whether the respective release case containers actually comply with the currently applicable regulations.

[0024] Accordingly, the compliance assessment can not only include a percentage or a number from a defined range of values ​​to describe the extent to which the respective regulations are complied with, but the evaluation module can also establish a direct link between the current release case container and the regulations. For example, the compliance assessment can include a textual explanation for some or all of the regulations under consideration, explaining why and to what extent this regulation is or is not complied with. The textual explanation can also be enhanced with graphics or diagrams.

[0025] The regulations database can be updated. This allows new regulations to be added and / or old regulations to be updated. Direct consideration of the regulations further increases the reliability that the large language model correctly evaluates the respective release case containers.

[0026] A further advantageous embodiment of the data release system according to the invention further provides that the evaluation module is configured to re-evaluate a historical release case container by generating the compliance assessment, taking into account new and / or amended regulations and / or a changed content of the respective historical release case container. If the regulations and / or the content of a historical release case container change, this may require a re-evaluation of an older data exchange case. According to the invention, the data release system is capable of re-evaluating historical release case containers, i.e., previous data exchange processes. This allows measures to be initiated if corresponding regulations are no longer complied with in a previous data exchange case.Such a data exchange scenario could, for example, involve the regular transmission of newly generated data to a corresponding third party. The next data transmission cycle to the third party can be paused, stopped, or adjusted accordingly.

[0027] The data release system can execute a script that finds relevant data exchange cases depending on changed regulations and / or a changed content of the historical release case container.

[0028] According to a further advantageous embodiment of the data release system according to the invention, the evaluation module is configured to formulate regulations based on evaluation patterns identified by processing a release case container and / or a historical release case container, to compare self-formulated regulations with the regulations stored in the regulations database, and to supplement a self-formulated regulation in the regulations database if the self-formulated regulation deviates beyond a specified degree from all regulations previously stored in the regulations database. As already mentioned, the large language model is capable of evaluating release case containers solely by comparing them with historical release case containers. This is made possible by identifying corresponding evaluation patterns.These evaluation patterns can be formulated into new regulations by the large language model, which can then be added to the regulations database. It is also conceivable to revise existing regulations using the large language model. This allows the operator of the data release system to reliably and transparently uncover relevant regulations of which the operator is not yet sufficiently aware and to incorporate them into the regulations database with minimal effort. Such self-formulated regulations are preferably only included in the regulations database if the respective self-formulated regulations differ fundamentally from already known regulations, and not merely due to minor details such as typos, grammatical deviations, or the like.

[0029] In a method for training a large language model usable by an evaluation module of a data release system described above, the invention provides that the large language model reads the self-generated compliance assessment and the user-generated compliance statement for the same release case container, with the large language model considering the compliance statement as ground truth. This enables continuous training of the large language model. The large language model compares the self-generated compliance assessment with the compliance statement and is thus able to understand which aspects of a check to determine whether or not a respective release case container complies with the applicable regulations were correctly applied and which were not. The large language model can thus optimize its approach for evaluating or classifying release case containers.

[0030] A method for training a reinforcement learning model that can be used by a learning module of a data release system described above comprises the following method steps: - Reading the user-generated compliance statement into the reinforcement learning model; - modifying the prompt formulated for the large language model if the compliance statement involves a change or rejection of the compliance assessment generated by the large language model; and - repeating the two previous procedural steps until the compliance statement includes confirmation of the compliance assessment.

[0031] This allows not only the large language model itself to be further trained, but also the functionality of the large language model to be gradually adapted with the help of the reinforcement learning model. Through training, the reinforcement learning model is specifically enabled to formulate increasingly accurate input prompts for the large language model.

[0032] Further advantageous embodiments of the data release system according to the invention and of the method for training the large language model and the reinforcement learning model also emerge from the exemplary embodiments which are described in more detail below with reference to the figures.

[0033] Showing: Fig. 1 a schematic representation of the process sequence carried out by a data release system according to the invention; and Fig. 2 a schematic representation of the data release system according to the invention.

[0034] The transfer of internal company data to an external party is subject to certain regulations. With the help of a Fig. The data sharing system 1 shown in Figure 2 can automate the sharing of such data with third parties and reliably ensure compliance with regulations. Fig. 1 illustrates the process flow executed by the data release system 1.

[0035] First, a release case container is created, comprising a data set to be forwarded to a third party, as well as a data set description and a data set context. The release case container is sent to the data release system 1 in step 101.

[0036] The release case container is stored in a case database 3 in a step 102. Previous release case containers, referred to in this context as historical release case containers, are stored in the case database 3.

[0037] Historical release case containers already contain an assessment of the extent to which the respective regulations are complied with. This description can be supplemented by an explanation describing why the corresponding assessment was made to this extent.

[0038] In a step 103, the release case container is evaluated using a large language model (LLM) executed by an evaluation module 4. The large language model (LLM) is capable of identifying corresponding evaluation patterns by analyzing historical release case containers and applying them to the current release case container. The result of the large language model (LLM) is summarized in the form of a regulatory compliance assessment.

[0039] In a step 104, the large language model (LLM) can generate reference information that refers to historical release case containers relevant to the current release case container. References are made to those historical release case containers whose evaluation scheme was applied to the current release case container. This allows an operator of the data release system 1 to quickly and easily understand the evaluation method of the large language model (LLM) and to re-examine previous data exchange cases.

[0040] In a step 105, the evaluation result of the large language model LLM can be reviewed by a user 7. User 7 is, in particular, a person familiar with the regulations, such as a lawyer. User 7 analyzes the regulatory compliance assessment and enriches it with information. This information describes whether the regulatory compliance assessment was rejected, confirmed, or corrected, and in particular, to what extent. This generates a regulatory compliance statement 11.

[0041] Compliance statement 11 can be fed into the large language model (LLM) for continuous training. A reinforcement learning model (REM) can also be provided, which serves to adapt the input prompt for the large language model (LLM). REM stands for reinforcement learning model.

[0042] In step 106, the large language model LLM can formulate the self-identified evaluation patterns in the form of new rules and supplement them in a rules database 9. The large language model LLM can also revise existing rules.

[0043] In step 107, a final assessment of the release case container is obtained. In addition to the degree of compliance with the relevant regulations, this information can also include a risk assessment and / or an audit log. In step 108, the release assessment of the release case container is completed, and if the respective release case container satisfies the relevant regulations, it can be automatically forwarded to a third party. If necessary, this must first be manually confirmed.

[0044] Fig. Figure 2 illustrates the structure of the data release system 1 according to the invention. The data release system 1 comprises a receiving interface 2, an evaluation module 4 executing or providing the large language model LLM, a learning module 8 for executing or providing the reinforcement learning model REM as well as the aforementioned case database 3 and regulations database 9. Furthermore, the data release system 1 comprises an output interface 5 for outputting corresponding evaluation results to an operator of the data release system 1 and for forwarding release case containers to third parties. Fig.In the preferred embodiment shown in Figure 2, the receiving interface 2 and the output interface 5 are integrated into a common logical unit. The receiving interface 2 and the output interface 5 can also be referred to as a service and ticketing system. This can be, for example, a web service that allows user interaction with the data release system 1. For example, a person can transmit a request 10 for authorizing a data exchange with a third party via the receiving interface 2. It is also conceivable that the output interface 5 is designed differently, so that corresponding release case containers released by the data release system 1 can be made available to third parties in a way other than using a corresponding service and ticketing system (not shown).

[0045] Release case containers received via the receiving interface 2 are then fed to the evaluation module 4 and evaluated by the large language model LLM. This evaluation, in the form of a regulatory compliance assessment, can then be forwarded to the user 7 and the learning module 8. The user 7 checks and, if necessary, corrects the assessment of the large language model LLM. This assessment is available in the form of a regulatory compliance statement 11. Information is passed to or received from the user 7 via a verification interface 6. The verification interface 6 could also be integrated into the service and ticketing system.

[0046] Compliance statement 11 can be fed to learning module 8 to further train the reinforcement learning model (REM). The reinforcement learning model (REM) then formulates new input prompts for the large language model (LLM).

[0047] The large language model LLM can access the regulations database 9 to retrieve the respective regulations and consider them for generating the regulatory compliance assessment. Even newly generated regulations can be added to the regulations database 9. The data release system 1 can execute a script 12 that detects the addition of new regulations to the regulations database 9 and, by comparing them with historical release case containers maintained in the case database 3, identifies those previous exchange cases that need to be reassessed based on the new regulations.

[0048] Indicated by an arrow 201, users can read the case database 3 via the service and ticketing system and thus manually review all information relevant to the respective exchange cases, even retrospectively. Indicated by an arrow 202, new requests 10 for the aforementioned historical release case containers are transmitted as needed (also automatically by the script 12).

[0049] With the help of the data release system according to the invention, the effort required to evaluate corresponding data exchange cases can be reduced to a minimum. New data exchange cases can thus be quickly reviewed by experts. To avoid errors, it may also be necessary to first manually approve a data exchange case approved by the data release system 1. Data exchange cases can be reliably evaluated because, due to the ability to process large amounts of data in a short time, a large number of previous evaluations of historical release case containers can be taken into account. Furthermore, it is possible to optimize the evaluation scheme of the large language model LLM through continuous training. Furthermore, new regulations can be derived by the data release system 1 almost in real time. Changes to applicable law can be taken into account quickly and reliably. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] US 2023 / 0315981 A1

[0006]

Claims

[1] Data sharing system (1) for the transfer of data to a third party in accordance with the provisions governing the respective transfer, comprising means for processing the data to be transferred using artificial intelligence, characterized by - a receiving interface (2) configured to receive a release case container comprising a data set, a data set description, and a data set context, wherein the data set description comprises a description of the data set and the data set context comprises a reason for forwarding the data set to a third party, each in computer-processable form; - a case database (3) comprising historical release case containers, wherein the receiving interface (2) is configured to store a received release case container in the case database (3); - an evaluation module (4) configured to process the release case container and at least one historical release case container by means of a large language model (LLM), wherein the large language model (LLM) is trained to generate a compliance assessment by comparing the release case container with the at least one historical release case container, comprising a description of the extent to which the release case container complies with the regulations relating to the transfer; and - an output interface (5) configured to forward at least the data set to the third party taking into account the compliance assessment if the release case container complies with the regulations at least up to a specified compliance threshold. [2] Data release system (1) according to claim 1, characterized by - a review interface (6) configured to output the release case container and the compliance assessment to a user (7) and to receive a compliance statement (11) from the user (7), wherein the compliance statement (11) comprises a compliance assessment confirmed, rejected, or corrected by the user (7); and - a learning module (8) configured to include the compliance statement (11) in the release case container and, by using a reinforcement learning model (REM), to formulate an input request for the large language model (LLM) depending on the content of the compliance statement (11) and to transmit this to the evaluation module (4) for the next evaluation of a release case container. [3] Data release system (1) according to claim 2, characterized bythat the evaluation module (4) is configured to generate the compliance assessment for a release case container taking into account an input request formulated by the learning module (8) and / or a compliance statement (11) included in the release case container or a historical release case container. [4] Data release system (1) according to one of claims 1 to 3, characterized by in that the evaluation module (4) is configured to generate, by means of the large language model (LLM), a reference information item in the course of generating the regulatory compliance assessment, comprising at least one reference to a historical release case container, on the basis of whose extent of compliance with the regulations the large language model (LLM) generates the regulatory compliance assessment for the release case container. [5] Data release system (1) according to one of claims 1 to 4, characterized bya regulations database (9) comprising regulations relating to the transfer of data, wherein the regulations database (9) is configured to include new regulations and / or to adapt an already stored regulation by means of an updated regulation, and wherein the evaluation module (4) is configured to generate the regulation compliance assessment for a release case container taking into account the regulations held in the regulations database (9). [6] Data release system (1) according to claim 5, characterized by that the evaluation module (4) is configured to re-evaluate a historical release case container by generating the compliance assessment taking into account new and / or amended regulations and / or a changed content of the respective historical release case container. [7] Data release system (1) according to claim 5 or 6, characterized bythat the evaluation module (4) is set up to formulate regulations depending on evaluation patterns identified by processing a release case container and / or a historical release case container, to compare self-formulated regulations with the regulations held in the regulations database (9) and to supplement a self-formulated regulation in the regulations database (9) if the self-formulated regulation deviates by a specified amount from all regulations previously held in the regulations database (9). [8] Method for training a large language model (LLM) usable by an evaluation module (4) of a data release system (1) according to one of claims 2 to 7, characterized bythat the large language model (LLM) reads the self-generated compliance score and the user-generated compliance statement (11) for the same release case container, whereby the large language model (LLM) considers the compliance statement (11) as ground truth. [9] Method for training a reinforcement learning (REM) model that can be used by a learning module (8) of a data release system (1) according to one of claims 2 to 7, characterized by the following procedural steps: - reading the compliance statement (11) generated by the user (7) into the reinforcement learning model (REM); - modifying the prompt formulated for the large language model (LLM) if the compliance statement (11) includes a change or rejection of the compliance assessment generated by the large language model (LLM); and - repeating the two previous procedural steps until the compliance statement (11) includes confirmation of the compliance assessment.