A method for assigning semantic concepts to unseen data assets.

The method automates semantic concept assignment for unseen data assets, enhancing data integration by using enrichment and neural networks, thus improving efficiency and accuracy in integrating new data sources into dynamic systems.

JP2026182840APending Publication Date: 2026-11-19ROBERT BOSCH GMBH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
JP2026078579
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-08
Filing Date
2026-05-07
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Current data management systems face challenges in harmonizing and integrating data from different sources due to custom schemas, leading to costly and time-consuming manual adapter creation for data integration.

Method used

A computer-implemented method for assigning semantic concepts to unseen data assets using enrichment operations, such as tokenization and acronym expansion, combined with neural networks for natural language processing, to automatically generate meaningful descriptions and rankings of semantic concepts, facilitating dynamic data integration.

Benefits of technology

Enables rapid integration of new data sources into applications like root cause analysis and anomaly detection, reducing manual effort and improving accuracy, efficiency, and robustness in data harmonization.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2026182840000001_ABST
    Figure 2026182840000001_ABST
Patent Text Reader

Abstract

This concerns computer-implemented methods for assigning semantic concepts to new data assets. [Solution] This method assigns semantic concepts to new data assets by enriching data asset metadata (1) (tokenizing labels (A1), expanding acronyms (A2), utilizing knowledge graphs (11) / databases (14) (A3), and generating descriptors (A4)), obtaining concept candidates and similarity scores by comparing enriched metadata embeddings with concept descriptor embeddings, for example, by searching a vector database (13) (2), determining a first ranking based on the similarity scores (3), determining at least a second ranking using a neural network (12b) that processes the first ranking and enriched metadata (4), and storing a proposed mapping (O) based on at least the second ranking in a knowledge graph (11).
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a computer-implemented method for assigning semantic concepts to unseen (previously unseen) data assets, a corresponding data processing system, a computer program, and a machine-readable storage medium.

Background Art

[0002] Prior Art A major challenge in enterprise data management and data-driven applications is data harmonization and integration. Data owners and system developers often use custom schemas for data naming and storage. This keeps data-driven applications (e.g., root cause analysis tools) and reusable software agents away from easy processing and utilization of data. Currently, data engineers build custom adapters or data integration pipelines for each data source, which is costly and wastes time when integrating new sources.

Summary of the Invention

Means for Solving the Problems

[0003] Disclosure of the Invention According to a first aspect, the present invention relates to a computer-implemented method for assigning semantic concepts to new data assets, i.e., unseen data assets. The data asset includes metadata, which includes at least a data asset label. The method comprises the following steps: In one method step, enriched metadata is determined by performing at least one enrichment operation. The latter enrichment operation includes at least one of the following: - tokenizing the data asset label, and - expanding acronyms in the metadata. Expanding acronyms in the metadata may be understood as expanding the data asset label into a known, human and / or machine-readable and understandable name based on known acronyms. The acronyms included in the data asset label may be expanded by referring to acronym-to-representation mappings stored in a knowledge graph. In other words, the acronym expansion may be performed based on prior knowledge, i.e., mappings from short sentences (acronyms) to long sentences, preferably stored in a knowledge graph. Generally, an acronym may be understood as an abbreviation formed from the (initial) components of a word or group of words. The latter enrichment operation involves deriving known assignments between (known) data assets and semantic concepts from the knowledge graph and / or external databases (for example, acronym expansions (expansion from short sentences to long sentences) may be automatically extracted from known mappings (e.g., already contained in the knowledge graph), and such acronym expansions may be used for new mapping requests). In other words, the context of a data asset may be derived, for example, from prior mappings stored in the knowledge graph (for example, the fact that data assets belong to the same container (e.g., one column to a table) can be used to extract data domains, for example, where data assets in the same table can be assumed to belong to the same data domain). Alternatively or additionally, missing information on the data asset may be obtained from vendor documentation or from web searches.The latter enrichment operation involves generating a text description of a data asset, which includes a predetermined limited number of words or characters. Such a text description of a data asset may be determined, for example, by providing a large-scale language model, i.e., a neural network for natural language processing, by providing, along with data asset metadata, derived known assignments between (known) data assets and semantic concepts from a knowledge graph and / or an external database as input. Additionally, the neural network may be prompted to generate a text description of the data asset based on the provided input. The neural network can then determine the text description of the data asset as an output.

[0004] The enrichment operation may include one or more of these distinct suboperations described above. Each of these distinct suboperations may be performed, for example, by a distinct agent, i.e., by a distinct software program. In the next step, multiple semantic concept candidates may be obtained, preferably with corresponding similarity scores. In this context, the similarity score can provide a measure of similarity between a semantic concept candidate and a particular part of the metadata (e.g., a label, a text description, etc.), where, for example, a weighted similarity score may be a real number between 0 and 1, with higher scores indicating greater similarity. Obtaining semantic concept candidates and determining their corresponding similarity scores may be done by performing at least one similarity search. For example, an approximate nearest neighbor search may be performed as an efficient method of similarity search. Within the similarity search, the text embeddings of the enriched metadata may be compared with the (multiple) text embeddings of the semantic concept candidate descriptors. The (multiple) semantic concept candidates may be obtained, for example, from a vector database, where the text embeddings of the semantic concept candidate descriptors may be stored in the vector database. In the next step, a first ranking of the semantic concept candidates may be determined based on at least the similarity score of at least one similarity search. In general, semantic concept candidates with higher scores may be ranked higher in the ranking. Subsequently, a second ranking of at least one semantic concept candidate may be determined by at least one neural network configured for natural language processing. The at least one neural network can receive the first ranking along with enriched metadata as input and determine the second ranking as output. In the next step, at least one semantic concept may be selected based on the second ranking, and the combination of data assets and the selected at least one semantic concept may be stored in the knowledge graph as a proposed mapping.Alternatively, multiple semantic concepts may be selected, and each of the multiple combinations of semantic concepts and data assets, i.e., multiple mappings, may be stored in the knowledge graph. In the repeated iterations of the above and below method steps for assigning semantic concepts to (other) new / unseen data assets, mappings selected by algorithms stored in the knowledge graph (alternatively or additionally, by manual approval by the user), or mappings that generally have high confidence, may be used to improve new / future assignments.

[0005] Advantageously, the proposed method replaces the costly manual creation of data mapping with an automated, dynamic approach. This allows applications such as root cause analysis / monitoring systems to rapidly integrate new data sources, such as new sensors or machines on a manufacturing line. Manual data mapping is eliminated when the application is configured to handle the new data assets. The proposed method automatically provides the necessary semantic concepts, enabling applications to integrate the new data into their calculations for monitoring, root cause analysis, and / or anomaly detection.

[0006] Additionally, the proposed invention can be used in any dynamic data-driven system or data management context that requires data assets having different data schemas to be arranged / harmonized or integrated according to a standard data model or a semantic data model. Conventionally, this requires data engineers / data stewards to manually develop custom adapters or custom mappings, which can be a very time-consuming and costly effort in complex systems. The method proposed herein provides smooth, real-time, i.e., fully automated data-to-semantic mapping support in the presence of dynamic data-driven systems and applications.

[0007] Preferably, the asset metadata further includes information (additional to the data asset label) that includes at least one of the following: asset type, text description of the asset, data type of the asset, data source, data application, and data domain. The asset type may be, for example, a column, table, or attribute. The data source, i.e., where the data asset may be stored, may be, for example, an RDBMS or a Data Lake, and an example of the data application (the IT application from which the data asset was originally created) may be SAP. The above information regarding the data application may be used, for example, to find a description of the data asset in embedded vendor documentation or via a web search (online database search). Data assets and semantic concepts may also often be associated with data domains (e.g., manufacturing, engineering, sales, etc.). This information regarding data domains may be used to narrow the search space for semantic concepts, because they may optionally be associated with data domains.

[0008] Advantageously, by providing more metadata information, the mapping process is transformed from a label (i.e., keyword)-based approach to a context-aware approach, resulting in significant improvements in accuracy, efficiency, and robustness.

[0009] Preferably, depending on the additional information included / provided in the metadata as described above, separate similarity searches may be performed for data asset label similarity, text description similarity, and contextual correspondence of semantic concepts, each with corresponding similarity score determinations. In this context, the contextual correspondence of semantic concepts may be determined, for example, together with the data domain from which the semantic concepts are derived. For example, the contextual correspondence can determine how well the identified semantic concepts match the context derived from the additional information in the metadata.

[0010] Advantageously, by performing separate similarity searches according to the available metadata information, i.e., the metadata aspects, one can assign different weights to each aspect. For example, a text description may provide more information than a label, or vice versa, and the corresponding similarity score may be given a higher weight.

[0011] Preferably, the text description of the data asset may be determined based on at least one of the data asset documentation and / or the documentation of the application providing the data asset, along with documentation provided in the form of unstructured data within a vector database. Alternatively or additionally, the text description of the data asset may be determined based on retrieving information within an online database.

[0012] Preferably, in addition to the similarity score, at least one of the relevance score and the maturity score is determined. The relevance score may be determined by using semantic concept metadata to compare the semantic concept metadata of a semantic concept candidate with data asset metadata (e.g., using text embeddings and an appropriate similarity measure), and assigning a higher relevance score to the semantic concept candidate if the data domain associated with the data asset matches the data domain associated with the semantic concept candidate, and / or if the semantic concept type associated with the data asset matches the semantic concept type associated with the semantic concept candidate. In other words, the relevance score may be determined by matching where the potential semantic concept matches the data domain of the asset, and / or where the semantic concept type (e.g., class, property) matches the data asset type (e.g., column or attribute), where a semantic concept candidate with higher relevance receives a higher relevance score.

[0013] Similarly, the maturity score may be determined based on semantic concept metadata relating to usage, where (a) the registration status of a semantic concept candidate, where a more advanced registration status determines a higher maturity score, or (b) the usage scope of a semantic concept candidate, where a wider usage scope results in / determines a higher maturity score, where the usage scope is determined based on at least one of (i) the organizational level to which the potential semantic concept is proposed, where a higher organizational level results in a wider usage scope, or (ii) the standardization status of a semantic concept candidate, where a standardized semantic concept has a wider usage scope than an unstandardized semantic concept.

[0014] In other words, semantic concept candidates can acquire a maturity score based on their registration status (e.g., candidate, eligible, proposed standard, etc.) and the scope of their ontology / glossary / taxonomic methodology (low = proposed by individual users / departments, medium = proposed by departments / domains, high = proposed by company-wide standardization teams). Semantic concept candidates with a higher maturity score will receive a higher maturity score. Generally, the maturity score can indicate how formalized and widely adopted the concept is.

[0015] Preferably, semantic concept candidates may be ranked in their respective rankings by determining a weighted combination of a similarity score (potentially determined separately, depending on the aspects covered in the metadata) and, optionally, an optionally determined relevance score and / or maturity score. The weights in the weighted combinations may be predetermined.

[0016] Advantageously, weighted combinations can allow for balancing different contributions, such as data asset label information and data asset descriptions. On the other hand, this can also allow for particular emphasis on prioritization contributions that may be known, such as past assignments of semantic concepts to data assets.

[0017] Preferably, before the proposed mapping is selected, the additional neural network for natural language processing may be provided with at least one second ranking and enrichment metadata. Furthermore, the additional neural network may be provided with text instructions as additional input, which include instructions for the additional neural network to evaluate and potentially reorder the ranking of semantic concepts provided by the second ranking. The additional neural network can then determine a third ranking based on the input. The proposed mapping (to be stored in knowledge) may then be selected based on at least the third ranking.

[0018] Advantageously, additional neural networks can perform more complex inferences using enriched metadata and initial rankings as context. They can take into account relationships between concepts, domain-specific knowledge, and even the nuances of the words used in descriptions. By providing the neural network with instructions for evaluating the initial rankings, it can perform a form of "self-reflection" or validation. This allows it to identify and correct potential errors or inconsistencies in the initial rankings. For example, it might be possible to detect that two highly similar concepts are ranked far apart due to the biases of the similarity scale and adjust their rankings accordingly.

[0019] Preferably, semantic concepts may be added to a first, second, or third ranking if their respective overall scores exceed a predetermined threshold. Each overall score may be determined based on a similarity score and, optionally, a relevance score and / or maturity score. Additionally, there may be a predetermined maximum number of semantic concepts that should be included in each of the first, second, or third rankings. Advantageously, this acts as a filter, progressively refining the selection of semantic concept candidates by gradually removing less relevant concepts in subsequent rankings. This multi-ranking approach allows for a more thorough examination of semantic concept candidates while ensuring greater relevance and less noise-induced scattering in each subsequent (second or third) ranking.

[0020] Preferably, the method may further include a step of determining at least one overall ranking of semantic concept candidates. This overall ranking may be determined by a weighted combination of all previously determined scores, including similarity scores, relevance scores, and maturity scores, and if each ranking is determined, it may be determined based on previously determined rankings, i.e., the first, second, and third rankings. The proposed mapping (to be stored in knowledge) may then be selected based on the overall ranking.

[0021] Preferably, the method may further include the following steps, in which a proposed / one map from the proposed map (e.g., the highest-ranked semantic concept candidate according to a first, second, third, or overall ranking along with the data assets) is provided to an IT application. The IT application may be, for example, a root cause analysis or anomaly detection software program on a production line during manufacturing. The IT application may be configured to process the data assets according to the associated semantic concepts (meaning). In this context, the data assets may include data from a new sensor or new data source added to the IT system (e.g., on a production line), where the IT application controls the IT system by taking action (e.g., issuing an alarm) or making a decision (e.g., discarding a manufactured part).

[0022] Advantageously, these additional steps can enable real-time adjustments and automation in manufacturing, particularly by rapidly integrating new data sources (sensors, machinery) into production line applications for root cause analysis and anomaly detection. The data is often processed on the fly, thereby allowing applications to gain immediate insights. This enables automated adjustments that can improve flexibility, leading to faster anomaly detection, root cause identification, and minimized downtime.

[0023] In a further aspect, the present invention relates to a data processing system comprising a processor configured to carry out a method as described herein.

[0024] According to a further aspect, the present invention relates to a computer program comprising machine-readable instructions that, when executed by a computer, cause the computer to perform one of the computer-implemented methods described above and hereinafter. Furthermore, according to another aspect, the present invention relates to a machine-readable storage medium storing the above computer program.

[0025] Embodiments of the present invention will be described in more detail with reference to the following drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] [Figure 1] This is a flowchart showing an exemplary embodiment. [Figure 2] This is a flowchart showing an exemplary embodiment. DETAILED DESCRIPTION OF EMBODIMENTS OF THE INVENTION

[0027] Description of Embodiments Figure 1 illustrates an exemplary embodiment of a method 1000 for assigning semantic concepts to unseen / unknown / new data assets. This flowchart shows information exchanged between different agents, that is, respective parts of a software program. The initial input I includes metadata of the new data asset. In general, the metadata of a data asset may be understood as a description of the data asset (for example, sensor data). Such metadata may consist of the following elements: - Label (for example, MATNR), and further optionally - Type (for example, column, table, attribute, object...) - Description (for example, "material number of ~") - Data type + sample (for example, xsd:string) - Context (for example, data source=SAP_1, container=MAST, data application, data field, data domain of the asset, etc.) It may consist of these.

[0028] This method determines, as output O, a proposed mapping that links data assets (e.g., by label) to semantic concepts (e.g., uri). As a non-restrictive example, the data asset label "MATNR" can be mapped to "http: / / company.com / man-core-ont / MaterialNumber". Note that, in general, a semantic concept may simply be a vocabulary (e.g., glossary), a taxonomic concept, or a data model entity.

[0029] In Figure 1, in the first step or phase 1, the input data I is enriched by at least one enrichment operation, depending on the metadata available / provided in I. Agent A1 can perform label tokenization by converting data asset labels into individual tokens. For example, the data asset label "PrdId" can be converted into tokens "prd" and "id". Similarly, "MATNR" can be tokenized into "mat" and "nr". Agent A2 can perform token decomposition by converting the tokens (obtained as the output of Agent A1) into more meaningful / human or machine-understandable labels, for example, based on known acronyms stored in the knowledge graph 11. For this purpose, Agent A2 can obtain information from the knowledge graph 11, which stores prior knowledge in the form of already known mappings (previously proposed and optionally approved, for example, by a human user), as well as domain knowledge in the form of known acronyms and / or acronym decompositions 111 and data asset descriptions 112. For example, A2 can receive A1's tokenized output "prd id" as input and convert it to a "product identifier". Optionally, agent A3 can derive / query missing contextual information within the knowledge graph 11 from known association maps already stored within the knowledge graph 11. For example, a data asset map for a container asset can reveal the data domain or data source of the data asset.

[0030] Furthermore, optionally, agent A4 may generate a description if a meaningful description of the data asset is not provided in the input data I. For this purpose, A4 may: Known descriptions of a given data asset from knowledge graph 11, Data asset labels, as well as information about data assets from other external databases such as online sources / online-accessible databases 14, based on type and optional context (if available), Information about data assets in the available vendor documentation 113 is obtained by using a similarity search based on data asset labels and a vector database 13 containing vendor documentation 113 for each data application. It is possible to attempt to obtain / acquire this. In general, the vector database 13 described above can store text embeddings of various types (e.g., semantic concept labels and descriptions) in different indices, and furthermore, this enables the agent to discover similar concepts based on data asset labels or descriptions.

[0031] A semantic concept repository 10, including a glossary, taxonomic methods, and ontology, may be provided, and the embedding of semantic concept names and corresponding descriptions from the semantic concept repository 10 may be stored, for example, in a vector database 13.

[0032] Ultimately, agent A4 can facilitate neural network 12a (i.e., LLM) for natural language processing in RAG format by providing the neural network with pre-obtained information. Therefore, the task for neural network 12a would be to generate meaningful and concise descriptions of data assets, given the pre-obtained information. If the neural network has high confidence in the descriptions, the generated descriptions may be stored in the knowledge graph for future use. It should be noted that agent A4's contribution may be particularly important because meaningful descriptions of data assets (if not provided as part of input I) can be crucial in increasing the likelihood of discovering semantic conceptual goals.

[0033] Enriched metadata is obtained as an output of Phase 1.

[0034] In the second step or phase 2, multiple semantic concept candidates may be retrieved / determined from the vector database along with their corresponding similarity scores by performing at least one similarity search that compares the text embeddings of the enriched metadata with the text embeddings of semantic concept candidate descriptors stored in the vector database.

[0035] To this end, agent A5 can determine similar semantic concepts based on the labels and context of data assets by embedding labels in the data assets and performing a similarity search against an index of a vector database 13 for semantic concept names (i.e., labels). This allows semantically more similar labels to receive higher similarity scores. To limit / restrict / reduce the number of concepts, the search can include the top X concepts (e.g., 100-500), but a minimum similarity threshold can also be used (i.e., semantic concepts with similarity scores below a predetermined threshold may be discarded). Additionally, to reduce the search space within the index, reliable contextual information about the data assets (e.g., the semantic concept type if known) may be used as a filter in the similarity search. This increases the likelihood of filtering out false positive candidates.

[0036] Furthermore, Agent A6 can retrieve similar semantic concepts based on the descriptions and contexts derived in Phase 1. By also taking into account the provided context information, Agent A6 can perform a similarity search based on the data asset descriptions obtained by Agent A4. Similar to Agent A5, Agent A6 can now only incorporate the top X concepts (e.g., 100-500) identified in the search, and additionally, a minimum similarity threshold can be used.

[0037] Agent A7 may optionally score semantic concepts based on relevance and maturity. Based on semantic concept metadata, for example, semantic concepts with matching data domains or expected concept types can receive a higher relevance score. Similarly, based on semantic concept metadata, for example, semantic concepts with higher maturity (based on registration status and collection scope, e.g., application, domain, core) can receive a higher maturity score.

[0038] Therefore, the output of Phase 2 is a set of semantic concepts having corresponding similarity scores and (optionally) relevance and / or maturity scores. Note that in Phase 2, only one or two of agents A5, A6, and A7 may be used; that is, in different embodiments, not all agents may be used. Therefore, in different embodiments, the semantic concepts and their corresponding (similarity) scores may be determined solely based on, for example, a label similarity search by agent A5.

[0039] In Phase 3, a first ranking of semantic concept candidates (obtained as the output of Phase 2) is obtained as an output based on at least one similarity score from the similarity search in Phase 2. For this purpose, agent A8 can combine the scores determined in Phase 2 based on a pre-configured scoring metric or per-agent weights. This allows agent A6 to calculate intermediate scores for the final selection of semantic concepts and determine the first ranking from these intermediate scores. Agent A6 can select / further reduce the number of semantic concept candidates to, for example, 50-100 concepts by, in particular, incorporating only the top 50-100 concepts and / or concepts with intermediate scores above a predetermined threshold. For example, given scores s1, s2, and s3 for a given semantic concept candidate from Phase 2, the intermediate score s for this semantic concept int is, s int =AVG(w1 * s1+w2 * s2+w3 * It is often determined as s3), where AVG represents the average calculation and wi represents predetermined weights that weight the different contributions of the agent's score in phase 2.

[0040] In Phase 4, a second ranking of at least one semantic concept candidate is determined by at least one neural network 12b configured for natural language processing. The at least one neural network 12b receives the first ranking, i.e., the output from Phase 3, along with enriched metadata, as input and determines the second ranking as output.

[0041] In this Phase 4, agent A9 can prompt one or more (e.g., three) neural networks 12b to perform language processing (e.g., by additionally using different facilitation strategies) and request the neural networks 12b to generate a second ranking of semantic concept candidates selected based on the first ranking. To this end, agent A9 can provide all information regarding the input data assets determined in Phases 1-4 and the selected semantic concept candidates in a RAG ("retrieval augmented generation") manner to at least one or more neural networks 12b in the prompts. This may be done in the same way as for agent A8.

[0042] Optionally, several short-term learning techniques based on realistic examples of tasks in the prompt may be used for each of the neural networks 12b. For each facilitated neural network, agent A9 can convert the ranking of each semantic concept into a score for each semantic concept (e.g., a score between 0 and 100).

[0043] Optionally, in Phase 4, Agent A10 may be used to generate a re-ranking based on the LLM using a self-reflection prompt. For this purpose, Agent A10 can take Agent A9's second ranking as input and provide this ranking as input to one or more (different) additional neural networks 12c (i.e., in this case, the LLM) along with an input text prompt that reflects this in the ranking result (given the same input as Agent A9 and its result). For example, this may also result in a semantic conceptual score between 0 and 100 obtained as the output from each of the additional neural networks 12c. Thus, the third ranking may be obtained from A10 based on the LLM re-ranking (as a result of the self-reflection prompt provided to the LLM 12c).

[0044] Agent A11 can receive the output of at least one of Agents A9 and A10 and calculate an overall score for each of the selected semantic concepts. For this purpose, Agent A11 can use scores from all available scoring metrics. The scoring function may be the same as for Agent A8, whereas Agent A11 can use metrics from Agent A9 (one per neural network) and Agent A10 (again, one per neural network). It is worth noting that, compared to Agent A8, a different set of pre-configured weights may be used. After an overall score has been calculated for each semantic concept candidate, the top 5-10 scoring concepts may be selected to further reduce the number of semantic concept candidates, and the overall ranking may be determined based on the selected semantic concept candidates along with those scoring determined by Agent A11. The overall score obtained from Agent A11 may then be used to determine the overall ranking of the semantic concept candidates.

[0045] In a further step, agent A12 may select at least one semantic concept as a proposed mapping based on the overall ranking received as output from A11. The at least one proposed mapping may then be stored in the knowledge graph for future reference. On the one hand, this can provide a means to provide immediate answers to future requests without the need to recalculate all scores. On the other hand, the proposed mappings and their data asset information may be used to improve new / similar mapping requests.

[0046] Optionally, the final output O may be a list of proposed mappings, and optionally include a total score determined by agent A11. This additionally allows the user to determine how reliable the proposed mappings are. Mappings above a given threshold are likely to be very good, while those below that threshold should always be used with caution.

[0047] Figure 2 shows a flowchart of another exemplary embodiment of a method for assigning semantic concepts to unseen / unknown / new data assets. In step 100, enriched metadata is determined from the metadata of the data asset by performing at least one enrichment operation. Such enrichment operation includes, as a substep or suboperation, at least one of the following: tokenizing the data asset label, expanding acronyms in the metadata, deriving known assignments between (known) data assets and semantic concepts from a knowledge graph and / or an external database, and generating a text description of the data asset. These substeps or suboperations may be performed depending on the availability of the respective information in the metadata. At least one data asset label should be available. Optionally, information regarding at least one of the data asset type, data asset description, data type, and / or further context of the form of text information may be included in the metadata of the data asset.

[0048] In step 200, multiple semantic concept candidates and their corresponding similarity scores are retrieved from the vector database by performing at least one similarity search. This similarity search compares the text embeddings of the enriched metadata with the text embeddings of the semantic concept candidate descriptors stored in the vector database. If the enriched metadata contains more than the expanded data asset labels, separate similarity searches may be performed based on the semantic concept descriptors and at least the (expanded) data asset labels, and based on the determined descriptions, if each of these pieces of information is available.

[0049] In step 300, a first ranking of semantic concept candidates is determined based on at least one similarity score from a similarity search. If multiple similarity scores, as well as optional relevance and / or maturity scores, are determined, the first ranking may be obtained by weighting each score, which has a predetermined weight.

[0050] Next, in step 400, a second ranking of at least one semantic concept candidate is determined by at least one neural network configured for natural language processing. This neural network receives the first ranking along with enriched metadata as input and determines the second ranking.

[0051] In step 500, at least one semantic concept is selected and stored in the knowledge graph as a proposed mapping along with data asset labels. This selection is based on at least a second ranking. Optionally, before a semantic concept is selected and combined with data asset labels to form the proposed mapping, an additional neural network for natural language processing may be provided with at least one second ranking and enrichment metadata as input, and additionally, a text command may be provided as input, containing instructions to the additional neural network to evaluate and potentially reorder the rankings of semantic concepts provided by the second ranking. The additional neural network can determine a third ranking based on the input. The proposed mapping may be selected from this third ranking.

[0052] Alternatively, in step 500, before at least one semantic concept is selected, the overall ranking of the semantic concept candidates may be determined by a weighted combination of all previously determined scores, including similarity scores, relevance scores, and maturity scores, and based on the first and second rankings determined. In this case, at least one proposed mapping may be selected based on the overall ranking.

[0053] In an optional step 600, the proposed mapping (from step 500) may be provided to an IT application along with a (new) data asset. This application may be configured to process the data asset according to associated semantic concepts. The (new) data asset may be data from a new sensor or a new data source added to the IT system, and the application is configured to control the IT system by performing actions or making decisions. In a further optional step 700, the IT system is controlled by the IT application, taking into account the information from the new data asset.

Claims

1. A computer-implemented method for assigning semantic concepts to an unseen data asset, wherein the data asset includes metadata, at least including a data asset label, in the method, The aforementioned method, - Determining enriched metadata by performing at least one enrichment operation, the enrichment operation is: Tokenizing the aforementioned data asset labels, Expanding the acronyms within the aforementioned metadata, Deriving known assignments between (known) data assets and semantic concepts from knowledge graphs and / or external databases, To generate a text description of the aforementioned data asset, It includes at least one of the following, - By performing at least one similarity search that compares the text embedding of the enriched metadata with the text embedding of the semantic concept candidate descriptor, multiple semantic concept candidates are obtained along with their corresponding similarity scores. - Determining a first ranking of the semantic concept candidates based on at least the similarity score of the at least one similarity search, - Determining at least one second ranking of the semantic concept candidates by at least one neural network configured for natural language processing, wherein the at least one neural network receives the first ranking along with the enriched metadata as input and determines the second ranking as output, - At least one semantic concept is selected as a proposed mapping based on the at least one second ranking, and at least one of the proposed mappings is stored in the knowledge graph. Methods that include...

2. The metadata of the asset further includes additional information, including at least one of the following: asset type, text description of the asset, data type of the asset, data source, data application, and data domain. The method according to claim 1.

3. Depending on the additional information contained in the metadata, separate similarity searches are performed for data asset label similarity, text description similarity, and / or contextual correspondences of the semantic concepts, each involving the determination of a corresponding similarity score. The method according to claim 2.

4. The text description of the aforementioned data asset is, - Documentation of the data assets, and / or documentation of the application providing the data assets, accompanied by documentation provided in the form of unstructured data within a vector database. and / or, - Searching for information in online databases Determined based on at least one of the following: The method according to any one of claims 1 to 3.

5. In addition to the similarity score, at least one of the relevance score and the maturity score is determined. - The relevance score is determined by using semantic concept metadata to compare the semantic concept metadata of a semantic concept candidate with the data asset metadata, and assigning a higher relevance score to the semantic concept candidate if the data domain associated with the data asset matches the data domain associated with the semantic concept candidate, and / or if the semantic concept type associated with the data asset matches the semantic concept type associated with the semantic concept candidate. - The maturity score is, (a) The registration status of the semantic concept candidate, in which a more advanced registration status determines a higher maturity score, Or, (b) A range of use of the semantic concept candidate, wherein a broader range of use results in / determines a higher maturity score, the range of use being: (i) A potential semantic concept is proposed at an organizational level where a higher organizational level results in a wider range of use. Alternatively, (ii) The standardization status of the candidate semantic concept such that the standardized semantic concept has a broader range of use than the unstandardized semantic concept. The scope of use is determined based on at least one of the following. Determined based on semantic conceptual metadata relating to The method according to any one of claims 1 to 4.

6. The aforementioned semantic concept candidates are ranked by determining a weighted combination of a determined similarity score and, optionally, an optionally determined relevance score and / or maturity score. The method according to any one of claims 1 to 5.

7. Before the proposed mapping is selected, an additional neural network for natural language processing is provided with at least one second ranking and the enriched metadata, and additionally, a text command is provided as input, which includes instructions to the neural network for evaluating and potentially reordering the ranking of semantic concepts provided by the second ranking. The aforementioned additional neural network determines a third ranking based on the input, The proposed mapping is selected from the third ranking. The method according to any one of claims 1 to 6.

8. Semantic concepts are added to the first, second, or third ranking if their respective overall scores exceed their respective predetermined thresholds. The overall score is determined based on the similarity score and, optionally, the relevance score and / or the maturity score. The method according to any one of claims 1 to 7.

9. The above method, before selecting at least one semantic concept, further involves the following steps, namely: - The steps include determining a weighted combination of all previously determined scores, including the similarity score, the relevance score, and the maturity score, and determining an overall ranking of the semantic concept candidates based on the first ranking, the second ranking, and / or the third ranking of the determined rankings, The at least one proposed mapping is selected based on the overall ranking. The method according to any one of claims 1 to 8.

10. The above method further includes the following steps, namely, - A step (600) of providing an IT application with a proposed mapping along with the aforementioned data assets, The aforementioned IT application is configured to process data assets according to the associated semantic concepts, The aforementioned data assets are data from new sensors or new data sources added to the IT system. The aforementioned application controls the IT system by performing actions or making decisions. Step (600), - The steps (700) of controlling the IT system using the IT application, including, The method according to any one of claims 1 to 9.

11. A data processing system comprising at least one processor configured to carry out the method described in any one of claims 1 to 8.

12. A computer program, which, when the program is executed by a computer, includes instructions to cause the computer to perform the method described in any one of claims 1 to 8.

13. A computer-readable data carrier, wherein the computer program described in claim 10 is stored on the computer-readable data carrier.