Apparatus and method for formalizing uncontrolled data format input

Abstract

A computer-implemented method for formalizing an uncontrolled data format input comprises the steps of: obtaining (S10) an uncontrolled data format input through an input interface; determining (S20), by an input processing algorithm, an embedding (E) of the acquired uncontrolled data format input in an embedding space (30) resulting from the plurality of pre-processing embedding, where the embedding (E) is represented by an embedding vector, where the embedding vector comprises a plurality of vector entries specifying the embedding (E); formalizing (S30) the controlled data format output using embedding (E) by an output processing algorithm; and outputting (S40) the controlled data format output through the output interface.

Description

Technical Field

[0001] The present disclosure relates to computer-implemented methods and apparatus for formalizing uncontrolled data format input. Background Art

[0002] Today, as an example of engineering and operations related content, uncontrolled engineering intent is manually translated or in other words converted into a controlled representation, such as into an ontological engineering intent model representation, for further processing based thereon, such as in the context of a process automation tool workflow.

[0003] Nevertheless, only manual efforts of informed process and automation engineering domain experts and ontology experts are needed to produce representations of ontology engineering intent or otherwise formalize uncontrolled data format inputs into controlled data format outputs that can be further processed. Nevertheless, even for experts, this is prone to errors.

[0004] Therefore, an improved method is needed to formalize such uncontrolled data format input, such as engineering intent. Summary of the invention

[0005] The invention is defined by the subject matter of the independent claims. Additional features of the invention are given in the dependent claims.

[0006] According to one aspect of the invention, a computer-implemented method for formalizing an uncontrolled data format input comprises the following steps: obtaining an uncontrolled data format input through an input interface; determining, through an input processing algorithm, an embedding of the obtained uncontrolled data format input in an embedding space generated by a plurality of pre-processed embeddings, wherein the embedding is represented by an embedding vector, wherein the embedding vector includes a plurality of vector entries specifying the embedding; formalizing, through an output processing algorithm, a controlled data format output using the embedding (determined above); and outputting the controlled data format output through an output interface.

[0007] The term "formalization" as used herein describes the transformation or conversion of an uncontrolled data format input (particularly in the form of an embedding vector) into a controlled data format output. For example, formalization includes converting an embedding vector of engineering intent into an ontology representation that can be further processed in an engineering workflow.

[0008] The term "controlled data format" as used herein includes, but is not limited to: RDF, OWL, XML, JSON, AutomationML, modular type packages, MTP, different relational database formats.

[0009] In other words, the term "embedding" as used herein is an abstract representation of an input, in particular its latent semantic representation. More preferably, an embedding is a vector representation of the input, referred to as an embedding vector. For example, an embedding is a meaningful vector representation of a word, sentence, and / or formula with respect to engineering intent. The embedding is preferably acquired or learned through natural language processing techniques or from a language model.

[0010] Preferably, the controlled data format output comprises a data format, a schema or an ontology. The controlled data format output is a known target format that can be further processed automatically, in particular by other tools in the engineering workflow. In other words, the method allows formalizing the controlled data format output in a way that represents the content of the uncontrolled data format input. Furthermore, the uncontrolled data format input may also include information represented in the controlled data format, but which is provided as an uncontrolled input, such as the format and well-structured content of a JSON file, but provided in plain text.

[0011] Preferably, the input processing algorithm comprises a neural network, referred to as the input layer.

[0012] Preferably, the output processing algorithm comprises a neural network, referred to as the output layer.

[0013] Preferably, the embedding space is (a) pre-trained state-of-the-art open source embeddings obtained through a natural language processing algorithm (e.g., a large language model (LLM)), or (b) proprietary trained embeddings obtained using domain-specific input and standard NLP algorithms, or (c) (a) or (b) plus a set of pre-processed domain-specific concepts / terms / inputs, represented by abstract / latent semantic representations that have been determined by the input processing algorithm. In particular, these pre-processed embeddings in the embedding space can represent a controlled / structured (e.g., ontology) representation of the engineering or operational content. In addition, for example, these pre-processed embeddings in the embedding space can be stored in an embedding database.

[0014] Preferably, specifying an embedding (i.e. assigning it an n-dimensional vector in the embedding space (i.e., an ordering of locations)) can also be referred to as classifying the input (as a "location" / as a "region" / "bubble" around a "location").

[0015] Preferably, the vector is represented as a normalized vector.

[0016] Preferably, a transformer model is provided, which transformer model also preferably includes an encoder model and a decoder model. The determination of the embedding of the acquired uncontrolled data format input by the input processing algorithm corresponds to the encoder model. The abstract / latent semantic embedding representation is then used as input to the decoder model to determine the controlled data format output.

[0017] Therefore, the method provides a pipeline for analyzing, classifying / computing and formalizing uncontrolled data format inputs (such as engineering or operational content) into controlled data format outputs (such as ontological representations of engineering and operational content) via abstract / latent semantic representations (embeddings), which controlled data format outputs can be further used in engineering workflows in particular.

[0018] Therefore, an improved automated way to formalize uncontrolled data format input is provided.

[0019] In a preferred embodiment, the uncontrolled data format input includes engineering or operations related content.

[0020] Preferably, the engineering or operation related content includes any one of engineering intent, process engineering specification, automation engineering specification, engineering recipe and operation specification.

[0021] In other words, engineering or operations-related content includes any engineering knowledge within the entire "life cycle" of the engineering workflow. For example, this starts from process engineering (e.g., chemical engineering), through automation engineering (e.g., automation and control engineering), to operations (i.e., when the plant is built and operating). In other words, engineering or operations-related content includes content related to both the engineering phase and / or the operations phase of an industrial process plant.

[0022] For example, operational specifications include starting and / or stopping a service, or changing the recipe of a product.

[0023] Therefore, engineering or operation related content can be formalized into a controlled data format output, which can be further used in the engineering workflow. In addition, the engineering or operation content is provided in a format that any specific stakeholder in the engineering process can process and understand. Therefore, the information gap between stakeholders (e.g., automation engineers and process engineers) can be eliminated.

[0024] Thus, an improved automated way to formalize uncontrolled data format input in the form of engineering or operational content is provided.

[0025] In a preferred embodiment, determining, by the input processing algorithm, the embedding of the acquired uncontrolled data format input comprises: identifying engineering or operational content in the uncontrolled data format input, and determining the embedding based on the identified engineering or operational content.

[0026] Preferably, the engineering or operation content is extracted from the uncontrolled data format input by a machine learning model. For example, if the engineering or operation content is contained in a natural language input or an uncontrolled text input, the engineering or operation content is identified from the uncontrolled data format input using natural language processing techniques. The engineering or operation content is then used to determine the value of a corresponding embedding vector representing the engineering or operation content.

[0027] Thus, an improved method is provided for formalizing uncontrolled data format input in the form of engineering or operational content.

[0028] In a preferred embodiment, the uncontrolled data format input includes natural language input or uncontrolled text input.

[0029] In other words, the uncontrolled data format input includes uncontrolled input in the form of text and / or voice.

[0030] For example, the method transforms engineering or operational content, preferably natural language engineering intent, into a formalized ontology engineering intent representation.

[0031] Additionally, uncontrolled data format input may also include information that is represented in a controlled data format but is provided as uncontrolled input, such as the format and well-structured content of a JSON file but provided in plain text.

[0032] For example, the text also contains tables with numbers or other symbols in addition to letters.

[0033] Thus, uncontrolled data format input in the form of text may include text in poetry or controlled narratives, but also unstructured text (i.e., any text that is not an image / audio / video) with numbers, variables, tables, etc.

[0034] Thus, an improved and automated way to formalize uncontrolled data format input in the form of engineering or operational content is provided.

[0035] In a preferred embodiment, determining, by the input processing algorithm, an embedding of the acquired uncontrolled data format input comprises identifying, in an embedding space, a pre-processed embedding that is similar to the embedding in terms of a predetermined metric, and determining the embedding based on the similar pre-processed embedding.

[0036] Preferably, the metric comprises "cosine similarity" or other common NLP-related metrics that capture semantic similarity / closeness of natural language building blocks / structures / terms.

[0037] In other words, determining the embedding includes analyzing the acquired uncontrolled data format input to find the most similar embedding among the pre-processed embeddings.

[0038] For example, an embedding is determined, and then analyzed according to its position in the overall embedding space (e.g., particularly with respect to which other pre-processed embedding vectors are closest).

[0039] Preferably, the predetermined metric is pre-selected from a pool of predetermined metrics.

[0040] Preferably, the pre-processing embeds a controlled / structured (eg, ontology) representation representing the engineering or operational content.

[0041] In a preferred embodiment, the predetermined metric comprises a semantically meaningful metric.

[0042] Preferably, the metric includes semantics, where similar pre-processed embeddings are semantically similar to the embedding, or in other words, close to the embedding. Thus, the most suitable, most similar embedding can be found among the pre-processed embeddings in the embedding space.

[0043] For example, in terms of this semantically meaningful metric, two embeddings that are close to each other indicate that the underlying words / sentences / contents of these two embeddings have similar meanings.

[0044] Preferably, the terms "close" or "near" relate to the metric, whereas the term "similar" relates to the semantics of the uncontrolled format data input.

[0045] In a preferred embodiment, the dimension of the embedding vector is greater than 10, preferably greater than 50, and further preferably greater than 100. The larger the dimension of the embedding vector or the embedding space, the more specific the abstract semantic representation of the concept to be represented.

[0046] Each embedding dimension (and accordingly, each embedding vector component value) can be viewed as a feature / characteristic (or, accordingly, a scale on this feature dimension), e.g., “the relationship of the concept to the oil and gas industry versus the pharmaceutical and chemical industries,” “the relationship of the module function to separating and mixing some liquid materials,” etc. Thus, a value of 0.9 (on a normalized scale of -1 to +1) for the first component indicates that the current concept is strongly related to oil and gas (as opposed to pharmaceuticals / chemicals), and a value of -0.2 for the second component indicates that the concept is related to mixing rather than separation. Additionally, this value may also indicate that there is no clear distinction here, which indicates that the relationship of the concept to separation / mixing may not be that important to the semantic meaning of the concept.

[0047] In practice, these are more abstract features that an embedding model / NLP algorithm is pre-trained or learned, optionally on large amounts of text (on the internet or Wikipedia) or domain-specific document content.

[0048] In a preferred embodiment, formalizing the controlled data format output using the embedding vector includes mapping the embedding vector to the controlled data format output.

[0049] In other words, similar embedding vectors in the embedding space may be grouped together and mapped to a single controlled data format output. Thus, mapping the embedding vectors to the controlled data format output preferably includes identifying such a set of similar embedding vectors, and mapping the embedding vectors to the controlled data format output associated with the set of similar embedded vectors. In particular, an embedding vector in such a set of embedding vectors may be identified and obtained from pre-processed embeddings in the embedding space, whereby these pre-processed embeddings in the embedding space may already represent a controlled / structured (e.g., ontology) representation of the engineering or operational content.

[0050] Preferably, the mapping comprises an explicit mapping, obtained in particular by a domain expert from a predefined domain knowledge representation (eg in an ontology format).

[0051] Preferably, the mapping comprises a learned mapping learned in a supervised learning setting. For example, based on pairwise embeddings, learning which inputs (ie, embeddings) lead to similar outputs, so that a controlled data format output can be learned.

[0052] Preferably, the groups are subdivided so that the target space of controlled data format outputs is reduced and fewer erroneous / inaccurate controlled data format outputs are determined.

[0053] In a preferred embodiment, the output processing algorithm includes an output machine learning model that formalizes the controlled data format output.

[0054] Preferably, the output machine learning model comprises a decoder neural network.

[0055] Preferably, the output machine learning model is trained to formalize the controlled data format output using embeddings.

[0056] In a preferred embodiment, the input processing algorithm includes determining an embedded input machine learning model.

[0057] Preferably, the input machine learning model comprises an encoder neural network.

[0058] Preferably, the input machine learning model is trained to determine the embedding of the acquired uncontrolled data format input in the embedding space.

[0059] Preferably, the input machine learning model is trained such that embeddings of semantically similar inputs are close to each other in the embedding space.

[0060] Optionally, the input machine learning model uses few-shot learning, where a limited number of labeled data for the new classification can be used to make predictions for the new classification. Further preferably, the input machine learning model uses zero-shot learning, where no labeled data for the new classification can be used to make predictions for the new classification.

[0061] New uncontrolled data format inputs (such as engineering or operational content, such as engineering intent) are represented in the same latent embedding space near similar previously processed representations. Therefore, new uncontrolled data format inputs will be identified and classified accordingly.

[0062] In a preferred embodiment, the input machine learning model is based on a Large Language Model (LLM).

[0063] Preferably, LLM is based on Natural Language Processing (NLP) techniques.

[0064] Therefore, using NLP techniques, it is possible to learn and obtain embeddings for inputs in uncontrolled data formats, such as engineering intent.

[0065] Thus, the system is trained and can be used to generate ontological engineering intent representations based on natural language engineering intent input and its abstract / latent semantic intermediate representation (i.e., embedding).

[0066] The trained models can be used in both directions, i.e. from uncontrolled data format input, such as natural language, and to determine formalized controlled data format output, such as ontology representation, and vice versa.

[0067] For example, this includes approaching the embedding from the "other side" and thereby converting ontological representations of engineering or operational content (such as engineering intent) into natural language representations of those engineering or operational content.

[0068] A use case for this is viewing controlled data format information, such as data / documents / structures built on the eBase format / tool: for example, the process automation industry uses eBase to handle its customer data. However, eBase representations vary from user to user and application to application, and customer data is also different, and cannot always be simply replicated even with the same tool. Therefore, a trained machine learning model converts the eBase representation(s) into a latent embedding space representation, and for example computes similarity scores in this latent embedding space (followed by engineering based on the automated controlled output format).

[0069] For example, LLM based on NLP technology takes natural language input, such as engineering intent, processes and transforms it, and produces output in a formal controlled data format, such as an ontology representation of engineering intent.

[0070] In a further extension, the resulting formal controlled data format output (e.g., ontology engineering intent representation) can be checked for consistency with the underlying target ontology model (e.g., underlying P&AE or eBase ontology model). These consistency checks can evaluate whether the resulting formal controlled data format output (e.g., ontology engineering intent representation) is available in the underlying P&AE or eBase ontology model (as concepts or relations or instances), and if they are not available there, this information can be used as feedback for the output machine learning model, for example by using standard penalty methods in mathematical optimization.

[0071] More preferably, if part of the ontology representation (e.g., concepts or relations or instances) is not available in the underlying P&AE or eBase ontology model, a constitutional AI approach (i.e., for example, supervising the underlying core LLM by another ML model in an extended supervised learning sense) can be used to penalize the formalization of the controlled data format output (e.g., the ontology engineering intent representation).

[0072] Additionally, the machine learning model (input and / or output) may include a pre-trained open source LLM, such as SentenceBERT or Dolly, for computing embeddings, etc. Alternatively, the machine learning model (input and / or output) may include an open source instruction-tuned LLM that may be fine-tuned (e.g., on an input-output model) to produce an ontology engineering intent representation. Alternatively, the machine learning model (input and / or output) may be obtained by fine-tuning a pre-trained instruction-tuned LLM, which may be fine-tuned using P&A engineering domain-specific engineering data, control narratives, P&IDs, eBase content, documents, etc.

[0073] Using pre-trained LLM-based systems (pre-trained ML models and embeddings) can achieve relatively good classification results (e.g., for intent topic modeling) with little / no overhead (i.e., fine-tuning using supervised ML and domain-specific P&AE intent vocabularies).

[0074] Optionally, after the embeddings are determined, principal component analysis is used to reduce the dimensionality of the embeddings and thereby allow them to be plotted on fewer dimensions. Based on this, in one example, a first / preliminary (transparent / interpretable) 2D visualization of the system output (i.e., formalization of the engineering intent based on the mapping of the embeddings) is provided to the user in order to determine whether his intent is correctly classified into similar intent formulations or clusters of similar suitable intents. Thus, if the engineer already believes here that his intent is incorrectly processed or classified, he can reformulate it before the system develops further in the wrong direction.

[0075] Alternatively, one can consider the entire engineering intent (e.g., sentence) or just individual words. For example, word embeddings can sometimes more easily match the embeddings / representations of ontology concepts.

[0076] Therefore, an improved method is provided to formalize uncontrolled data format input.

[0077] In a preferred embodiment, the input machine learning model and / or the output machine learning model comprises a neural network.

[0078] The neural network preferably comprises a transformer neural network having an underlying encoder-decoder architecture and one or more latent semantic representation layers, i.e. embedding layer(s).

[0079] According to one aspect of the present invention, a device for formalizing an uncontrolled data format input includes: an input interface configured to obtain an uncontrolled data format input; an input processing algorithm configured to determine an embedding of the obtained uncontrolled data format input in an embedding space of multiple pre-processed embeddings, wherein the embedding is represented by an embedding vector, wherein the embedding vector includes multiple vector entries specifying the embedding; an output processing algorithm configured to formalize a controlled data format output using the embedding; and an output interface configured to output the controlled data format output.

[0080] The input processing algorithm and the output processing algorithm are preferably provided by a processing unit.

[0081] According to one aspect of the invention, a computer program comprises instructions which, when executed by a computer, cause the computer to perform the steps of the method described herein.

[0082] According to one aspect of the invention, a computer-readable medium comprises instructions which, when executed by a computer, cause the computer to perform the steps of a method as described herein. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Exemplary embodiments of the present invention will be described below with reference to the accompanying drawings:

[0084] Figure 1 A schematic diagram illustrating a common workflow pipeline between process engineers and automation engineers;

[0085] Figure 2 A schematic diagram illustrating the main elements of the intended representation;

[0086] Figure 3 An exemplary instantiation of an intended representation is schematically shown;

[0087] Figure 4 schematically illustrates example steps for formalizing an uncontrolled data format input into a formalized controlled data format output;

[0088] Figure 5 schematically illustrates a transformer architecture for formalizing an uncontrolled data format input; and

[0089] Figure 6 Schematically shown is a very low-dimensional (2D) embedding space with multiple embedding vectors (visualized as 2D points). DETAILED DESCRIPTION

[0090] Figure 1 A schematic diagram of a generic workflow pipeline between a process engineer 100 and an automation engineer 200 is illustrated. The process engineer 100 determines engineering or operational content, such as engineering intent, during his workflow. This so-called intent-based engineering 310 includes, for example, intent modeling, intent formalization, abstract services and module representation, and output specifications as input to module pipeline generation 320. In the next step of the workflow pipeline, process engineering is connected to automation engineering and module pipeline generation 320 is performed. For example, this includes generating module pipelines, optimization and parameter conversion, and determining output specifications describing plant topology as input to an orchestration designer 330. In the next step of the workflow pipeline, an orchestration designer tool 330 is executed. For example, this includes further processing and manipulation of modules and plant topology (alarms, parameterization, etc.), as well as providing functions for creating module instances and exporting all content to a target control system (e.g., 800×A or Freelance).

[0091] In order to execute the above-mentioned workflow pipeline, the semantic facilitation and integration layer 340 provides a common language, terminology and relationships, for example in the form of an underlying ontology, and may therefore also include a common knowledge graph and a common API, so that the workflow pipeline between the process engineer 100 and the automation engineer 200 can work without any inconsistency and without losing any information.

[0092] However, to date, this workflow pipeline requires manual input from both the engineering domain expert 100 and the ontology expert to produce the ontology engineering intent representation in 310 , which is still prone to errors.

[0093] Figure 2 A schematic diagram illustrating the main elements of an intent representation R is illustrated. The intent representation R may contain information about an engineering intent defined by the properties and interrelationships of the intent elements of the engineering intent, the goals of the engineering intent, the requirements of the engineering intent, and the realization of the engineering intent. Thus, the intent representation R is the basis for exchanging engineering intent information between different stakeholders of an engineering workflow.

[0094] Figure 3Schematically, an exemplary instantiation of an intended representation R is shown. Compared to an overall view of the main elements of the intended representation R, Figure 3 An illustrative example of the engineering intent of a process engineer who wishes to maintain a constant pressure at a gas outlet is shown. The intent representation R shows how complex the engineering intent of a process engineer is by showing the interrelationships and dependencies between goals (including, for example, sub-goals), different necessary realizations, and different requirements.

[0095] Figure 4 Example steps of a process of formalizing an uncontrolled data format input into a formalized controlled data format output are schematically shown. In this case, the uncontrolled data format input includes engineering or operational content in the form of engineering intent expressed in natural language. For example, this could be: "I want to build an oil-gas-water separator" or "I want to maintain pressure."

[0096] In general, a new procedure is provided that automatically processes an uncontrolled data format input and formalizes it into a controlled data format output, in this case an intent representation R, which can be represented by Figure 1 All stakeholders of the workflow pipeline shown are used. In other words, the uncontrolled data format input is analyzed and the content of the uncontrolled data format input is classified. Then the engineering intent is extracted and the content of the uncontrolled data format input is rewritten to determine the intent representation R.

[0097] Therefore, several NLP model functions are merged into a higher-level process for engineering intent identification, classification, and formalization. This should allow automatic understanding of engineering intent, classification of engineering intent, and formalization of engineering intent, i.e., for example, conversion of engineering intent into an ontology representation.

[0098] Figure 5A transformer architecture for formalizing an uncontrolled data format input is schematically shown. The transformer architecture includes an input interface configured to obtain an uncontrolled data format input, which in this case includes engineering or operational content in the form of engineering intent I. An input processing algorithm executed by a processing unit provides an encoder Enc, which determines an embedding E of the obtained uncontrolled data format input I in an embedding space 30 generated by multiple pre-processed embeddings, wherein the embedding E is represented by an embedding vector, wherein the embedding vector includes multiple vector entries specifying the embedding E. An output processing algorithm executed by the processing unit provides a decoder Dec, which formalizes a controlled data format output using the embedding. The encoder Enc is provided using an input machine learning model 10 (e.g., including a large language model) to determine the embedding vector from the uncontrolled data format input. The decoder Dec is provided using an output machine learning model 20 to map the embedding vector of the uncontrolled data format input to a corresponding controlled data format output, such as an ontology representation R of the engineering intent I.

[0099] Figure 6 A very low dimensional (2D) embedding space 30 of embeddings E in the form of multiple embedding vectors (visualized as 2D points) is schematically shown.

[0100] In this case, the embedding space 30 includes six embedding vectors E1-E6. The first embedding E1 is a representation of an uncontrolled data format input, and its engineering intent is "I want to build an oil-gas-water separator". The second embedding E2 is a representation of an uncontrolled data format input, and its engineering intent is "I want to separate oil, gas and water". The third embedding E3 is a representation of an uncontrolled data format input, and its engineering intent is "I want to separate oil and gas". The fourth embedding E4 is a representation of an uncontrolled data format input, and its engineering intent is "I want to maintain pressure". The fifth embedding E5 is a representation of an uncontrolled data format input, and its engineering intent is "I want to control pressure". The sixth embedding E6 is a representation of an uncontrolled data format input, and its engineering intent is "pressure controller".

[0101] Based on the values ​​of different entries of the corresponding embedding vectors, the first embedding E1 to the third embedding E3 are similar to each other in a semantic sense, and therefore are close to each other in the embedding space. This group of first embedding E1 to third embedding E3 can be classified into one category (because "related to the intention of separating oil, gas and water"), and can be mapped to a similar controlled data format output, in particular, an ontological engineering intent representation. Similarly, the fourth embedding E4 to the sixth embedding E6 are similar to each other in a semantic sense, and therefore are also close to each other in the embedding space. This group of fourth embedding E4 to sixth embedding E6 can be classified into one category (because "related to the concept of pressure controller"), and when further processed in the above process, they can be converted into similar controlled data format outputs, in particular, converted into an ontological engineering intent representation again.

[0102] Therefore, determining the embedding in the form of an embedding vector in the embedding space allows for automatically formalizing a controlled data format output using the embedding based on an uncontrolled data format input.

Claims

1. A computer-implemented method for formalizing an uncontrolled data format input, comprising the steps of: Obtaining (S10) an uncontrolled data format input through an input interface; determining (S20) by an input processing algorithm an embedding (E) of the acquired uncontrolled data format input in an embedding space (30) generated by a plurality of pre-processed embeddings, wherein the embedding (E) is represented by an embedding vector, wherein the embedding vector comprises a plurality of vector entries specifying the embedding (E); formalizing a controlled data format output using the embedding (E) through an output processing algorithm (S30); The controlled data format is outputted through an output interface (S40).

2. The method according to claim 1, The uncontrolled data format input includes engineering or operational content.

3. The method according to claim 2, Wherein the embedding (E) of the acquired uncontrolled data format input is determined by the input processing algorithm comprises: identifying the engineering or operational content in the uncontrolled data format input; The embedding (E) is determined based on the identified engineering or operational content.

4. The method according to any one of the preceding claims, The uncontrolled data format input includes natural language input or uncontrolled text input.

5. The method according to any one of the preceding claims, Wherein the embedding (E) of the acquired uncontrolled data format input is determined by the input processing algorithm comprises: identifying in the embedding space (30) a preprocessed embedding that is similar to the embedding (E) in terms of a predetermined metric; The embedding (E) is determined based on similarity of the pre-processed embeddings.

6. The method according to claim 5, The predetermined metrics include semantically meaningful metrics.

7. The method according to any one of the preceding claims, The dimension of the embedding vector is greater than 10, preferably greater than 50, and further preferably greater than 100.

8. The method according to any one of the preceding claims, The step of formalizing the controlled data format output using the embedding vector comprises: The embedded vector is mapped to the controlled data format output.

9. The method according to any one of the preceding claims, Wherein the output processing algorithm includes an output machine learning model (30) that formalizes the controlled data format output.

10. The method according to any one of the preceding claims, Wherein the input processing algorithm comprises an input machine learning model (10) for determining the embedding (E).

11. The method according to claim 10, The input machine learning model (10) is based on a large language model.

12. The method according to any one of claims 9 to 11, Wherein the input machine learning model (10) and / or the output machine learning model (30) comprises a neural network.

13. A device for formalizing an input of an uncontrolled data format, comprising: An input interface configured to obtain an uncontrolled data format input; an input processing algorithm configured to determine an embedding (E) of the acquired uncontrolled data format input in an embedding space (30) of a plurality of pre-processed embeddings, wherein the embedding (E) is represented by an embedding vector, wherein the embedding vector comprises a plurality of vector entries specifying the embedding (E); an output processing algorithm configured to formalize a controlled data format output using said embedding (E); The output interface is configured to output the controlled data format.

14. A computer program comprising instructions which, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 12.

15. A computer readable medium comprising instructions which, when executed by a computer, cause the computer to perform the steps of the method according to claim 14.

Images