Method and apparatus for training a word-embedding method

Abstract

The invention relates to a method for training a word-embedding method (105), comprising the following steps – generating at least one training text dataset (10) comprising a multiplicity of lists (11

Description

The proposed solution relates to a method with the features of claim 1 and a corresponding apparatus with the features of claim 18. Search functions and methods are fundamental functionalities of operating systems, database systems, and information systems, which are used in particular in content and document management systems, information retrieval systems of libraries and archives, and search functions of websites in intranets and extranets. These search functions and methods relate to electronic documents (hereinafter referred to as documents) that contain at least some text and have been created or converted into file form by digitization (conversion into binary code). Without search functions, it would be almost impossible to search through extensive document collections, such as patent specifications. Search functions, methods, and engines are based on information technology principles of information and document retrieval (Introduction to Information Retrieval, Christopher D. Manning, Prabhakar Raghavan, and Hinrich Schiitze, Cambridge University Press. 2008), such as algorithms for converting and syntactically analyzing documents, efficient data structures for indexing document content, access algorithms optimized for these index structures, the avoidance of repeated calculations by caching results (see DE 10029644) and measurement methods that can be used to measure the degree of correspondence (referred to as “relevance”) of documents with regard to a search query. Conventional methods of information retrieval from unstructured textual information evaluate the “relevance” of documents based on the occurrence of search terms using statistical, probabilistic, and information-theoretical evaluation. An essential characteristic of search engines is the interpretation of the type of link between entered keywords. In practice, two types of links have become established: AND and ANDOR. With AND, only documents that contain all search terms are searched. With ANDOR, on the other hand, the search query is interpreted as a disjunctive link, but the result documents are weighted according to the number of search terms found per document in order to be able to also find similar documents. These conventional methods are usually based on term vectors, which symbolically represent documents as vectors in a high-dimensional space (e.g., with thousands to hundreds of thousands of dimensions). Each dimension of such a vector space represents a word. All dimensions together form the orthonormal basis of the vector space. File vectors, or document vectors, are formed as a linear combination of word frequencies or normalized word frequencies over the orthonormal basis. Since documents usually consist of only a fraction of all possible words, document vectors are a) usually “sparse” (only sparsely populated, many of the vector components are zero), b) discrete (each dimension only captures the meaning of one word), and c) this representation tends to produce “obstinate” documents (documents that are found as results for a wide variety of queries) solely due to the structure of high-dimensional spaces. (On the existence of obstinate results in vector space models, Milos Radovanovic, Alexandros Nanopoulos, Miijana Ivanovic, Proceeding of the 33rd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2010, Geneva, Switzerland, July 19-23, 2010, DOI: 10 1145 / 1835449.1835482) In particular, the discrete nature of this symbolic representation means that words with similar meanings are mapped to independent dimensions of the orthonormal basis and thus to independent components of the document vectors. In order to take term dependencies into account, this form of representation therefore requires additional knowledge to enrich the document vectors with information about similar terms and the degree of these term similarities. So-called semantic search methods determine the topics underlying the documents on a probability basis (US4839853, Latent Dirichlet allocation. David M. Biei, Andrew Y Ng, Michael I. Jordan in: Journal of Machine Learning Research, Vol. 3 (2003), pp. 993-1022, http: / / jmlr.csail.mit.edu / papers / v3 / blei03a.html, (last accessed June 6, 2023) and its variants) or 2 determine similarities between documents based on explicitly specified knowledge models in the form of conceptual models (linguistic models, semantic networks, word networks, taxonomies, thesauri, topic maps, ontologies, knowledge graphs). The topics identified by the first group of semantic search methods, also known as topic modeling methods, usually appear artificial, are rarely interpretable by humans, and often generate search results that are difficult to classify. The second form of semantic search methods uses predefined knowledge models to map the documents and queries to a common controlled vocabulary defined by the knowledge model [EP 2199926 A3 / US 000008156142 B2], thereby simplifying the search. The mappings of documents to the knowledge model are referred to as annotations, which may be enriched with additional terms from the knowledge model based on conceptual similarities. Knowledge models are used to enrich annotations, for example, to identify additional synonymous terms, generic terms of subordinate terms, or terms that have other relationships to the terms in the annotation. The degree of term similarity can be determined based on semantic distance (Conceptual Graph Matching for Semantic Search. Zhong J., Zhu H., Li J., Yu Y. in: Priss U., Corbett D., Angelova G. (eds) Conceptual Structures: Integration and Interfaces. ICCS 2002. Lecture Notes in Computer Science, vol. 2393. Springer, Berlin, Heidelberg) or the length of these implication chains from the knowledge models. The annotations enriched with additional terms correspond to an enrichment of the document vector with additional vector components representing these additional terms, provided that they were not already included in the vocabulary of the document corpus, or to an assignment of the vector components corresponding to the additional terms, provided that these terms were already included in the vocabulary of the document corpus. Search methods based on term models are currently the most widely used form of semantic search due to the high quality of the search results and the potential explainability of the results based on the network structure [EP 2562695A3, EP 2045728A1, EP 2400400A1, EP 2199926A2, US20060271584A1, US20070208726A1, US20090076839A1, WO2008027503A9, WO2008131607A1, WO2017173104]. However, there are several disadvantages associated with this last type of semantic search method 1) The methods are dependent on explicitly specified concept models. 2) If these models do not exist for a particular area of application, they must first be modeled. 3) The quality of the search results also depends on the quality of these models. 4) Due to this model dependency, these semantic search methods cannot be transferred to other areas of application. 5) These methods generally fail in the case of spelling errors and terms that are not included in the term models. 6) Since misspelled terms are generally not part of the term models and unknown terms cannot be part of the term models, these methods must be supplemented by additional methods for spelling error detection or correction and by conventional full-text search. The applications DE 102019212421 Al, EP 2020073304 deal with the problem of “semantic information retrieval based on word embeddings” (SIR), which consists of implementing a search function that works without explicitly specified background knowledge. The search should be carried out as efficiently as conventional information retrieval methods across any number of documents. It should output suitable documents sorted by similarity, taking into account the similarity of the terms used in them. And it should limit the number of results to such an extent that only truly comparable documents are considered. In addition, the results obtained should be comprehensible to a user. And the solution should be able to be used both for comparison with a user profile formulated in terms of the documents and for comparison of documents with each other. The principle of word embedding is well known. Well-known methods, such as Word2Vec (including its variants Paragraph2Vec, Doc2Vec, etc.), GloVe, and fastText, determine the semantics of individual words / terms and can thus replace explicitly specified term models. Words 4 in a language can be understood as connected character strings (alphanumeric characters, hyphens). A term can be understood as a superset of words, which may include additional punctuation marks or printable special characters, or may consist of several related words and terms. Reference is made to the following sources. Word2Vec: Efficient Estimation of Word Representations in Vector Space, Tomas Mikolov, Kai Chen, Greg Corrado, Jeffrey Dean , https: / / arxiv.org / abs / 1301.3781 (last accessed June 6, 2023). GloVe: Global Vectors for Word Representation, Jeffrey Pennington, Richard Socher, Christopher D. Manning, Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532-1543, October 25-29, 2014, Doha, Qatar.), fastText: Facebook's Artificial Intelligence Research lab releases open source fastText on GitHub, Mannes, John, https: / / techcrunch.com / 2016 / 08 / 18 / facebooks-artificial-intelligence-research-lab-releases-open-source-fasttext-on-github / (last accessed June 6, 2023). These methods are based on continuous - as opposed to discrete - term vectors (A Neural Probabilistic Language Mode, Yoshua Bengio, Rejean Ducharme, Pascal Vincent, Christian Jauvin; Journal of Machine Learning Research 3 (2003) 1137-1155). In these methods, terms / words are represented by a low-dimensional numerical vector, which usually comprises only a few hundred dimensions but, unlike a discrete term vector, uses all vector components. While in discrete representation the individual dimensions correspond to the orthonormal basis of the vector space, thus representing terms symbolically and documents as linear combinations of the orthonormal vectors, in continuous representation, words are represented as points (or vectors) in a space whose orthonormal basis can be interpreted as a subsymbolic representation of latent meanings (the words are quasi-embedded in the space of latent meanings). Words and documents in discrete representation lie on the hyper-edges and hyper-surfaces of a high-dimensional space due to their sparseness, but in continuous representation they usually lie in the middle of the space or its low-dimensional subspaces. To determine the positions of words in the vector space of continuous representation, the wordembedding methods described above use unsupervised machine learning methods. These learning methods use the context of words in the texts of a text corpus - i.e., their surrounding words - to determine the position of the word in the vector space. This has the effect that terms that occur in texts in the same or similar contexts are located close to each other in the vector space (see illustration in Fig. 1). From the word embeddings trained in this way, terms with similar content can be determined using different distance measures, such as Euclidean distance or cosine distance. Another measure is the so-called cosine similarity (see Manning et al. above), which is used to determine the similarity of vectors via their scalar product. The cosine similarity A can be used to determine whether two vectors point in the same direction (A = 1), point in similar directions (0.7 <A <1), are orthogonal (A = 0), or point in opposite directions (-1 <= A <0). While the cosine similarity A of conventional term vectors can only lie in the interval [0,1], it can lie in the interval [-1,1] for word embeddings. Doc2Vec or Paragraph2Vec (Distributed Representations of Sentences and Documents, Quoc Le, Tomas Mikolov Proceedings of the 31st International Conference on Machine Learning, Beijing, China, 2014. JMLR: W&CP volume 32. https: / / cs.Stanford.edu / ~quocle / paragraph vector.pdf (last accessed June 6, 2023)) extends the Word2Vec approach by taking into account document identifiers, which are used as separate terms during training. Like other terms, these identifiers are also embedded in the same vector space and can only be distinguished from words by the syntax of their identifiers. In contrast, documents and queries in the SIR method are represented by a linear combination of the word embeddings of their words and are represented in a separate document space of the same dimensionality. Document embeddings and query embeddings are generated by adding all word embeddings of the words in a document or query and then normalizing them with respect to the document or query length. A query embedding vector is known from Zamani, Croft; Estimating 6 embedding vectors for queries, in Proceedings of the 2016 ACM International Conference of the Theory of Information Retrieval, pp. 123-132, DOI, 10.1145 / 2970398.2970403. While the Word2Vec and Doc2Vec approaches view the words to be represented as atomic, the fastText approach (see Facebook's Artificial Intelligence above) goes one step further and represents words by the set of their N-grams (the set of all sequences of N consecutive sub-strings of the word). This extension allows morphological similarities between words (such as prefixes, suffixes, inflections, plural forms, spelling variations, etc.) to be included in the calculation of the position of the word vectors, so that the position of previously unknown words (“out-of-vocabulary” terms) can also be determined in the vector space. The fastText approach is therefore tolerant of spelling errors and unknown words to a limited extent. Due to the fastText approach of using N-grams, an approach based on it does have a certain tolerance for spelling errors and unknown words. However, it is not sensitive to well-formed words and allows even nonsensical character combinations to be compared as long as they contain at least one N-gram that also occurs in the training set. The main problem with these approaches, however, is that when words are represented in a continuous, subsymbolic vector space, each word has a distance to all other words, and all words are similar to each other, albeit to varying degrees. For example, the word “car” will be in close spatial proximity to “automobile,” “motor vehicle,” and “vehicle,” or their angles will be small and thus their cosine similarity will be large, the distance will increase to “vehicle,” “means of transport,” and “airplane,” the angle will be larger, and the cosine similarity will be smaller. However, this word will also have a distance to the words “chicken broth,” “plane,” “velvety,” “keel,” and “Ouagadougou” and form a very large angle with their vectors. There is therefore no criterion by which the “most similar” terms can be distinguished from the “unsimilar” ones. If the word embeddings of the words in a query or document are combined as described to form query or document embeddings, this problem is transferred: All documents are similar to all other documents, and a query is similar to all documents, but to varying degrees. In the publication “Soft Similarity and Soft Cosine Measure: Similarity of Features in Vector Space Model”. Sidorov, Grigori; Gelbukh, Alexander; Gomez-Adorno, Helena; Pinto, David, Computaciony Sistemas. 18 (3): 491 504. doi: 10.13053 / CyS-l8-3-2043. (last accessed February 6, 2019), the “soft cosine similarity” measure was introduced, which allows an additional weighting factor to be included in the calculation of the vector components of cosine similarity. This weighting factor can be used to include the similarity of individual words into the calculation of the similarity of document vectors. In principle, the cosine similarity of word embeddings could be used as word similarity. However, this is not feasible for a search function at runtime for efficiency reasons, since the scalar product of the document vectors would require comparing every word in a document or query with all words in another document. This can be circumvented at runtime by pre-calculating word similarities, but pre-calculation involves a quadratic effort with n*(n-l) / 2 comparisons. Even if each calculation only took one millisecond with a vocabulary of 100,000 words, calculating all similarities would take around 57.9 hours. Parallelization of the calculation would in fact be possible, but would require additional hardware. Even a method based on “soft cosine similarity” would also suffer from the problem that “everything is similar to everything else to varying degrees.” While a purely Boolean retrieval function (see Manning et al. above) can use the hard criterion of a term being contained in a document to limit the number of documents to those that contain the term, no approach based on term similarities provides an analogous hard criterion. Although KR 102018058449A describes a system and method for semantic search using word vectors, which also appears to be based on a similarity measure related to cosine similarity, it remains unclear whether this method is designed for discrete term vectors or continuous word embeddings. It seems reasonable to assume that this approach is subject to the similarity problem described above and returns all documents. US 20180336241 Al describes a method for calculating the similarity of search queries to job titles, which calculates query and document vectors from word embeddings, and a search engine that is used to find similar job offers, limited to the application area of job title searches. The specific structure of the search engine is not described, nor is the similarity problem addressed or how the number of search results can be limited. WO 2018126325 Al describes an approach to learning document embeddings from word embeddings using a convolutional neural network. Document embeddings of the presented invention, on the other hand, are calculated by linear combination of word embeddings. WO 2017007740 Al describes a system that uses contextual and, in contrast to the structural N-grams of fastText, morphological similarities in a special form of “Knowledge-powered neural NETwork” (KNET) to deal with rare words or words that do not appear in the document corpus. KNET can be considered an alternative approach to the use of Word2Vec, GloVe, or fastText in the present invention. US 20180113938 Al describes a recommender system for (semi-)structured data based on word embeddings. The determination of document embeddings follows a different principle. Here, too, the similarity problem is not addressed. The applications DE 102019212421 Al, EP 2020073304 describe another method of semantic search based on the similarity of word embeddings representing words, whose cosine similarity is in a dense vector space, and the clustering of the most similar and most important words. The method in applications DE 102019212421 Al, EP 2020073304 uses documents that contain tokenized character strings. It is therefore assumed that documents and queries have already been pre-processed and are available as tokenized sequences of character strings in a uniform character encoding. Tokenization means breaking down a text into individually processable components (words, terms, and punctuation marks). In a first step, an indexing procedure is used to calculate an inverted index (also known as a reverse index) for at least a subset of the documents. This means that a file or data structure is created in which each tokenized string indicates the documents in which it is contained. Next, word embeddings are calculated for the at least one subset of documents, i.e., the character strings are mapped to a vector with real numbers. Then, a document embedding is calculated for the at least one subset of documents by adding the word embeddings of all character strings, in particular words of the document, for each document and normalizing them with the number of character strings, in particular words, wherein SimSet groups of similar character strings are calculated before, after, or in parallel with the calculated word embeddings using a clustering procedure. Subsequently, in a query phase, a query expansion is performed in which i) query terms that occur in SimSet groups, or ii) query terms that do not occur in the SimSet groups but do occur in the documents, or iii) query terms that do not occur in the documents, in particular misspelled query terms are used for a preselection (in particular by means of the inverted index for the subset of documents) of the documents in order to achieve a restriction of the number of hits. Then a query embedding is performed. After that, a comparison of the query embedding with the document embeddings is performed using the previously calculated SimSet groups to quantitatively limit the number of document embeddings and preselected documents to be compared in order to automatically determine a ranking of the similarity of the documents and to display and / or save them. Based on this ranking, the documents most similar to the query or another document can be determined, for example. It should be noted that SimSet groups do not comprise documents, but words. In one embodiment of the method from applications DE 102019212421 Al, EP 2020073304, a CBOW model or a skip-gram model is used for word embedding. In a further embodiment of the method from applications DE 102019212421 Al, EP 2020073304, a non-parametric clustering method is used so that no a priori assumptions need to 10 be made. Hierarchical methods, in particular divisive or agglomerative clustering methods, can be used as clustering methods. It is also possible for the clustering method to be designed as a densitybased method, in particular DBSCAN or OPTICS. Alternatively, the clustering method can be designed as a graph-based method, in particular spectral clustering or Louvain. To restrict the search space, in one embodiment of the method from applications DE 102019212421 Al, EP 2020073304, a cosine similarity, a term frequency, and / or an inverse document frequency can be used as a threshold value for cluster formation. The method from applications DE 102019212421 Al, EP 2020073304 is based solely on the content of the documents / text documents provided; thus in the sense of (Hoppe 2020: Semantische Suche, Thomas Hoppe, Springer Vieweg, 2020, ISBN 978-3-658-30426-3) it only takes into account the actual language use of the authors, in the sense of Wittgenstein (Wittgenstein 1958 / 1977: Philosophische Untersuchungen, Ludwig Wittgenstein, suhrkamp taschenbuch wissenschaft, 203, 1977, translation of Philosophical Investigations, Basil Blackwell 1958) only their “language games.” Although a method based solely on the content of the documents / text documents provided is generally useful as an initial approximation / initial solution for use in a new area of application, two important components that influence the quality of the results are missing: 1) consideration of the language used by users and 2) consideration of the language used in standard vocabularies (such as KldB, NAIS, ECOS, MeSH, SNOMED, ICD, etc.) defined by standardization bodies. Traditional semantic search engines that use explicitly formulated vocabularies (such as taxonomies, classifications, thesauri, and ontologies), in which these standard vocabularies are often formulated (Hoppe 2020), have an advantage over the method described in applications DE 102019212421 Al, EP 2020073304. However, they have the disadvantage that they can only address the language usage of the documents to a limited extent (e.g., by using ontology-based entity recognition (OER), as in Hoppe, Thomas, Al Qundus, Jamal, Peikert, Silvio, Ontology-based Entity Recognition and Annotation, https: / / ceur-ws.org / Vol-2535 / paper 4.pdf, last accessed: June 11 6, 2023 in: Proceedings of the Conference on Digital Curation Technologies (Qurator 2020), Berlin, Germany, January 20th to 21st, 2020, https: / / ceur-ws.org / Vol-2535 / , last accessed: June 6,2023) and only recognize the terms that are specified in the respective vocabulary. Taking into account the language usage of users is problematic for both approaches, as this can only be learned gradually during the runtime of the system. Taxonomies and classifications are usually represented in a “tree” data structure. Thesauri and ontologies are represented as “directed acyclic graphs” (DAGs). The difference between these two symbolic, hierarchical data structures is that in trees, a node can only have a maximum of one predecessor node, while nodes in DAGs can have multiple predecessor nodes. By traversing these hierarchies, all paths from root nodes to leaf nodes (which no longer have successors) can be determined. In traditional semantic search engines, the similarity of terms is generally determined by distance measures in the symbolic, hierarchical data structures. Since these structures are principally based on graphs (consisting of nodes and edges), the path length between nodes of the graph is usually used to determine the distance. In the simplest version, each edge between the nodes is assigned a length of 1 and the path length is determined as the sum of all edge weights. In an extended version, the edges are weighted depending on other, mostly local criteria (number of outgoing edges, number of sibling nodes, depth of the node in the hierarchy, etc.) and / or the path length is determined as a weighted sum. If one now wanted to combine the path-length-based similarities of symbolic, hierarchical data structures with the similarities of word embeddings determined using the cosine measure in order to incorporate background knowledge (in the form of standards or the language usage of users) into the word embeddings in this way, one would face at least three fundamental problems: 1) Both similarity measures are based on different mathematical models and thus different assumptions that are not compatible. In particular, the similarity defined by path lengths is based on local information, while the similarity based on the cosine measure takes into account information about the similarity of word contexts encoded in the word embeddings. Furthermore, a mathematical model for mutual mapping is not yet known. 2) The training of word embeddings is based on artificial neural networks, which can be performed extremely efficiently using matrix calculations on GPUs. The calculations of path lengths in graphs, on the other hand, require either a. traversing the path to determine all path lengths (and thus similarities) between all node pairs, which must be precalculated and stored in a cache for efficiency reasons. b. or mapping the graph to an adjacency matrix. As experts know, both require quadratic growth in memory space O(nA2) in the worst case. 3) Technically, it is unclear how the cached similarities (2a.) and the adjacency matrices (2b.) can be calculated with the word embeddings on GPUs. An efficient calculation of the similarities of word embeddings and the representations of the similarities via path lengths using GPUs or parallel processing is not known thus far. The task is solved by a method with the features of claim 1. In a first step, at least one training text dataset is generated, which comprises a plurality of lists. The generation of the at least one training text dataset is based on at least one structured text dataset. Each of the at least one structured text datasets comprises at least one hierarchical data structure. Each of the at least one hierarchical data structures has the form of a directed, acyclic graph on a set of nodes assigned to the graph. One or more text elements are assigned to each node in the set of nodes. A text element may comprise, in particular, a word and / or a term. In other words, the respective structure of each of the hierarchical data structures of each of the at least one structured text dataset is based on a directed, acyclic graph. In particular, each of the at least one structured text dataset may comprise one or more hierarchical data structures, each of which arranges text elements by means of a graph structure (in the form of a directed, acyclic graph) or relates them to one another. Each of the plurality of lists of the at least one training text dataset is formed on the basis of a longest path of one of the directed, acyclic graphs of one of the at least one structured text dataset. The entries in a respective list comprise text elements that are assigned to the nodes of the represented longest path of the directed, acyclic graph. More precisely, the respective list contains an entry for each node of the associated longest path, wherein each entry in the respective list comprises at least one text element that is assigned to the node on which the respective entry is based. In a second step, the at least one training text dataset is used to train a word-embedding method. The word-embedding method assigns a representation, in particular a real-valued vector, to at least one subset of the set of text elements comprising the entries of the plurality of lists. In principle, one or more text elements can be assigned to each node of the set of nodes of one of the directed, acyclic graphs. In one embodiment, at least one of the directed acyclic graphs comprises at least one node to which several, i.e., at least two, text elements are assigned. The text elements of the at least one node can each represent alternative designations for each other. Alternative designations can sometimes include synonyms and abbreviations. For example, the text elements “Fraunhofer” and “Fraunhofer Society” could be assigned to a node. In this context, the word “Fraunhofer” would be used as an abbreviation for “Fraunhofer Society.” In one embodiment, all text elements assigned to the same node of a graph are alternative designations for each other. In addition or alternatively, the text elements of the at least one node may represent synonymous designations in different languages. In one embodiment, the plurality of lists of the at least one training text dataset comprises exactly one list for each of the directed acyclic graphs of at least one of the at least one structured text datasets and for each longest path of the respective directed acyclic graph. In particular, the plurality of lists may comprise exactly one list for each of the at least one structured text datasets and for each of the directed acyclic graphs of the respective structured text datasets and for each longest path of the respective directed acyclic graph. In a further embodiment, the at least one training text dataset may comprise more than one list for a longest path in one of the directed acyclic graphs. This is particularly the case if several text elements are assigned to at least one node of the associated graph. In this case, the at least one training text dataset may comprise a separate list for each of these text elements, each of which contains one of these text elements. In one embodiment, the plurality of lists for each of the directed acyclic graphs comprises at least one of the at least one structured text datasets for each longest path of the respective directed acyclic graph and for each selection of one text element along the nodes of the respective longest path, exactly one list whose entries correspond to the respective selection of text elements. In particular, the plurality of lists may comprise, for each of the at least one structured text datasets and for each of the directed acyclic graphs of the respective structured text dataset and for each longest path of the respective directed acyclic graph and for each selection of one text element along the nodes of the respective longest path, exactly one list whose entries correspond to the respective selection of text elements. The lists of the at least one training text dataset can each be understood as a “pseudo-sentence,” wherein the “words” of the “pseudo-sentence” correspond to the entries in the list. By means of the list structure, the at least one training text dataset can be understood as a tokenized document. The at least one hierarchical data structure of the at least one structured text dataset can relate text elements to each other. For example, the at least one structured text dataset can comprise at least one thesaurus, a word network, and / or an ontology. In addition or alternatively, the underlying directed acyclic graph of at least one of the hierarchical data structures of the at least one structured text data record may comprise a tree structure. In one embodiment, the at least one directed acyclic graph has exactly an edge from a first node to a second node (i.e., another node different from the first node) of the at least one directed acyclic graph if - a first text element is assigned to the first node, which represents a generic term of a second text element assigned to the second node, and - none of the other text elements of the same directed, acyclic graph represents both a generic term of the second text element and a subordinate term of the first text element. In another embodiment, the at least one directed acyclic graph has an edge from a first node to a second node (i.e., another node) of the at least one directed acyclic graph if - a first text element is assigned to the first node, which comprises a superclass of a second text element assigned to the second node, and - none of the other text elements of the same directed, acyclic graph represents both a superclass of the second text element and a subclass of the first text element. In one embodiment, the at least one structured text dataset is created by modeling a thesaurus or ontology from a user's search queries within one or more, in particular all available, user sessions. The modeling of a thesaurus or ontology is described, for example, in Hoppe / Tolksdorf 2018 (Guide for Pragmatical Modelling of Ontologies in Corporate Settings, Thomas Hoppe, Robert Tolksdorf in: Semantic Applications - Methodology, Technology, Corporate Use, Thomas Hoppe, Bernhard Humm, Anatol Reibold (eds.), Springer Vieweg, 1018). In a further embodiment, the at least one structured text dataset is created by combining a user's search queries made within one or more user sessions into a sequence. This sequence contains all of a user's consecutive search queries. In these sequences, individual words / terms can basically occur multiple times in direct succession. When creating the sequences, directly consecutive duplicates can be removed so that the word-embedding methods used obtain sentence-like contexts. In this embodiment, the at least one hierarchical data structure of the at least one structured text dataset is determined by the chronological sequence of the search queries made. The order of the entries in each of the plurality of lists can be predetermined by the order of the nodes in the associated longest path. In particular, the order of the entries in each of the plurality of lists can follow the order of the associated path, wherein the first entry in the respective list corresponds to a text element of a root node and the last entry in the respective list corresponds to a text element of a leaf node. Alternatively, the order of the entries in each of the plurality of lists may be opposite to the order of the associated path. In this case, the first entry in each of the respective lists corresponds to a text element of a leaf node and the last entry in each of the respective lists corresponds to a text element of a root node. The word-embedding method may comprise at least one of the methods known per se, namely Word2Vec, GloVe, FastText, Gauss2Vec, or Bayesian Skip-gram. In addition or alternatively, a CBOW model or a Skip-gram model may be used for the wordembedding method. The word embedding algorithm can basically be trained based solely on the at least one training text dataset. In addition, the training set of the word embedding algorithm may comprise further documents. The proposed method can be embedded in a method for preselecting and identifying similar documents from applications DE 102019212421 Al, EP 2020073304. The starting point of the method from applications DE 102019212421 Al, EP 2020073304 is a subset of a set of documents that contain tokenized character strings. In the proposed extension of the method from applications DE 102019212421 Al, EP 2020073304, the subset of documents for training the word-embedding method is supplemented with at least one training text dataset. In a further step, an inverted index is calculated for the at least one subset of the documents using an indexing procedure. In a further step, a document embedding is calculated for each of these documents for at least one subset of the documents by adding the word embeddings of all character strings, in particular words of the document, for each document and normalizing the result with the number of character strings, in particular words. In a further step, SimSet groups of similar character strings are calculated before, after, or in parallel with the calculated word embeddings using a clustering procedure. Subsequently, in a query phase, a query expansion is performed in which query terms that occur in SimSet groups or query terms that do not occur in the SimSet groups but do occur in the documents, or query terms that do not occur in the documents, in particular misspelled query terms, are used for a preselection of the documents in order to achieve a restriction of the number of hits. In a next, optional step, a query embedding is determined. Subsequently, a comparison of the query embedding with the document embeddings is performed using the SimSet groups formed with the clustering method to quantitatively limit the number of document embeddings of preselected documents to be compared in order to automatically determine a ranking of the similarity of the documents and to display and / or save these and / or metadata associated with the documents, in particular bibliographic data. The task is also solved by an apparatus having the features of claim 18. In one aspect, the proposed solution concerns the extension of word meanings represented by word embeddings and calculated by computer-based implementations with symbolically represented background knowledge in the form of classifications, thesauri, or ontologies. This problem is solved by supplementing the at least one training text dataset provided to the word embedding process with a preprocessing step. In this preprocessing step, lists (sequences) of terms are derived from symbolically represented background knowledge, which are considered “pseudo-sentences.” These “pseudo- sentences” are used to supplement the at least one training text dataset. This translates the knowledge symbolically represented in the background knowledge into a form that can be processed by the neural networks underlying the word embedding process. This step allows the background knowledge encoded in this way to be efficiently combined with the information from the document texts through implementations on conventional computer hardware, such as CPUs, or through matrix calculations on GPUs. The derivation of “pseudo- sentences” from symbolic representations means that their semantic relationships can be converted into a form that word-embedding methods can use to take vector representations into account when processing document texts using artificial neural networks, and these calculations can be performed efficiently using GPU hardware. In principle, all hierarchical data structures, or the graphs underlying the respective hierarchical data structures, of the at least one structured text dataset can have a similar / comparable structure or have been formed according to a similar scheme. For example, all hierarchical data structures of the same structured text dataset (or all structured text datasets on the basis of which the at least one training text dataset is generated) can each have a thesaurus. Alternatively, all hierarchical data structures of the same structured text dataset can each have a word network. However, the at least one structured text dataset may also comprise hierarchical data structures, each of which has different structures. The at least one training text dataset can basically be formed from one or more structured text datasets. The plurality of lists of the at least one training dataset can contain lists which are each based on different hierarchical data structures or on the graphs underlying the different hierarchical data structures. The hierarchical data structures can belong to the same structured text dataset or to different structured text datasets. The plurality of lists can be created on the basis of a subset of the hierarchical data structures of one or more structured text datasets. Each of the structured text datasets or hierarchical data structures can basically be used to create one or more training text datasets. In this sense, a structured text dataset can be reused. Embodiments are shown in connection with the following Figs. These show: Fig. 1 an example of clusters of similar terms in a set of approx. 73,500 documents; Fig. 2 an example of a structured text dataset; Fig. 2A a training text dataset generated from the structured text dataset in Fig. 2; Fig. 2B a further example of a structured text dataset; Fig. 3 a further example of a structured text dataset in the form of a thesaurus; Fig. 4 a further example of a structured text dataset; Fig. 4A an example of a training text dataset generated from the structured text dataset of Fig. 4; Fig. 5 a schematic representation of the process steps of the proposed solution; Fig. 6 a schematic representation of an indexing phase in one embodiment of the process; Fig. 6A examples of a word embedding and a document embedding; Fig. 7 a schematic representation of the determination of SimSet groups; Fig. 8A-C a determination of the most similar word embeddings for restricting a similarity graph; Fig. 8D an example of a SimSet for the example from Fig. 6A; Fig. 9 a schematic representation of the generation of a similarity graph as part of a clustering process; Fig. 10 a schematic representation of a query preparation; Fig. 11 a schematic representation of a case distinction of a query expansion; Fig. 12 a schematic representation of a document retrieval. Fig. 2 shows an embodiment of a structured text dataset 20. In the embodiment shown, the structured text dataset 20 consists of a hierarchical data structure. In principle, however, the structured text dataset could also comprise several hierarchical data structures. The hierarchical data structure is based on a graph structure consisting of a set of nodes 30-1 to 30-7 and a set of edges. Using the underlying graph, words / terms (also referred to generally as text elements) are related to each other in the hierarchical data structure of the structured text dataset 20. In the embodiment shown, exactly one text element is assigned to each node of the underlying graph. In the embodiment shown, for example, the text element 40-n is assigned to a node 30-n, where n is a natural number between 1 and 7. In principle, however, several text elements can also be assigned to a node. The edges of the graph are used to capture the relationships between the text elements. The edges of the graph connect two nodes of the graph, in each case. The edges of the graph are directed. This means that each edge of the graph points from a start node to an end node. This is referred to as a directed graph. In the embodiment shown, an edge is represented by an arrow pointing from the start node to the end node. For example, edge 50 points from start node 30-3 to end node 30-5. A root node of a graph is a node that has no predecessor nodes (parent nodes). This means that no edge from any other node to a root node is shown. A leaf node is a node that has no successor nodes, meaning that no edge from the leaf node to any other node is shown. In the embodiment shown, node 30-1 is a root node and nodes 30-2, 30-5, and 30-7 are leaf nodes. A route or path in a graph is a sequence of nodes vi,... ,vn in which each two consecutive nodes are connected by an edge. If there is no path with vi=vn in the graph, then it is called acyclic. A path or route is referred to as a longest route / path if it can no longer be extended by adding further nodes. A longest path in a directed, acyclic graph is always a path from a root node to a leaf node, in other words a path with a sequence of nodes that includes a root node as its first link and a leaf node as its last link. In the embodiment shown, for example, the node sequences 30-1, 30-2, and 30-1, 30-4, 30-6 are each paths. The path 30-1, 30-2 is a longest path. The path 30-1, 30-4, 30-6 is not a longest path because it can be supplemented by the node 30-7. The longest paths in the graph of structured text dataset 20 in Fig. 2 are given by the following four node sequences: 30-1, 30-2; 30-1, 30-3, 30-5; 30-1, 30-4, 30-5; 30-1, 30-4, 30-6, 30-7 The set of text elements along a path comprises those text elements that are assigned to the nodes of the path. For example, the text elements 40-1 and 40-2 are assigned to the nodes along the path 30-1, 30-2. Fig. 2A shows a training text dataset 10 that was generated from the structured text dataset 20 in Fig. 2. The training text dataset 10 comprises 4 lists 11-1, 11-2, 11-3, and 11-4. Each of the lists 11-1, 11-2, 11-3, and 11-4 comprises as entries the text elements of the nodes along a longest path in the graph underlying the structured text dataset 20. The training text dataset 10 is generated by traversing each longest path of the graph underlying the structured text dataset 20 and listing the text elements along the respective path as entries in a list. In one embodiment, the text elements are listed in the order of the associated nodes in the lists. In this embodiment, the lists each begin with a text element belonging to a root node and end with a text element belonging to a leaf node. Alternatively, all lists can be listed in reverse order. In this case, the respective lists begin with a text element belonging to a leaf node and end with a text element belonging to a root node. In other words, the training text dataset 10 is generated by traversing a hierarchical data structure, the structured text dataset 20. In doing so, the sequence of all nodes of a path from the root node to the leaf nodes is determined and the designation (the associated text element) of each node of the path is collected in a list. This list can be interpreted as a “pseudo- sentence.” These “sentences” place each term in the context of their predecessor or successor terms. By also taking into account the alternative designations and synonyms of normative designations and constructing their “pseudo- sentences” for alternative designations and synonyms, they can be embedded in the same contexts as the normative designations. It should be noted that, in general, a node of a graph underlying one of the hierarchical data structures in the structured text dataset is not determined by the text element(s) assigned to it. In principle, different nodes in the same or different graphs of the structured text dataset can be assigned the same text elements (see the example in Fig. 4). Different structured text datasets can, in principle, generate the same training text dataset when the method is applied. For example, structured text dataset 20, which is shown in Fig. 2B, leads to training text dataset 10, which is shown in Fig. 2A, just like the structured text dataset in Fig. 2. The structured text datasets in Fig. 2 and Fig. 2B are largely identical, with the only difference being that node 30-5 of the structured text dataset in Fig. 2 has been “duplicated” into nodes 30-5 and 30-8 of the structured text dataset 20 in Fig. 2B. However, this change in the graph structure ultimately does not alter the resulting longest paths and thus also does not alter the resulting training text dataset 10. Fig. 3 shows an embodiment of a structured text dataset 20 in the form of a thesaurus. The thesaurus relates words on the subject area “education / profession” to each other according to the criteria “is a generic term of’ or “is a subordinate term of’. The relationship between the words / terms is captured by the underlying graph. For example, “education” is a generic term of “vocational training.” This is captured by an edge from the node with the assigned word “education” to the node with the assigned word “vocational training.” Fig. 4 shows a structured text dataset 20 which, similar to the structured text dataset in Fig. 3, represents an (excerpt) from at least one thesaurus and relates words to the subject area “education / profession” according to the criteria “is generic term of’ or “is subordinate term of.” The structured text dataset 20 in Fig. 4 comprises two hierarchical data structures 21-1 and 21-2. The graphs underlying the hierarchical data structures 21-1 and 21-2 each have two nodes, each of which is assigned more than one text element. For example, the three words “plumber,” “tinsmith,” and “sheet metal worker,” which are alternative terms, are assigned to node 31 of the graph underlying hierarchical data structure 21-1. The four words “plumber,” “tinsmith,” “sheet metal worker,” and “fitter” are assigned to node 32 of the graph underlying hierarchical data structure 21-2. If several text elements are assigned to a node, the respective text elements are embedded in the same context. This is because the context of a text element of a node is determined within a hierarchical data structure by the text elements of the predecessor and successor nodes of the graph underlying the hierarchical data structure. Fig. 4A shows a training text dataset 10, which is generated from the structured text dataset 20 described with reference to Fig. 4. Each list of the training text dataset 10 is created on the basis of a longest path in one of the graphs of Fig. 4. In the embodiment shown, a list comprises one entry for each node of the associated longest path. The entry for a node comprises exactly one of the text elements assigned to the node. Since the nodes of the graphs in Fig. 4 are sometimes assigned several text elements, different lists (i.e., lists whose entries differ) can be created based on a longest path, since only one text element per entry is selected in the lists. In the embodiment shown, the training text dataset 10 includes all lists of such possible combinations for each longest path. In this way, the training text dataset 10 in the embodiment shown takes alternative designations into account. If two alternative designations are assigned to the same node of one of the graphs, the training text dataset 10 includes at least a first and a second list for each longest path of the graph that passes through this node, which differ only in the entry belonging to the node. The first list contains the first of the two alternative designations for this entry, and the second list contains the second alternative designation for the entry. As a result, alternative designations in the lists (“pseudo-sentences”) of the training text dataset 10 are listed in the same contexts in each case. Another embodiment for creating the lists of a training text dataset “ignores” alternative designations. In this embodiment, multiple text elements, for example in the form of alternative designations, can in principle also be assigned to the nodes of one of the graphs. However, only one list is provided in the training text dataset for each of the alternative designations. In this embodiment, only one list is formed along the nodes of the path for each longest path in the graph of the structured text dataset 20. In this case, one of the text elements assigned to the nodes of the path is selected for each entry. In an alternative embodiment, at least a subset of the nodes of one of the graphs of the structured text dataset 20 is assigned several text elements, which represent foreign-language designations of the same term. Analogous to the embodiments described above, foreign-language designations can be taken into account or ignored when creating the training dataset 20. Depending on whether separate lists are formed for all text elements of the nodes (in this variant, foreign-language designations are taken into account) or whether exactly one list is to be created for each longest path (in this variant, foreign-language designations are ignored). The proposed solution also includes hybrid forms in which text elements assigned to the same node of one of the graphs comprise both alternative designations (in particular synonyms or abbreviations) and foreign-language designations for the same term. By taking alternative designations and / or foreign-language designations into account, the training text dataset 10 grows exponentially. However, this expansion of the training text dataset 10 is advantageous depending on the area of application. In the method for determining similar documents described below with reference to Figs. 6 to 12, for example, this also enables foreign-language queries to be answered. Fig. 5 schematically illustrates how a word-embedding method 105 is trained using the proposed method. The starting point is a structured text dataset 20. In principle, however, the method can also be carried out based on several structured text datasets. The structured text dataset comprises at least one hierarchical data structure 21, which relates text elements to each other by means of a (hierarchical) graph structure. In one embodiment, the structured text dataset 20 comprises standard vocabularies, which are available, for example, in the form of a thesaurus, a word network, or an ontology. In addition or alternatively, the structured text dataset 20, or the at least one hierarchical data structure 21 of the structured text dataset 20, can be specially created in the course of the process, for example to represent the language use of users. For this purpose, a sequence of search queries / search terms entered by a user in a usage session can be used, for example. As described in Hoppe / Tolksdorf 2018, a thesaurus or ontology can be modeled from the search queries / search terms used in order to generate / supplement the structured text dataset 20. In addition or alternatively, all search queries made by a user can be combined as a sequence in the order in which they are made within one or more usage sessions (with directly consecutive duplicates being removed). The structured text dataset 20 can also be generated in this way. The hierarchical graph structure of the structured text dataset 20 is given by the chronological sequence of the search queries made. A training text dataset 10 is generated from the structured text dataset 20, as described, for example, with reference to Figs. 2A and 4A. In principle, however, several training text datasets can also be generated from this. The training text dataset 10 comprises at least one list 11 whose entries comprise a subset of the text elements of the structured text dataset 20. The entries of list 11 correspond to text elements along a longest path in the graph, which underlies the at least one hierarchical data structure 21 of the structured text dataset 20. Since multiple text elements can generally be assigned to the nodes of a graph, different lists can generally also be created based on a longest path, for example by selecting only one of the text elements of a node for each entry in a list. In principle, when generating the training text dataset 10, all longest paths of the graph underlying the at least one hierarchical data structure 21 of the structured text dataset 20 can be recorded and a list can be generated for each of these longest paths, which list comprises text elements along the respective longest path. However, when generating the training text dataset 10, several, in particular all, hierarchical data structures of the structured text dataset 20 can also be used. In the sense that the longest paths on which lists of the training text dataset 10 are based belong (at least in part) to graphs of different hierarchical data structures of the structured text dataset 10. Hierarchical data structures of different structured text datasets can also be used to generate a training text dataset. The lists of the training text dataset 10 can be understood as “pseudo-sentences.” Each entry in list 11 corresponds, for example, to a word / term in a sentence of a natural language. Consequently, 26 when applying a word-embedding method, the training text dataset 10 can be treated just like any other (text) document. The entries in the lists of the training text dataset 10 divide the “pseudo-sentences” into text elements. This division corresponds to the tokenization of a document. Consequently, the training text dataset 10 can be treated as a tokenized document. The training text dataset 10 is then used to train a word-embedding method 105. The wordembedding method 105 assigns a representation, in particular a real-valued vector, to text elements. To calculate word embeddings 105, methods known per se such as Word2Vec, GloVe, FastText, Gauss2Vec, or Bayesian Skip-gram are available, which calculate word embeddings 105 for a lowdimensional, continuous word vector space. Word embedding 105 is the collective term for a series of language modeling and feature learning techniques in natural language processing (NLP) in which character strings from a vocabulary, in particular a lexicon, are mapped to vectors of real numbers, which are referred to as word embeddings. Conceptually, this involves the mathematical embedding of a space with many dimensions into a continuous vector space with a lower dimension. In one embodiment, the word embedding algorithm 105 is trained solely on the basis of the training text dataset 10. In addition or alternatively, the training text dataset 10 can be used together with a set of additional (tokenized) documents to train a word-embedding method 105. Basically, word embeddings use context information from words within a sentence to arrange vector representations of the words based on their context words using an artificial neural network with a gradient method in close spatial proximity to the vector space. Words within the “pseudosentences” are therefore arranged in close spatial proximity using the word-embedding method, provided they are embedded in the same context words. This not only allows adequate word embeddings of alternative terms and synonyms to be calculated, but also allows them to be related to the terms used in document texts, provided they are used in similar contexts. Word-embedding methods are based on analyzing the context of a word in a sentence (the word usage) and using this context to adjust the vector representation of the word, thereby positioning words that are used in the same or similar contexts in the same regions of the vector space. The context of a node in a hierarchical data structure is formed by its predecessor and successor nodes. By collecting the names of the nodes on the path in the order of the predecessor / successor nodes, the “pseudo-sentence” [education, vocational training, craftsman, metalworker, fitter, master fitter] can be represented with the interpretation “education is a generic term for vocational training is a generic term for craftsman is a generic term for metalworker is a generic term for fitter is a generic term for master fitter (cf. Fig. 3). By taking into account all combinations of alternative terms and synonyms, their common contexts of meaning can also be captured, as shown, for example, in the training text dataset in Fig. 4A. By converting all potential longest paths of the hierarchy into “pseudo-sentences” in this way, all different combinations of interpretations of the terms can be mapped. With reference to Figs. 6 to 12, the method from applications DE 102019212421 Al, EP 2020073304 is described, which is extended with the help of the embodiments described above. For the sake of simplicity, the description assumes a structured text dataset with a hierarchical data structure, which is used to generate a training text dataset. In principle, however, the method can also be applied to several structured text datasets. Each structured text dataset can also comprise several hierarchical data structures. In principle, several training text datasets can also be used. With the help of the proposed solution, the search results of the method described in DE 102019212421 Al, EP 2020073304 are improved by adapting the word-embedding models derived from the document texts to the language usage of any existing standard vocabularies and / or to the language usage of users. This is achieved by using at least one additional training text dataset to train the word-embedding models, which captures any existing standard vocabularies and / or the 28 language usage of users. These adjustments are then also taken into account in the SimSets (central innovation in DE 102019212421 Al, EP 2020073304). Only the proposed solution enables this adjustment of the word-embedding models. This is because the proposed solution uses any existing standard vocabularies and / or user language usage to create a data structure that can be used by word-embedding methods. In addition, newly added documents and the language usage of searchers can also be continuously taken into account and / or the word-embedding models and SimSets can be adjusted at regular intervals. The method described in applications DE 102019212421 Al, EP 2020073304 takes place in two phases: the indexing phase and the query phase. The indexing phase serves to build efficient data structures, while the query phase is used to search for documents in these data structures. These two phases can optionally be supplemented by a third phase, the recommendation phase. Indexing phase The sequence of processing steps in the indexing phase is shown schematically in Fig. 6. The starting point is a set of documents 101, each of which is available as tokenized sequences of character strings. The documents in set 101 can each be linked to metadata, in particular bibliographic data. The respective documents can be identified, for example, on the basis of their metadata. An indexing process 102 is used to calculate an inverted index 103 for these documents 101. This inverse index 103 enables rapid access to all documents 101 containing given character strings, such as words and / or terms, based on the character strings contained in the documents 101. In a further step, a word embedding process 104 is trained / calculated on the basis of the set of documents 101 supplemented with at least one training text dataset 10, as already described with reference to Fig. 5. Supplementing the documents 101 with the at least one training text dataset 10 has an influence on the resulting word embeddings 105. In particular, words / terms that were not present in the documents 101 may occur in the at least one training text dataset 10. Consequently, the scope of the word embeddings 105 can be expanded. Furthermore, a word / term that already occurs in one of the documents 101 may appear in the at least one training text dataset 10 in new contexts compared to the documents 101, given by the “pseudo-sentences” in which the word / term occurs. This can influence the resulting word embedding 105 of the respective word / term. In one embodiment, the CBOW model is used to calculate the word embeddings 105, which enables the prediction of words based on context words. In another embodiment, a skip-gram model can be used instead of CBOW, which enables the prediction of context words for a word. These calculation methods ensure that the word vectors of similar terms (terms that are frequently used in the same context) are arranged in close spatial proximity to each other in the word vector space. For the documents in document set 101, document embeddings 107 are also calculated, 106, by adding the word embeddings 105 of all character strings for each document and normalizing them with the number of words. This avoids numerical overflows and dependencies of the document embeddings 107 on the document length, so that documents of different lengths can still be meaningfully compared with each other. Since documents that use the same or very similar words (i.e., character strings) are highly likely to deal with similar or related topics, adding their word embeddings 105 results in their document embeddings 107 being arranged in close spatial proximity to each other in the document vector space. Fig. 6A shows examples of a word embedding 105 and a document embedding 107. In this example, the document set to be examined contains only one sentence: “A police officer is an official.” This results in four vectors for the word embedding 105 and one vector for the document embedding 107. In a further step, a clustering process 108 is used to determine groups of very similar character strings / words from the word embeddings 105, which are referred to below as SimSet groups 109. This step can also be performed before, after, or in parallel with the document embedding 107 determination step. Since the number of potential groups of similar words is unknown, a non-parametric clustering method 108 is used, in which the number of clusters does not have to be specified. The methods that can be used include hierarchical methods such as divisive clustering, agglomerative clustering, and density-based methods such as DBSCAN, OPTICS, and their various extensions. In one embodiment, graph-based methods such as spectral clustering and Louvain can also be used. This embodiment for calculating SimSets 109 is shown in Fig. 7. For graph-based clustering of word embeddings 105, the similarities between all word embeddings 105 are considered as weighted edges in a graph - referred to as a similarity graph -108.4, whose nodes are formed by the word embeddings 105. In one embodiment, both the words / terms that occur in at least one of the documents in the document set 101 and the text elements of the supplementary training text dataset 10 are taken into account. The weighting of the edges corresponds here to the degree of similarity. In a naive solution, this graph would be completely linked, since each word embedding has a distance or similarity to all others. The graph would therefore comprise n*(n-l) / 2 edges, and clustering would require searching through an exponential number of clusters (potentially 2n subsets). Determining the optimal clusters would therefore be NP-hard. With two restrictions, both the number of nodes in the similarity graph to be considered and the number of edges to be taken into account can be drastically reduced. In the context of a search that takes into account similar words in addition to the actual query, it is sufficient to consider the character strings / words that fall into a special form of clusters -referred to as SimSets 109. These strings / words should a) occur frequently in the text corpus (measured by term frequency, TF, see Manning et al.), b) have a high information content (measured by inverse document frequency, IDF, see Manning), and c) be very similar to each other. The term frequency mentioned under a) measures the frequency of words / strings / text elements both in the documents of document set 101 and in the training text dataset 10. It is basically not necessary to distinguish whether the respective word / text element or character string originates from one of the documents in document set 101 or from training text dataset 10. To calculate the inverse document frequency mentioned under b), the “pseudo-sentences” of the training text dataset 10 can each be considered as separate “pseudo-documents.” Since term frequencies in a corpus follow a power distribution, it is sufficient to satisfy the Pareto principle and select those terms that comprise, for example, 80% - 95% of all terms with the largest combined TFIDF (term frequency multiplied by inverse document frequency) (cases a and b above combined) of the corpus. The specific value can be used as an importance threshold to control the number of SimSets 109. In one embodiment, the set of nodes of the similarity graph is determined as follows: When determining the number of nodes, all text elements / character strings / words that appear in at least one of the documents in document set 101 or training text dataset 10 are first considered, as a matter of principle. The term frequency (TF) is determined for each of these text elements / character strings / words. The term frequency indicates how often the text element / character string / word occurs in total in document set 101 (or a subset thereof) or in training text dataset 10. In addition, the inverse document frequency (IDF) is determined for each of these text elements / character strings / words. The inverse document frequency indicates which of these text elements / character strings / words have a high information content. In one embodiment, each “pseudo-sentence” of the training text dataset 10 is evaluated as a separate “pseudo-document.” Forthose strings / words / text elements of the documents in document set 101 or the training text dataset, whose combined TFIDF (term frequency multiplied by inverse document frequency) is high in comparison to other words / terms or text elements, a node is created in the similarity graph. As a result, the similarity graph in the method adapted with the proposed solution basically includes both nodes for character strings / words / terms of the documents in document set 101 and nodes for additional words / terms (text elements) of the training text dataset 10. The similarity measurement of word embeddings 105 using cosine similarity (above under c) is shown in Fig. 8A-C. Fig. 8 A shows the similarity of all word embeddings 105 to a given word embedding (dashed reference vector). Then, for example, all word embeddings 105 with negative similarity - cosine similarity <0, angle >90° - can be excluded (Fig. 8B, hatched half-plane). One could also set a similarity threshold based on the cosine similarity for similarity in a range of less than 0.87 to 0.7, and thus ignore all word embeddings with an angle between 90° and 45° to 60° as dissimilar (Fig. 8C, bold hatched segments). What remains are the word embeddings 105 that are most similar to the dashed reference vector, those with an angle of at most 30° - 45°. These can then be used as nodes of the similarity graph. Edges can be created in the similarity graph for each of the associated nodes. The specific value of the similarity threshold controls the size - in terms of the number of terms - of the SimSets 109. Fig. 8D shows the calculation of the cosine similarity for the example set from Fig. 6A. The shading in the individual cells corresponds to the hatching in Figs. 8A A Fig. 8D shows the numerical values for the cosine similarity, which are arranged symmetrically. On the main diagonal, the similarity values are naturally 1. In a first step, the negative similarities (e.g., police officer - is) can be sorted out, which corresponds to the situation in Fig. 8B; i.e., only the positive half-plane is considered. Positive numerical values below a similarity threshold (here 0.75) are highlighted in dark gray and correspond to the narrowing of the angle range in Fig. 8C. The word “one”, for example, has only a slight similarity to the words “police officer,” “is,” and “official.” This leaves the word pair “police officer” and “official” as the relevant value above a similarity threshold of 0.75 (and outside the main diagonal) with a similarity of 0.7533. These two words then form a SimSet group 109 for the sample document set. Based on this observation (and with reference to Figs. 7 and 9), the similarity graph 108.4 can be constructed as follows 108.3: For each word in the document set 101 or each text element in the training text dataset 10, the combined TFIDF measure is calculated and sorted, 108.1, and a reduced word list (i.e., list of character strings) 108.2 sorted in descending TFIDF order is obtained. To extract the similarity graph 108.3, these words / character strings / text elements are gone through in sequence and for each word / character string / text element with a TFIDF above the importance threshold, the first decision process shown in Fig. 9 is performed. In the event of a negative result in one of the three comparisons, the respective string, word, term, or text element is discarded (not shown in Fig. 9). For each word / character string / text element, the most similar words / strings whose cosine similarity exceeds the similarity threshold are determined for its word embedding 105 (second decision process in Fig. 9). For the corresponding words, undirected edges and / or nodes are created in the similarity graph, if they do not already exist. As a result, the similarity graph always includes an undirected edge between two nodes if the cosine similarity of the associated text elements / words / character strings exceeds a similarity threshold. Each undirected edge is assigned a weight. The weight of an undirected edge corresponds to the specific cosine similarity between the words / text elements / character strings whose associated nodes connect the undirected edge (step 108.3 in Fig. 9). The similarity graph constructed in this way contains all nodes with high TFIDF values and / or that have a similarity to each other greater than or equal to the similarity threshold, and connects those nodes by undirected edges that have a similarity to each other greater than or equal to the similarity threshold. This graph has the property that all nodes that are in close spatial proximity in the word vector space are more closely connected to each other than to nodes that are further away. In the similarity graph 108.4, a graph-based clustering method such as Louvain (“Fast unfolding of communities in large networks”. Blondel, Vincent D; Guillaume, Jean-Loup; Lambiotte, Renaud; Lefebvre, Etienne, Journal of Statistical Mechanics: Theory and Experiment. 2008 (10): P10008. arXiv:0803.0476. Bibcode:2008 JSMTE..10..008B. doi: 10.1088 / 1742-5468 / 2008 / 10 / P10008 (last accessed February 6, 2019)), clusters of words / character strings can be identified that are highly similar to each other and are distinct from clusters of words / character strings to which they are less similar. These clusters of similar words are stored as SimSets 109 for further use. In one embodiment, the SimSets 109 are made accessible for efficient retrieval via another inverted index. This allows for quick identification of whether a given word is contained in a SimSet 109 and, if so, in which one. The same mechanism (an inverted index) can be used for this as for determining the documents that contain a given word. Query phase Responding to a search query for similar documents using the data determined in the indexing phase is done in two steps. In the first step, query preparation, a query 201, which is available as a tokenized sequence of character strings, is prepared by calculating a query embedding 205 for it, analogous to a normal document. In the second step, retrieval, this query embedding 205 is compared with the document embeddings 107 of potentially relevant, preselected documents 204, and these documents and / or document embeddings are sorted according to their similarity in order to then be displayed and / or 35 stored. This comparison is made with the Sim Set groups 109 formed in the clustering process in order to quantitatively limit the number of document embeddings 107 to be compared. A ranking of the similarity of the documents is then automatically determined, displayed, and / or stored. Query preparation The query preparation process is shown in Fig. 10. Query preparation consists of several parts: the calculation of the query embedding 104 for a query 201, which is analogous to the calculation of the document embeddings 106 and results in a query embedding 205, a query expansion 202, and a document selection 203. Since every document in the document vector space is similar to all others (albeit to varying degrees), this also applies to the analogously constructed query embedding 104. However, this would mean that all documents would always be found for every query, as there is no strict selection criterion. To construct an appropriate selection criterion, a query expansion 202 is performed for query 201. In query expansion (see Fig. 11), a distinction is made between a) query terms that occur in SimSets 109, b) query terms that do not occur in the SimSets but do occur in the corpus (i.e., the documents 101), c) query terms that do not occur in the corpus. This also includes misspelled query terms. In case a), query expansion consists of preselecting, for each SimSet 109 containing a query term, those documents that contain at least one of the SimSet terms (202.1 in Fig. 11). Although this approach has the disadvantage of ignoring documents that contain terms with a lower degree of similarity, the advantage lies in a greatly reduced number of hits (analogous to a Boolean search) and the explainability of the hits via the terms of the SimSets. In case c), when implemented using an embodiment that uses Word2Vec word embeddings, the preselected documents can be set to the empty set (202.3 in Fig. 11). When implemented using an embodiment that uses fastText word embeddings, such preselection of documents would not be possible in cases b) and c), as this variant can handle 36 spelling errors and “out-of-vocabulary” terms. In order to achieve a reduction in the number of hits for these cases, the following consideration implies a solution: As described, SimSets 109 consist of terms that 1) have a high TFIDF and 2) are very similar to each other. This means that there may be individual query terms that are contained in the corpus but not in a SimSet 109 and yet have a similarity above the similarity threshold to a query term. In case b) and in the embodiment that uses text word embeddings, also in case c), the document terms that have a similarity above the similarity threshold but are not contained in the SimSets can be determined for these query terms via the word embeddings 105 (202.2 in Fig. 11). These terms can then also be used for query expansion to perform a preselection 203 of the documents. The preselected documents 204 are transferred to the retrieval for comparison with the query embedding 205. If possible, SimSets are used in query expansion 202 to expand queries in a manner analogous to conventional semantic search (see Fig. 11). Since the expanded queries are used to retrieve document candidates from the inverted index, the method delivers an expanded result set analogous to a conventional search, but without running into the described problem of unrestricted retrieval that a purely word embedding-based approach would entail. This method thus delivers expanded but quantitatively limited results compared to a full-text search. Retrieval After the documents have been preselected 204 and the query embedding 205 has been calculated, the retrieval takes place, as shown in Fig. 12. For this purpose, the document embeddings 107 of the preselected documents 204 are compiled with the selected document embeddings 302. For each document embedding 302, the cosine similarity to the query embedding is calculated using the cosine similarity measure, and the documents are sorted in descending order of similarity to the document ranking 304. In one embodiment, the calculation can be parallelized using a known map-reduce architecture in order to efficiently process even very large amounts of documents. Since, as described, the cosine similarity of a continuous vector space representation can also take negative values, an additional filter criterion can be used during document ranking 304 to further restrict the set of search hits. Search results whose document embeddings have a negative cosine similarity to the query embedding can be filtered out because they would be, so to speak, contrary to the query. Since even small cosine similarities of angles greater than 60° indicate very dissimilar vectors, it is also useful - in a further embodiment of 303 - to filter the documents in 302 using a minimum similarity threshold. In a further embodiment, instead of query embedding 205, an embedding of user profiles can also be used, which can be constructed analogously to a query 205 or document embedding 107 from a description of the user or his interests. Recommendation phase In a further embodiment, in an optional recommendation phase, any document embedding 107 can also be used instead of query embedding 205 to calculate the cosine similarity and to rank the documents among themselves in order to determine the documents most similar to a document. The applications DE 102019212421 Al, EP 2020073304 describe a method in which term meanings do not have to be specified by a term model, as in conventional search methods, but can be determined directly from the context of the words / character strings within the documents. On the other hand, determining the SimSets 109 on the basis of the specific term meaning not only allows the set of documents to be compared at the time of the query to be efficiently limited, but also provides the user with reasons for the hit determination based on the term similarities precalculated in the SimSets, thus supporting the traceability of the search result. Using the solution proposed here, the method described in applications DE 102019212421 Al, EP 2020073304 can be supplemented by conventional semantic search means based on the use of background knowledge in the form of conceptual models such as taxonomies, thesauri, ontologies, word networks, and / or knowledge graphs. In addition or alternatively, the solution proposed here in the method from applications DE 102019212421 Al, EP 2020073304 allows the language usage of users to be taken into account in the form of a sequence of search queries. The structured text dataset 20 and / or the training text dataset 10 are formed on the basis of a sequence of search queries from a user. When searching in a search engine, users usually have a goal in mind. This means that even if users make several queries, these are not unrelated, but form a broader context. As long as the user's need for information is not satisfied or they give up, they will continue the search process with query variants. These variants can be distinguished as follows: a) Additions to the previous query b) Complete reformulations of the query terms In case a), the previous query is supplemented with new terms, thus supplementing the context of the original query. In case b), the user establishes a connection between the previous query and the reformulated query through the goal they are pursuing. They essentially expand the context of a query with new terms. Therefore, the original query and the reformulated query are indirectly related through the broader context. Although the “pseudo-sentences” derived from these query sequences may be nonsensical in terms of content, they serve the purpose of describing the search term in ever greater detail. In these “pseudo-sentences,” nodes very high up in the hierarchy will occur very frequently, while nodes very low down will occur very rarely. This has direct consequences for the TFIDF of the terms in these “pseudo-sentences” (which is used to determine the importance of the terms in DE 102019212421 Al, EP 2020073304). The terms of the upper nodes receive a lower TFIDF, while the lower ones receive a higher TFIDF, depending on the number of their respective parent nodes in the directed, acyclic graph. This has direct consequences for the use of SimSetslO9, as these can now be expanded with the specific language usage of the users. Contrary to a naive use of word embeddings to implement a semantic search function, the SimSets concept makes it possible to filter the number of search hits - analogous to a purely Boolean exclusion criterion - and thus limit the result set for the user to the “most relevant” documents. Modifications to circumvent the inventions consist of using pre-trained models of word embeddings. General pre-trained models are already available from Google, Facebook, and others, for example. Instead of calculating the word embeddings with Word2Vec, GloVe, or fastText, KNET could be used to modify the invention. Possible applications of the embodiments can be found, for example, in content and document management systems, information systems, and information retrieval systems of libraries and archives. Update In one embodiment of the proposed method, the word embeddings 105 and / or the SimSets 109 are updated, in particular automatically updated or recalculated, depending on at least one update criterion. When the word embeddings 105 are updated, the previously used word-embedding model is adjusted on the basis of newly added tokenized documents and / or on the basis of newly added “pseudo-sentences” in a newly added and / or updated training text dataset 10. Based on the updated word embeddings 105, the SimSets can be updated / recalculated (in particular by first updating the similarity graph). Newly added documents or “pseudo-sentences” or even deleted documents can change the TFIDF of existing words / text elements. This can affect the set of nodes in the similarity graph and the SimSets 109. If new documents are to be added to the document set 101 and these are also to be included in the search, the inverted index for the document set is also updated and the document embeddings for the newly added documents are calculated. An update of the inverted index may also be indicated if documents are to be deleted from the document set 101. An update can, for example, be performed at predetermined, regular intervals - in particular automatically (e.g., daily). In addition or alternatively, an update criterion can be fulfilled if a predetermined number of user sessions by a searcher has been exceeded, thereby making new information on the searcher's language use available. In one embodiment, an update (update of the word embeddings 105, the SimSets 109, and / or the inverted index 103) is only performed if additional factors are also met. In one embodiment, the TFIDF of the words / text elements is calculated based on the changed document set 101 / the changed training text dataset 10. In one embodiment, an update is only performed if the proportion of words / text elements whose TFIDF changes significantly compared to the previous TFIDF value (for example, they slip below / above the TFIDF threshold for selection for the similarity graph) exceeds a threshold value. In one embodiment, an update is only performed if newly added words / text elements are assigned a relatively high TFIDF value. In principle, an update can be performed on a daily basis. If the amount of data is so large that the update cannot be completed during the phase in which the system is not in use, or if the update takes several days, it can be performed in the background (e.g., on a separate computer) and, once completed, the system can be restarted with new index data structures, SimSets, and word embeddings. In principle, an update can be performed “incrementally.” This is particularly true as long as it only involves adding training text data from user query sequences (these are generally not revised). Since vocabularies (such as standards, taxonomies, thesauri, word networks, ontologies, etc.) do not change very often, no “incremental” update is required. Changes to such vocabularies are usually more extensive and may also include deletions. However, deletions create additional work for documenting dependencies. In particular, there is no approach to “undo” the effects on the training of word embeddings. Document corpora change more frequently, and deletions occur here in particular, depending on the area of application (e.g., job advertisements, continuing education offers). For the above reason, recalculation is easier. In general, recalculations can be more efficient than incremental calculations. Reference sign list 10 Training text dataset 11, 11-1 to 11-4 List 20 Structured text dataset 21,21-1,21-2 Hierarchical data structure 30-1 to 30-8 Nodes 30-1 Root nodes 30-2, 30-5, 30-7, 30-8 Leaf nodes 31,32 Nodes to which multiple text elements are assigned 40-1 to 40-7 Text element 50 Edge 101 Documents 102 Indexing method 103 Inverted index 104 Calculation of word embeddings 105 Set of word embeddings 106 Calculation of document embeddings 107 Document embedding 108 Clustering method 108.1 Calculation of and sorting of TFIDF 108.2 Sorting of words in descending order according to TFIDF 108.3 Extraction of similarity graph 108.31 Creation of nodes and edges 108.4 Similarity graph 108.5 Graph clustering 109 SimSets / SimSet group (group of similar character strings / words) 201 Query 5 202 Query expansion 202.1 All documents contained in SimSet group 202.2 Documents containing terms whose similarity to the query is greater than the similarity threshold 202.3 Empty set 10 203 Preselection / document selection 204 Preselected documents 205 Query embedding 301 Doc embedding lookup 302 Selected doc embeddings 15 303 Ranking by cosine similarity 304 Document ranking

Claims

1. Method for training a word-embedding method (105), comprising the following steps- generating at least one training text dataset (10) comprising a multiplicity of lists (11, 11-1 to H-4),wherein the at least one training text dataset (10) is generated on the basis of at least one structured text dataset (20),wherein each of the at least one structured text datasets (20) has at least one hierarchic data structure (21, 21-1, 21-2) based on a directional, acyclic graph on a set of nodes (30-1 to 30-7),wherein each node of the set of nodes (30-1 to 30-7) has one or more associated text elements (40-1 to 40-7), and wherein each text element comprises in particular a word and / or a term,characterized in thateach list of the multiplicity of lists (11, 11-1 to 11-4) is formed on the basis of a longest path of one of the directional, acyclic graphs of one of the at least one structured text datasets (20), wherein the respective list contains an entry for each node of the associated longest path, wherein each entry of the respective list comprises at least one text element associated with the node on which the respective entry is based, and- the at least one training text dataset (10) is used to train a word-embedding method (105) that assigns a respective representation, in particular a real-value vector, to each text element of a subset of the set of text elements that comprises the entries of the multiplicity of lists (11, 11-1 to 11-4).

2. Method according to claim 1, characterized in that at least one of the directed, acyclic graphs of at least one of the at least one structured text datasets (20) comprises at least one node to which several text elements are assigned, and wherein the text elements of the at least one node each represent alternative designations for one another.

3. Method according to claim 1 or 2, characterized in that at least one of the directed, acyclic graphs of at least one of the at least one structured text datasets (20) comprises a node to whichseveral text elements are assigned, and wherein the text elements of the at least one node each represent synonymous terms in different languages.

4. Method according to at least one of the previous claims, characterized in that the plurality of lists of the at least one training text dataset (10) for each of the directed, acyclic graphs of at least one of the at least one structured text dataset (20) for each longest path in the respective directed acyclic graph, comprises exactly one list (11, 11-1 to 11-4).

5. Method according to at least one of the previous claims, characterized in that for each directed acyclic graph of at least one of the at least one structured text dataset (20), for each longest path in the respective directed acyclic graph and for each selection of one text element along the nodes of the respective longest path, the plurality of lists contains exactly one list whose entries correspond to the respectively selected text elements.

6. Method according to at least one of the previous claims, characterized in that at least one of the directed, acyclic graphs of at least one of the at least one structured text datasets (20) has an edge from a first node to a second node, wherein the first node is different from the second node, if- a first text element is assigned to the first node, which represents a generic term of a second text element, which is different from the first text element and is assigned to the second node, and- none of the other text elements of the same directed, acyclic graph represents both a generic term of the second text element and a subordinate term of the first text element.

7. Method according to at least one of the preceding claims, characterized in that at least one of the directed acyclic graphs of at least one of the at least one structured text datasets (20) has an edge from a first node to a second node, wherein the first node is different from the second node, if- the first node is associated with a first text element which represents a superclass of a second text element associated with the second node, and- none of the other text elements of the same directed, acyclic graph represents both a superclass of the second text element and a subclass of the first text element.

8. Method according to at least one of the preceding claims, characterized in that the at least one structured text data record (20) comprises at least- a thesaurus,- a word network, and / or- an ontology.

9. Method according to at least one of the preceding claims, characterized in that at least one of the at least one structured text dataset (20) comprises at least one hierarchical data structure (21, 21-1, 21-2), the underlying directed, acyclic graph of which has a tree structure.

10. Method according to at least one of the preceding claims, characterized in that at least one of the at least one structured text data record (20) is created by modeling a thesaurus and / or an ontology from search queries of a user within one or more user sessions.

11. Method according to at least one of the preceding claims, characterized in that at least one of the at least one structured text data record (20) is created by combining a user’s search queries made within one or more user sessions into a sequence, from which optionally directly consecutive duplicates are removed, and wherein the hierarchical data structure of the at least one structured text data record (20) is determined by the temporal sequence of the search queries made.

12. Method according to at least one of the preceding claims, characterized in that the order of the entries in each of the plurality of lists (11, 11-1 to 11-4) is predetermined by the order of the nodes in the longest path, on the basis of which the respective list was formed.

13. Method according to claim 12, characterized in that the order of the entries in each of the plurality of lists (11, 11-1 to 11-4) runs along the order of the nodes of the associated longest path, wherein the first entry of the respective list (11, 11-1 to 11-4) corresponds to a text element of a root node of the set of nodes and the last entry of the respective list (11, 11-1 to 11-4) corresponds to a text element of a leaf node of the set of nodes.

14. Method according to claim 12, characterized in that the order of the entries in each of the plurality of lists (11, 11-1 to 11-4) runs in the opposite direction to the order of the nodes of the associated longest path, wherein the first entry of the respective list (11, 11-1 to 11-4) corresponds 46to a text element of a leaf node of the set of nodes and the last entry of the respective list (11, 11-1 to 11-4) corresponds to a text element of a root node of the set of nodes.

15. Method according to at least one of the preceding claims, characterized in that the wordembedding method (105) comprises at least one of the methods Word2Vec, GloVe, FastText, Gauss2Vec, or Bayesian Skip-gram.

16. Method according to at least one of the preceding claims, characterized in that a CBOW model or a Skip-gram model is used as the word-embedding method (105).

17. Method according to at least one of the preceding claims, characterized in thata) the at least one training text dataset (10) is used in addition to a subset of a set of documents (101) to train the word-embedding method (105), wherein the documents (101) comprise tokenized character strings,b) an inverted index for the at least one subset of the documents (101) is calculated using an indexing method (102),c) for the at least one subset of the documents (101), a document embedding (107) is calculated for each of these documents (101) by adding the word embeddings (105) of all character strings, in particular words of the document (101), for each document (101) and normalized with the number of character strings, in particular words (106), wherein before, after or in paralleld) SimSet groups (109) of similar character strings are calculated with the calculated word embeddings (105) using a clustering method, andsubsequentlye) in a query phase (200), a query expansion (202) is performed, in whichi) query terms that occur in SimSet groups (109), orii) query terms that do not occur in the SimSet groups (109) but do occur in the documents (101), oriii) query terms that do not occur in the documents (101), in particular also misspelled query termsare used for a preselection (203) of the documents in order to achieve a restriction of the hit set, and then a query embedding (205) is first determinedand thenf) a comparison of the query embeddings (205) with the document embeddings (107) of the SimSet groups (109) formed in step d) using the clustering method to quantitatively limit the number of document embeddings (107), preselected documents is performed in order to automatically determine a ranking of the similarity of the documents (101) and to display and / or store documents and / or metadata linked to the documents, in particular bibliographic data.

18. Apparatus comprising- a means for generating at least one training text dataset (10), which comprises a plurality of lists (11, 11-1 to 11-4),wherein the generation of the at least one training text dataset (10) is performed on the basis of at least one structured text dataset (20),wherein the at least one structured text dataset (20) comprises at least one hierarchical data structure (21, 21-1, 21-2) based on a directed, acyclic graph on a set of nodes (30-1 to 30-7),wherein each node of the set of nodes (30-1 to 30-7) is assigned one or more text elements (40-1 to 40-7), wherein each text element comprises in particular a word and / or a term, andwherein each list of the plurality of lists (11, 11-1 to 11-4) is formed on the basis of a longest path of one of the directed, acyclic graphs, wherein the respective list contains an entry for each node of the associated longest path, wherein each entry of the respective list comprises at least one text element which is assigned to the node on which the respective entry is based,and- a means for training a word-embedding method (105) based on the at least one training text dataset (10), which assigns a representation, in particular a real-valued vector, to each text element of a subset of the set of text elements comprising the entries of the plurality of lists (11, 11 -1 to 11 -4).INTERNATIONAL SEARCH REPORT International application No. PCT / EP2024 / 067379A. CLASSIFICATION OF SUBJECT MATTER G06F 16 / 36(2019.01)1 According to International Patent Classification (IPC) or to both national classification and IPC B. FIELDS SEARCHED Minimum documentation searched (classification system followed by classification symbols) G06F Documentation searched other than minimum documentation to the extent that such documents are included in the fields searched Electronic data base consulted during the international search (name of data base and, where practicable, search terms used) EPO-Intemal, WPI Data C. DOCUMENTS CONSIDERED TO BE RELEVANT Category* Citation of document, with indication, where appropriate, of the relevant passages Relevant to claim No. X US 2022292123 Al (HOPPE THOMAS [DE]) 15 September 2022 (2022-09-15) 1-18 paragraphs [0042], [0092], [0111] | | Further documents are listed in the continuation of Box C. | V | See patent family annex. * Special categories of cited documents: “A” document defining the general state of the art which is not considered to be of particular relevance “E” earlier application or patent but published on or after the international filing date “L” document which may throw doubts on priority claim(s) or which is cited to establish the publication date of another citation or other special reason (as specified) “O” document referring to an oral disclosure, use, exhibition or other means “P” document published prior to the international filing date but later than the priority date claimed “T” later document published after the international filing date or priority date and not in conflict with the application but cited to understand the principle or theory underlying the invention “X” document of particular relevance; the claimed invention cannot be considered novel or cannot be considered to involve an inventive step when the document is taken alone “Y” document of particular relevance; the claimed invention cannot be considered to involve an inventive step when the document is combined with one or more other such documents, such combination being obvious to a person skilled in the ait document member of the same patent family Date of the actual completion of the international search Date of mailing of the international search report 09 August 2024 23 August 2024 Name and mailing address of the ISA / EP Authorized officer European Patent Office p.b. 5818, Patentlaan 2,2280 HV Rijswijk Netherlands (Kingdom of the) Telephone No. (+31-70)340-2040 Facsimile No. (+31-70)340-3016 Rameseder, Jonathan Telephone No.INTERNATIONAL SEARCH REPORT Information on patent family membersInternational application No.PCT / EP2024 / 067379Patent document cited in search report Publication date (day / month / year) Patent family member)s) Publication date (day / month / year) US 2022292123 Al 15 September 2022 CA 3151834 Al 25 February 2021 DE 102019212421 Al 25 February 2021 EP 3973412 Al 30 March 2022 US 2022292123 Al 15 September 2022 WO 2021032824 Al 25 February 2021

Images