WO2020201249A1 - Apprentissage par machine sur la base de définitions de déclencheurs - Google Patents

Apprentissage par machine sur la base de définitions de déclencheurs Download PDF

Info

Publication number
WO2020201249A1
WO2020201249A1 PCT/EP2020/059042 EP2020059042W WO2020201249A1 WO 2020201249 A1 WO2020201249 A1 WO 2020201249A1 EP 2020059042 W EP2020059042 W EP 2020059042W WO 2020201249 A1 WO2020201249 A1 WO 2020201249A1
Authority
WO
WIPO (PCT)
Prior art keywords
trigger
data
tokens
additional
index
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/EP2020/059042
Other languages
German (de)
English (en)
Inventor
Andreas Wilke
Ilya Komarov
Peter Palm
Manfred Paeschke
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Bundesdruckerei GmbH
Original Assignee
Bundesdruckerei GmbH
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Bundesdruckerei GmbH filed Critical Bundesdruckerei GmbH
Priority to EP20719125.5A priority Critical patent/EP3948578A1/fr
Publication of WO2020201249A1 publication Critical patent/WO2020201249A1/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/31—Indexing; Data structures therefor; Storage structures
    • G06F16/313—Selection or weighting of terms for indexing
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
    • G06F16/31—Indexing; Data structures therefor; Storage structures
    • G06F16/316—Indexing structures
    • G06F16/328—Management therefor
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90—Details of database functions independent of the retrieved data types
    • G06F16/906—Clustering; Classification
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning

Definitions

  • the invention relates to a method and a computer system for machine learning.
  • the invention is based on the object of creating an improved method for maschi nellen learning.
  • Embodiments include a computer-implemented method for machine learning, the method comprising:
  • the pre-trained learning module comprising a plurality of predetermined trigger definitions which define triggers for assigning tokens to classes of a first group of classes,
  • the database comprising a plurality of data records which are stored in a document-oriented data model, the stored data records each comprising one or more field values, the individual field values of the stored data records are each stored in a field,
  • the database further comprises a searchable index which is stored in a further data model, wherein the index comprises a plurality of tokens generated from the field values of the stored data records, wherein tokens in the index each point to one or more pointers one or more of the data records stored in the document-oriented data model are linked, from whose field values the corresponding token was generated,
  • first tokens in the index which are included in one of the trigger definitions as triggers, are each assigned to the corresponding trigger definition, with second tokens in the index each being assigned to one or more classes of the first group of classes, and the remaining ones To ken in the index for identifying the corresponding remaining tokens are assigned to a collecting class as unknown data, the assignment to the collecting class excluding an assignment to one of the trigger definitions as well as an assignment to one of the classes of the first group of classes ,
  • the learning module identifies the corresponding token as a trigger
  • the learning module Using the identified triggers for assigning one or more second additional tokens to one or more classes of the first group of classes by the learning module, if the corresponding second additional tokens from the additional data set in a combination with one or more of the identified triggers according to one of the trigger definitions are included, the corresponding triggers triggering a corresponding class assignment according to the corresponding trigger definition,
  • Embodiments can have the advantage that the learning module is a pre-trained learning module.
  • the pre-trained learning module comprises a plurality of inertially provided or defined trigger definitions.
  • the learning module is configured to use these inertially defined trigger definitions to classify all tokens included in the database or the index.
  • Embodiments can have the advantage that no coincidence enters the decision-making or classification process. Rather, the classification of tokens is based on the predetermined trigger definitions and can therefore be traced at any time. Even if the learning module advances, for example on the basis of the classification, and further patterns and regularities on the basis of which it learns additional patterns and regularities in the course of a learning transfer, the underlying classification goes back to the predetermined trigger definitions.
  • meta and / or context information on the classified tokens is provided in the form of the classification.
  • This meta and / or context information is identified on the basis of the trigger according to the trigger definitions and assigned to the corresponding token in the form of the class assignment.
  • the learning module can be configured to learn further patterns and laws using this meta and / or context information.
  • Embodiments can have the advantage that the data records received from the database are all stored in their original form in the document-oriented data model. This ensures that the full information content of these data records is retained.
  • those of the are stored in the document-oriented data model
  • Data sets comprised data provided in the form of an index.
  • This index includes the corresponding data of the document-oriented data model in the form of tokens.
  • the index comprises all of the elementary data elements included in the document-oriented data model in the form of elementary tokens.
  • the index additionally comprises a combination of the elementary data elements in the form of token combinations comprised by the document-oriented data model. These token combinations each include a combination of a plurality of elementary tokens.
  • the index comprises token combinations up to a predetermined complexity.
  • the complexity of a token combination is defined, for example, by the number and / or type of elementary tokens it encompasses.
  • the index comprises all combinations of elementary data elements included in the document-oriented data model in the form of corresponding token combinations.
  • the tokens included in the index can be triggers according to the predetermined trigger definitions, for example. If a corresponding token is generated for the first time, for example in the course of processing an additional data record, it is identified as a trigger using one of the trigger definitions, added to the index and assigned to the corresponding trigger definition. If the learning module recognizes the same token, which the index defines as a trigger, within a further data set, the learning module accesses the trigger definition assigned to triggering in the index and assigns one or more tokens following the corresponding trigger definition to a context of the token stored as a trigger in the further data record one or more classes of the first group of classes.
  • the index further comprises a plurality of tokens which are each assigned to one or more classes of the first group of classes.
  • the assignment to the classes provides meta and / or context information for the corresponding tokens.
  • the corresponding meta and / or context information can be used, for example, for processing the corresponding tokens and / or the data records comprising the corresponding tokens are used in the document-oriented data model.
  • the corresponding meta and / or context information is used in the course of a search query to identify relevant tokens and / or data records or in the course of a further method for machine learning that uses the index.
  • additional patterns and laws can be learned in the course of a further learning transfer.
  • This further method for machine learning is carried out, for example, by the learning module using the trigger definitions or by a further learning module.
  • the further method for machine learning is a KI method that is executed by a KI module.
  • the index includes tokens that do not fall under any of the predetermined trigger definitions. These tokens are neither triggers nor can they be assigned to the trigger classes defined by the trigger definitions. Rather, these tokens are unknown data that cannot be assigned and for which meta or context information is missing. These tokens are assigned to a trap class as unknown data. An assignment to the collection class excludes an assignment to one of the trigger definitions as well as an assignment to one of the classes of the first groups of classes.
  • Embodiments can have the advantage that the token assignments can be used to identify which tokens are unknown data and which tokens are known data, ie trigger or classifiable data.
  • search queries can be defined in such a way that they only take into account known data.
  • Additional learning algorithms can be configured, for example, in such a way that they work exclusively on known data.
  • the use of chance in a decision-making or classification process can be used, even if additional learning algorithms are used.
  • the initially defined triggers are available for all learning processes and / or KI processes, which are used to classify the data received from the database.
  • the predetermined trigger definitions provide a basis for supervised learning.
  • Embodiments can furthermore have the advantage that additional data records that are added to the database are each analyzed to determine which of the data they comprise is known data and which data is unknown data.
  • known data is understood to mean those data which are known as triggers, for which meta or context information is available and / or for which meta or context information is derived from the context of the data records using the trigger definitions can.
  • Data that are neither triggers nor data that can be classified using the trigger definitions are unknown data. Unknown data is assigned to the trap class.
  • Embodiments can have the advantage that a database system optimized for machine learning is used.
  • the corresponding database system includes all the data on which machine learning is based, i.e. both trigger definitions used to classify data and the data that are processed using the trigger definitions. This enables continuous learning by the learning module, taking into account all of the data seen by the system or the learning module.
  • the database saves all data records received in a document-oriented data model.
  • a document-oriented data model means that the data model does not impose any structural requirements on the data to be saved. Rather, the data is stored in documents or data containers in the form in which it is received. In this sense it concerns with the document-oriented data model, stored data to raw data.
  • Raw data means that the data are stored in the form in which they are received, without additional data processing by the database management system, in particular no restructuring of the data.
  • Embodiments can have the advantage that the entire information content of the received data can thus be (almost) completely retained without the assumptions of the database management system being included. Both the database management system and the learning module can access the original data at any time and take them into account in further processing.
  • An index is generated based on this data pool of raw data provided by the document-based data model. Structural information or contexts of meaning are only extracted from the data sets at this level. This structural information or context is taken into account in the form of class assignments of the indexed data.
  • the data records are broken down to an elementary level using tokenization, which takes into account the elementary components of the data records in the form of tokens.
  • the tokens are assigned by the learning module as a trigger to one of the trigger definitions or classified as using the trigger definitions. All tokens that are neither identified as triggers nor can be classified using one of the trigger definitions are assigned to the trap class as unknown data.
  • the learning module comprises a classifier and is configured to classify the tokens using the predetermined trigger definitions.
  • the corresponding classification can, for example, be part of a pattern recognition in which a feature extraction is implemented by the tokenization.
  • each token in the index is linked to one or more pointers that indicate in which data records the corresponding token occurs. This means that the raw data relevant to a token can be accessed at any time and this raw data can be used for evaluation with regard to this token.
  • the token assignments are differentiated according to known data, which represent secured facts, and unknown data.
  • Embodiments can have the advantage that the use of the structures and regularities determined by the learning module in the data records, which are reflected in the token assignments, is based on the use of the predetermined trigger definitions.
  • Unknown data are recorded as such and left out until they can also be classified and thus viewed as reliable facts.
  • Such an additional classification can be implemented, for example, by additional trigger definitions.
  • additional trigger definitions can be added to reduce the amount of tokens included in the collection class. The method thus enables learning and / or classification with reservations.
  • Embodiments can therefore have the advantage that they allow the learning module to work on the entire available database. In particular, they can have the advantage of enabling continuous learning that takes into account both additional data sets and data sets that have already been saved. Embodiments can therefore have the advantage that they are not restricted to an arbitrary subset being picked out of an available total amount of data in order to train on it. Rather, all of the data contained in the database are processed using the trigger definitions. By adding to the trigger definitions, it can also be achieved, according to embodiments, that all tokens are either identified as triggers or classified using the (added) trigger definitions. Are unknown data from searches and / or Excluded from further learning processes, this exclusion is not arbitrary, but based on the trigger definitions provided.
  • Embodiments can have the advantage that no random initialization is required, as is the case with known self-learning systems, e.g. neural networks. Rather, the initialization is based on the predetermined trigger definitions. Because of the random moment resulting from this random initialization, the decisions / classifications of a corresponding neural network are not transparent and cannot be traced. In contrast, embodiments can have the advantage of being completely deterministic.
  • Embodiments can have the advantage that an already trained system, e.g. the pre-trained learning module is retrained or trained further.
  • Trigger definitions can be added, removed or changed. In this way, for example, the classes used in the classification can also be added, removed or changed. If trigger definitions are added, removed or changed, then all assignments of tokens based on these to the corresponding trigger definitions or to one of the classes must be adapted accordingly.
  • trigger definitions for example, new structures can be recorded that have not yet been shown. In this way, the factual knowledge in particular can be expanded subsequently, i.e. Tokens assigned to the trap class can be assigned to other classes.
  • the learning module implements an algorithm for machine learning, the method not being restricted to a specific algorithm.
  • the machine learning algorithm comprises at least one classification algorithm for classifying tokens.
  • Machine learning can be monitored or unsupervised learning.
  • the machine learning can include a classification and / or a regression analysis.
  • a learning algorithm tries to find a hypothesis or a mapping which assigns the (presumed) output value to each input value. Are the output values to be assigned in a continuous distribution, the results of which can assume any quantitative values within a given range of values, is generally referred to as a regression problem. If, on the other hand, the output values to be assigned are available in discrete form or if the values are qualitative, this is generally referred to as a classification problem.
  • the machine learning relies on the classification of the indexed tokens.
  • the learning module comprises an algorithm specially developed for machine learning, such as, for example, but not limited to, a density-based multidimensional outlier detection (“local outlier detection”), a Random Forrest algorithm, a neural network, a support vector machine, a naive Bayes classifier or a feedback similar to the feedback of a linear or nonlinear controller.
  • a multi-model database is understood here to be a database which is configured to support a plurality of different data models.
  • a multi-model database is therefore configured to store, index and query data in more than one data model.
  • Data models are, for example, relational, column-oriented, document-oriented, graph-based, key-value-based etc.
  • a database model defines the structure in which data is stored in a database system, i.e. the form in which the data is organized, stored and processed.
  • a database is understood to be a (typically large) amount of data that is managed in a computer system by a database management system (DBMS) according to certain criteria.
  • the data is organized in a large number of data sets.
  • a database management system or DBMS is understood below to mean an electronic system for storing and retrieving data.
  • the data are preferably stored consistently and permanently in the DBMS and are efficiently made available to various application programs and users in a needs-based form.
  • a DBMS can typically contain one or more databases and manage the data records contained therein. With the DBMS it can be preferably be a field-oriented DBMS, that is to say a DBMS that is configured to store parts of individual data records, so-called field values, in several different fields.
  • a data record is understood to mean a coherent set of data made available to the database system, which is managed by the database management system as a coherent set of data.
  • a data record comprises, for example, a set of content-related data.
  • data sets are stored in the document-oriented data model as coherent data sets.
  • a single data set may refer to a particular physical object, e.g. a natural person or a device. The person can e.g. be an employee, a patient, a customer, etc.
  • the device can be, for example, a production device, a computer device, a computer or network element or a transport device.
  • the corresponding data record can contain a predefined set of attribute values of this person or device (e.g. name or pseudonym, age, height, weight, date of birth, ID numbers, security certificates, authentication codes, biometric data, identifier, date of entry, date of commissioning, configuration data, and so on other).
  • a data record can represent a group of content-related data fields (belonging to an object), e.g. B. Item number, item size, item color, item name or the like.
  • the classes' Name ‘, 'Address' and' Date of birth 'could e.g. form the logical structure of a data record for the "person" object type.
  • data is stored in the form of data records in databases, whereby they are the subject of the processing of computer programs and are generated, read, changed and deleted by these.
  • NoSQL (English for Not only SQL) DBMS is a DBMS that follows a non-relational approach to data storage and does not require any fixed table schemes.
  • the NoSQL DBMSs include in particular document-oriented DBMSs such as Apache Jackrabbit, BaseX, CouchDB, IBM No tes, MongoDB, graph databases such as Neo4j, OrientDB, InfoGrid, HyperGraphDB, Core Data, DEX, AllegroGraph, and 4store, distributed ACID DBMSs such as MySQL Cluster, key value databases such as Chordless, Google BigTable, GT.M, InterSystems Cache, Membase, Redis, sorted key-value memories, multivalue databases, object databases such as Db4o, ZODB, column-oriented databases and temporary databases such as Codex DB.
  • document-oriented DBMSs such as Apache Jackrabbit, BaseX, CouchDB, IBM No tes, MongoDB, graph databases such as Neo4j, OrientDB, InfoGrid, HyperGraphDB, Core Data, DE
  • An index is a data structure which accelerates a search for certain data values by a database management system.
  • An index consists of a collection of pointers (references) that define an order relation to several “indexed” data values (stored in the index). For example, B + trees are used for this.
  • Each indexed data value is linked to further pointers which refer to data records in which the indexed data value found is contained and which represented the database for creating the index.
  • Database management systems use indices to quickly identify the desired data records in response to a search query, in that the index is first searched along the pointer for a data value which is identical to a reference value contained in the search query. Without an index, the data values of a field managed by the DBMS would have to be searched sequentially, while a search using the index, e.g. of a B + tree, often only has logarithmic complexity.
  • the index also orders the indexed data, i. Token, classes, whereby the corresponding data is linked with meta or context information.
  • This meta or context information can be used in a search and / or in a machine learning process on the data in the database.
  • a field is an area on a logical or physical data carrier that is managed by a DBMS that is assigned to a previously defined field type and that is created and intended for storing a field value of a data record. So a field is an element for Storage of a field value of a data record according to the above definition. Fields of a data record are managed jointly by a DBMS.
  • a field value is a data value that is part of a data record and is stored in a field of the data record.
  • a field value can consist of a single word, a single number, or a combination of several words and / or numbers and / or other data formats, with different embodiments of the invention varying degrees of flexibility with regard to the type and combinability of data types within the the same field value.
  • a "tokenizer” is a program logic that receives data, for example a field value, as input, analyzes the data, e.g. to identify delimiters or other decomposition criteria and patterns, and then breaks the data down into one or more tokens as the result of the analysis and returns the tokens. It is also possible that not all data will be returned as tokens. For example, a full-text indexer can recognize and filter out semantically insignificant stop words so that they are not indexed. Alternatively, all data is returned as.
  • To “tokenize” a data value means breaking the data value into several components according to a certain scheme. The components represent the tokens. For example, natural-language texts can be divided up using predefined separators, e.g.
  • tokens spaces, periods or commas, and the components (words) generated in this way are used as tokens.
  • all tokens are used for indexing. It is also possible that some tokens are not used for indexing (e.g. stop words) or that the tokens are additionally processed prior to indexing (e.g. reducing words to the stem).
  • the search value is preferably processed in the same way by the client computer system or the server computer system to ensure that the search values of the search queries match the tokens contained in the index correspond.
  • a class defines a category or a type to which a token belongs. The class therefore assigns meta or context information to the token, for example in the form of a property.
  • a class can represent a certain attribute of a physical object in the form of a token.
  • data records to be saved that contain employee attributes, which classes such as "Name”, “Pseudonym”, “ID number”, “Access certificate for room R”, “Access certificate for device G”, “Access certificate for building GB", " Age "represent.
  • Each token can be assigned to one or more classes.
  • combinations of tokens can in turn be assigned to one or more further classes as independent tokens.
  • the data records received are stored using a document-oriented data model.
  • all field values of the stored data records are transferred as tokens to a multi-dimensional key / value memory (KeyA / alue store) or key value databases.
  • KeyA / alue store multi-dimensional key / value memory
  • key value databases key value databases
  • the transaction time and the validity time of the data records are also stored bit-temporally.
  • the transaction time indicates the point in time at which a change to a data object in the database occurs.
  • the validity time indicates a point in time or period in which a data object in the modeled image of the real world has the state described. If both validity and transaction time are relevant, one speaks of bitemporal.
  • a key-value data model enables storage, retrieval and management of associative data fields. Values are uniquely identified using a key.
  • a document-oriented data model also known as a document store
  • documents or data containers form the basic unit for storing the data.
  • a document-oriented data model enables document-oriented information, also known as semi-structured data, to be stored, accessed and managed.
  • Databases based on a document-oriented data model belong to the NoSQL databases and form a subclass of key-value stores.
  • key-value store the data is considered to be inherently opaque to the database, while a document-oriented database relies on internal structures in the stored documents to extract metadata.
  • the semi-structured data model is a database model in which there is no separation between the data and the schema and the scope of the structure used depends on the purpose of the database. Each document within the data model is addressed using a unique identifier.
  • a combination of the different database concepts enables data records to be saved as documents or containers (document disturbance) and additionally in the form of an index, e.g. of a key-value memory to be converted into the 6th normal form.
  • This key-value memory represents the entire amount of data in the document memory, while the original data records are retained.
  • selections are carried out exclusively in the key-value memory in the redundancy-free sixth normal form. Only the result is read from the document storage container. According to embodiments, in addition to reading and writing rights in the data records, a selection right is also implemented on the key-value memory. This means that you can work on the index alone without having to read out the underlying data.
  • the proposed multi-model database thus provides a complete normalization of the entire scope of data in the sixth normal form in addition to a schematic data storage based on a document memory.
  • Execution forms can have the advantage that the index includes data elements of the data records, ie tokens, as keys and each of these keys has one or several pointers are assigned as values which indicate in which data sets and / or fields of the data sets the corresponding key, ie token / data value, is stored as a field value.
  • This index therefore forms all fields of the data records and their contents, i.e. the field values from the entire database with all of the data records it encompasses, so that all queries are handled in the index and the data of the document-oriented data model stored without a schema are only used to output the search results.
  • the small size of the index compared to the schema-less data enables quick queries in any query combination.
  • a computer or computer system is understood here to mean a device which processes data by means of programmable arithmetic rules.
  • a program or program instructions is understood here, without restriction, to be any type of computer program which includes machine-readable instructions for controlling a functionality of a computer.
  • a computer or computer system can comprise a communication interface for connection to the network, wherein the network can be a private or public network, in particular the Internet or another communication network. Depending on the embodiment, this connection can also be established via a cellular network.
  • a computer system can be a stationary computer system, such as a personal computer (PC) or a client or server integrated in a client-server environment.
  • a computer system can be, for example, a mobile telecommunications device, in particular a smartphone, a portable computer such as a laptop PC or palmtop PC, a tablet PC, a personal digital assistant or the like.
  • a memory is understood here to mean both volatile and non-volatile electronic memories or digital storage media.
  • a non-volatile memory is understood here as an electronic memory for the permanent storage of data.
  • a non-volatile memory can be configured as a non-changeable memory, which is also referred to as read-only memory (ROM), or as a changeable memory, which is also referred to as non-volatile memory (NVM).
  • ROM read-only memory
  • NVM non-volatile memory
  • it can be an EEPROM, for example a Flash EEPROM, referred to as Flash for short.
  • Flash Flash for short.
  • a non-volatile memory is characterized by the fact that the data stored on it is retained even after the power supply has been switched off.
  • a volatile electronic memory is a memory for the temporary storage of data, which is characterized in that all data is lost after the power supply is switched off.
  • this can be a volatile random access memory, which is also referred to as a random access memory (RAM), or a volatile main memory of the processor.
  • RAM random access memory
  • a processor is understood here and in the following to be a logic circuit that is used to execute program instructions.
  • the logic circuit can be implemented on one or more discrete components, in particular on a chip.
  • a processor is understood to mean a microprocessor or a microprocessor system made up of a plurality of processor cores and / or a plurality of microprocessors.
  • supplementing the index includes:
  • Embodiments can have the advantage that data from additional data records can be efficiently inserted into the existing database and, in particular, into the index.
  • the tokens generated using the additional data set are compared with the index. All tokens that the index does not (yet) include are added to the index as additional tokens including their class assignments. Furthermore, the additional tokens are each linked to the pointer to the additional data record.
  • the corresponding class assignments are added.
  • the pointer to the additional data record is added to the index for these tokens.
  • Embodiments can have the advantage that it can always be ensured that the index has all tokens comprised by the data records in the database.
  • the index for all corresponding tokens includes all found class assignments.
  • each of the tokens in the index is linked to a pointer to all data records in the database that contain the corresponding token.
  • an initial set of predetermined trigger definitions is established.
  • data records are received and stored in the document-based data model.
  • the stored data records are tokenized and class assignments are determined for the resulting tokens using the initially set trigger definitions and an initial index is generated for the resulting tokens.
  • the initial index includes all triggers included in the trigger definitions as tokens.
  • tokens specified as triggers by the trigger definitions are only added to the index on the condition that they are included in one of the data records.
  • An assignment of a token to a class using a predetermined trigger function represents a fact secured by the corresponding predetermined trigger function. For tokens which are not triggers and which are not covered by any of the trigger definitions, there is a lack of such factual knowledge. Instead, the corresponding tokens are assigned to the trap class as unknown data.
  • Embodiments can thus have the advantage that, using initially established trigger definitions, new data can be converted into known data, i.e. Triggers or tokens classified using trigger definitions, and unknown data can be classified, i.e. Token assigned to the trap class.
  • the combinations of second additional tokens with one or more of the identified triggers that have triggered a class assignment according to one of the trigger definitions are identified in the index as classified combinations and class assignments are only made for combinations of second additional tokens and one or more identified triggers that are not marked as classified combinations.
  • Embodiments can have the advantage that for all token combinations for which a class assignment has already been taken into account or for which a class assignment has already been carried out, are each identified in the index as already classified. It can thus be avoided that the same classifications are carried out again for token combinations which the learning module has already seen and fully taken into account in the course of the classifications.
  • the system can thus be designed to be significantly more efficient.
  • the index includes all token combinations for which a classification has already been made, i.e. all token combinations which are to be marked as classified.
  • the corresponding token combinations in the index are each provided with a flag which indicates whether the corresponding token combinations are classified token combinations.
  • a comparison is first carried out with all token combinations marked as already classified. The classification is not repeated for these token combinations; rather, there is only a link with the pointer to the additional data record.
  • the corresponding pointer is also linked to all of the tokens included in the token combination in the index.
  • the comparison first takes place with the largest, ie most extensive, token combinations of the index. For all token combinations of the additional data record already recognized as classified, only the pointer to the corresponding data record is stored in the database. According to embodiments, the corresponding pointer is also linked to all of the tokens included in the token combination in the index.
  • a comparison with further token combinations takes place successively, whereby the size or scope of the further tokens used Combinations successively decreases. According to embodiments, only those further token combinations with a smaller size or scope are taken into account which, as part of a larger or extensive token combination, a match was not found in the course of the comparison.
  • Embodiments can have the advantage that for extensive token combinations which are recognized as already classified, no additional comparison is made for sub-combinations comprised by the corresponding token combination. Rather, a corresponding comparison only takes place if the corresponding sub-combination is included in the additional data record as an independent token combination independent of the corresponding more extensive token combination.
  • the method further comprises:
  • Embodiments can have the advantage that, based on the triggers identified by the initial trigger definitions, additional triggers can be identified in the form of trigger combinations. Based on these identified trigger combinations, combined trigger definitions can be determined from the initial trigger definitions, with which the majority of the predetermined trigger definitions can be expanded.
  • Embodiments can have the advantage that on the basis of combined trigger definitions token combinations can be identified in the index as classified combinations, thereby avoiding unnecessary repetitions of classifications of already classified token combinations.
  • the combination criterion includes a minimum frequency for the corresponding trigger combination to occur in the data records.
  • Embodiments can have the advantage that corresponding trigger combinations are only used to form a combined trigger definition when the corresponding trigger combination occurs in the data records with a minimum frequency. This prevents additional combined trigger definitions from being formed due to the accidental occurrence of triggers with different trigger definitions in one and the same data set. Such a random occurrence is to be expected from a certain size and / or complexity of the data records, without it being possible to draw conclusions about an underlying relationship between the triggers. However, if the corresponding trigger combinations occur more frequently, a connection can be concluded from them.
  • the minimum frequency defines an absolute frequency value of the occurrence in the data records.
  • the corresponding minimum frequency can be a minimum value for the occurrence of the corresponding trigger combination in all data records. The occurrence of the corresponding trigger combination is added up across all data records. If the resulting sum is greater than or equal to the minimum value, this is fulfilled.
  • the minimum frequency can be a minimum value for the occurrence in one of the data records. The occurrence of the corresponding trigger combination is summed up individually for the individual data records. If one of the resulting sums meets the minimum value, the minimum frequency is present.
  • the minimum value must be a predetermined value Number of records or a predetermined percentage of the records to be filled.
  • the corresponding predetermined percentage is either a percentage of all data records in the database or all data records which comprise the corresponding trigger combination.
  • the minimum value must be fulfilled by all data records and / or by all data records which comprise the corresponding trigger combination.
  • the corresponding minimum frequency can be a minimum value for an average frequency of occurrence of the corresponding trigger combination in all data records of the database or all data records which comprise the corresponding trigger combination.
  • the minimum frequency defines a relative frequency value of the occurrence in the data records.
  • the corresponding minimum frequency is dependent on the number of data records and / or the number of tokens and / or the size of the data comprised by the data records. For example, the frequency value determined by the minimum frequency increases with the number of data records and / or the number of tokens and / or the size of the data comprised by the data records.
  • the minimum frequency stipulates a relative frequency value of the occurrence in the data records relative to the frequencies of occurrence of one or more of the triggers comprised by the corresponding trigger combination in the data records.
  • the relative frequency value is dependent on the occurrence of the trigger with the highest frequency of occurrence, the trigger with the lowest frequency of occurrence and / or an average value of the occurrence of all triggers of the corresponding trigger combination.
  • Embodiments can have the advantage that, when a relative frequency value is taken into account, the frequency of occurrence of one or more of the triggers comprised by the corresponding trigger combination is included in the decision-making process as to whether an additional combined trigger definition based on the corresponding trigger combination is complementary, with flowing in.
  • the frequency of occurrence of the corresponding triggers can vary as before in the case of the absolute frequency value to an occurrence of the corresponding trigger in all data records, to an average occurrence in all data records, to a most frequent occurrence in one of the data records and / or to a minimum occurrence in one of the data records.
  • Embodiments can have the advantage that the relative frequency value is selected to be higher, the higher the frequencies of occurrence of the one or more corresponding triggers comprised by the trigger combination. It can thus be avoided that a trigger definition is generated on the basis of a trigger combination, the occurrence of which is random, i.e. whose triggers happen to be included in the same data set, without this indicating a connection between the corresponding triggers.
  • the combination criterion comprises one or more conditions at relative positions of the triggers of the corresponding trigger combination to one another within one of the data sets.
  • Embodiments can have the advantage that a relative position of the triggers of the corresponding trigger combination within the data set is taken into account for the combination criterion.
  • a corresponding relative position of data within data records results from or is dependent on contextual relationships. Corresponding contextual relationships can therefore be read from the relative position.
  • the relative position can be a relative position in a one-dimensional, i.e. sequential data structure, such as a text or voice file, a two-dimensional data structure, such as an image file, or a higher-dimensional, for example three-dimensional or n-dimensional, data structure.
  • the trigger definitions each include a definition of a trigger structure which is used for one or more triggers included in the corresponding trigger definition and one or more triggers in accordance with the corresponding Trigger definition of one of the tokens to be assigned to the classes defines relative positions to each other.
  • Embodiments can have the advantage that a corresponding trigger definition uses one or more triggers to define how one or more tokens are to be classified as a function of a relative position of the corresponding tokens to the corresponding triggers.
  • the corresponding relative position can be a relative position in a one-dimensional, two-dimensional or higher-dimensional, for example three-dimensional or n-dimensional, data space.
  • the definitions of the relative positions include at least one of the following definitions: the one or more tokens to be assigned are arranged according to a trigger included by the corresponding trigger definition, the one or more tokens to be assigned are before one of the corresponding trigger definition arranged triggers included, the one or more tokens to be assigned are each arranged between triggers included in the corresponding trigger definition.
  • a trigger can, for example, trigger a classification of preceding data, e.g. "[before] [trigger]”.
  • the occurrence of the trigger “Triggerl” triggers a classification of the preceding data “before”.
  • the trigger itself is part of the classification, i.e. the combination "[before] [trigger]” is classified.
  • the trigger “Triggerl”, if it is recognized, is assigned as a trigger to the corresponding trigger definition.
  • a trigger can, for example, trigger a classification of subsequent data, eg "[Trigger2] [afterl]".
  • the occurrence of the trigger “Trigger2” triggers a classification of the subsequent data “after”.
  • the trigger itself is part of the classification, ie it is classified Combination “[Trigger2] [after]”.
  • the trigger “Trigger2”, if it is recognized, is assigned to the corresponding trigger definition as a trigger.
  • a trigger can, for example, trigger a classification of preceding and succeeding data, e.g. "[before2] [trigger3] [after2]”.
  • the occurrence of the “Trigger3” trigger triggers a classification of the preceding data “before2” and the following data “after2”.
  • the trigger itself is part of the classification, i.e. the combination “[before2] [trigger3] [after2]” is classified.
  • the “Trig gers” trigger if it is recognized, is assigned to the corresponding trigger definition as a trigger.
  • a combination of two or more triggers can, for example, trigger a classification of preceding, following and data arranged between the triggers, e.g. "[before3] [trigger4] [in between] [trigger5] [after3]”.
  • the occurrence of the combination of the triggers “Trigger4” and “Trigger5” triggers a classification of the preceding data “before3”, the following data “after3” and the data in between “between 1”.
  • the triggers themselves are part of the classification, i.e. The entire combination “[before3] [trigger4] [in between] [trigger5] [after3]” is classified.
  • the triggers “Trigger4” and “Trigger5” are assigned as triggers to the corresponding trigger definition when it is recognized.
  • a trigger combination can comprise any number of triggers, for example ,, [before4] [trigger6] [between2] [trigger7]
  • a trigger combination of the triggers “Trigger6” to “Trigger6 + (N + 1)” triggers a classification of the preceding data “before4”, the following data “after4” and the data in between “between2” to “ in between 2 + N “.
  • the triggers themselves are part of the classification, ie the entire combination is classified, [before4] [trigger6] [between2] [Trigger7]
  • the triggers “Trigger6” to “Trigger6 + (N + 1)”, if it is recognized, are assigned as a trigger to the corresponding trigger definition.
  • the formulation "May over” is a first trigger [Triggerl] and with the formulation "and” around a second trigger [Trigger2]
  • the structure thus corresponds to a structure of the form [before] [Triggerl] [between] [Trigger2] [after].
  • previous data [before] classified as an identity, and intermediate data [in between] and subsequent data [after] are each classified as identities.
  • the formulation "The customer bears the damage" is a trigger [trigger].
  • the structure therefore corresponds to the structure [trigger] [after].
  • the following data [after] is classified as a condition.
  • a trigger definition can stipulate that a token which is located within a radius around a specific trigger in an n-dimensional data space is to be assigned to a specific class.
  • a token which is located within a radius around a specific trigger in an n-dimensional data space is to be assigned to a specific class.
  • a trigger definition can stipulate that a token which is arranged within a plurality of radii around one trigger of a plurality of triggers is to be assigned to a specific class.
  • the n-dimensional areas delimited by the individual radii overlap and delimit an n-dimensional or lower-dimensional intersection area in the n-dimensional data space.
  • a token, which is part of this n-dimensional or lower-dimensional intersection area, is assigned to a certain class, for example.
  • a maximum trigger distance is defined for the triggers in accordance with the trigger definitions, which defines a maximum distance relative to the corresponding trigger to which a trigger effect of the trigger is limited.
  • Embodiments can have the advantage that the corresponding maximum distance is a radius around the corresponding trigger in an n-dimensional data space.
  • the trigger effect is limited to the corresponding maximum trigger distance in front of and behind the corresponding trigger.
  • the trigger effect is limited to a two-dimensional circular area around the corresponding trigger.
  • the trigger effect is limited to a spherical volume around the corresponding trigger.
  • the trigger effect is limited to a volume of an n-dimensional sphere around the corresponding trigger.
  • the maximum distance can depend on the spatial direction and be set to be of different sizes in different spatial directions.
  • the maximum trigger spacing is identical for all triggers. According to embodiments, the maximum trigger spacing is identical for a subset of the triggers. According to embodiments, the maximum trigger distance is determined individually for each trigger.
  • the corresponding maximum trigger distance can be a distance in a specific unit, depending on the type of data. For example, a sequential sequence in time is a time interval measured in a time unit, such as milliseconds, seconds or minutes.
  • a one-dimensional, two-dimensional or three-dimensional spatial data structure is a spatial distance in a spatial unit, such as millimeters, centimeters, decimeters or meters.
  • the distance can be based on pixels or voxels, for example. A corresponding distance can thus be, for example, a number of pixels or a number of voxels.
  • the distance is a logical distance. This can for example be based on elementary data elements, such as elementary characters.
  • a corresponding distance can be a number of characters, for example.
  • the corresponding spacing can be a number of elementary data elements composed of elements, such as, for example, a number of words. For example, the number is limited to a certain part of speech.
  • the distance can be limited by logical elements in the data structure, such as a punctuation mark and / or a trigger.
  • the method further comprises: • Supplement the pre-trained learning module with one or more additional trigger definitions which define additional triggers for replacing assignments of tokens in the index to the receiving class with assignments to one or more classes of a second group of classes in the course of reclassification ,
  • the additional triggers to reclassify one or more tokens assigned to the trap class in the index to one or more classes of the second group of classes by the learning module if the corresponding tokens assigned to the trap class are from one of the data records in a combination with one or more of the additional triggers are included and the corresponding additional triggers trigger a corresponding assignment to the one or more classes of the second group of classes in accordance with the corresponding additional trigger definition.
  • Embodiments can have the advantage that by adding additional trigger definitions to the learning module, the number of tokens that are assigned to the interception class can be reduced. Additional trigger definitions can be supplemented in a targeted manner in order to reclassify those tokens that are assigned to the trap class. Additional trigger definitions can therefore be supplemented as a function of the data records which the database comprises and the unknown data which they comprise.
  • additional trigger definitions are added until all tokens of the trap class are reclassified.
  • corresponding additional trigger definitions are added according to predefined intervals.
  • Corresponding predefined intervals are, for example, defined in time, based on the number of tokens included in the collection class, the amount of data stored in the database and / or the amount of data added to the database since the last addition.
  • the second group comprises classes different from the classes in the first group.
  • Embodiments can have the advantage that additional classes are defined so that those tokens of the capture class can be classified for which the meta or context information corresponding to the classes of the first group cannot be used. Rather, additional meta or context information can be defined and used by the classes of the second group.
  • one or more classes of the second group are each identical to one of the classes of the first group.
  • Embodiments can have the advantage that the additional trigger definitions provide triggers which enable the tokens of the collection class to be assigned to classes of the first group of classes.
  • the trigger definitions to be supplemented are in each case dependent on a trigger definition already included in the learning module.
  • Embodiments can have the advantage that one or more of the supplementary trigger definitions are defined in the form of additions to the trigger definitions already included in the learning module.
  • the corresponding supplementary trigger definitions extend, for example, the trigger effect of already existing trigger definitions.
  • the supplementary trigger definitions form combined trigger definitions with the already existing trigger definitions.
  • the additions are carried out repeatedly following a recursive scheme, the trigger definitions to be added to each recursion level each comprising additions to a trigger definition of a preceding recursion level so that the recursive additions are tree structures which each include one of the predetermined trigger definition as a root node.
  • Embodiments can have the advantage that the trigger effect of the existing trigger definitions is successively expanded by a progressive recursion scheme until all tokens of the collection class have been reclassified.
  • the result of the corresponding additions to the already existing trigger functions can be tree structures, for example, which can be followed by a classification of tokens.
  • the additional trigger definitions to be supplemented are received by the learning module.
  • Embodiments can have the advantage that the corresponding trigger definitions can, for example, be provided externally, for example by an administrator. This means that the relevant administrator always has the option of controlling, correcting and adding to the classification.
  • an external fine adjustment can optionally or optionally take place, for example by an administrator.
  • additional trigger definitions are extracted from the class of the unknown data, ie the catch class, tokens and assigned to existing classes and / or new classes are generated to which extracted tokens are assigned.
  • an administrator provides additional trigger definitions analogous to the trigger definitions provided, which are applied to the collection class.
  • the additional triggers are applied exclusively to the collection class and to data received in the future in accordance with the additional trigger definitions.
  • the use of an additional trigger can be implemented as an IF condition. For example, if another trigger has already been successfully applied to a data record, e.g. a trigger, and the data record also includes data classified as unknown, where an additional trigger, eg a Trigger2, is applied according to one of the additional trigger definitions.
  • the corresponding threshold value can be an absolute value which is independent of the number of tokens comprised by the index and the amount of data comprised by the database.
  • the corresponding threshold value can be a relative value which is dependent on the number of tokens comprised by the index and / or the amount of data comprised by the database
  • trigger trees or decision trees can arise behind the initially defined triggers or trigger definitions, the number of levels depending on the number of recursions N, for example the number of levels is equal to N + 1.
  • each initial trigger or each initial trigger definition forms a root point of a corresponding trigger tree or decision tree.
  • a decision tree is understood here to mean ordered, directed trees that serve to represent decision rules. If a data record includes an initial trigger, which means that part of the tokens in the data record can be classified without all tokens in the data record being able to be classified at the same time, it is checked whether the data record also includes a trigger of the first recursion.
  • the additional trigger definitions to be supplemented are created by the learning module, which comprises a statistical model, the statistical model being used for a statistical analysis of the tokens included in the collection classes and their occurrence in the data records, the result of the statistical Analysis is used to create the additional trigger definitions to be supplemented.
  • Embodiments can have the advantage that the learning module can independently create supplementary additional trigger definitions.
  • the optional or facultative fine adjustment described above takes place using the statistical model.
  • the statistical model identifies, e.g. by frequency analyzes and correlation analyzes, triggers within the unknown data, which are then applied to the tokens classified as unknown in analogy to the procedure described above. According to execution forms, a recursive procedure using the statistical model can also take place.
  • the method further comprises:
  • an administrator can recognize errors in classified classes and correct them if necessary, for example by providing a corrected trigger definition, on the basis of which tokens are reclassified.
  • Embodiments can have the advantage that a correction of trigger Definitions at any point in the process is made possible.
  • the trigger definitions can be checked after training the learning module. If correction trigger definitions are identified, correspondingly corrected trigger definitions are provided.
  • Embodiments can have the advantage that corrected trigger definitions can also be provided at a later point in time when incorrect classifications are recognized. Administrative intervention in the learning and classification process is therefore possible at any time. This allows errors in the learning system to be corrected without having to convert the entire model.
  • the pointers with which the tokens are stored linked in the index each point to one or more of the field values in the stored data records.
  • Embodiments can have the advantage that a finer granularity can be achieved when determining the origin of tokens in the data records. Such a finer granularity also enables the relative relationships of the tokens within the data records to be broken down and taken into account in an analysis or other use of the index.
  • the field values of the additional data record include text data, image data, audio data and / or video data.
  • the method can be used, for example, for signal processing, such as 1 D audio recognition, 2D and 3D image processing, or ND data input from N sensors, etc.
  • the method can be used, for example, for an analysis of stream data (bitstream or . Bitstream).
  • a bit stream also known as a bit stream, designates a sequence of bits that represent a flow of information, ie a serial or sequential signal.
  • a bit stream is thus a sequence of bits of indefinite length in chronological order.
  • a bit stream for example, represents a data stream divided into logical structures, which is divided into more fundamental Small structures such as symbols of a fixed size, ie bits and bytes, and can be further divided into blocks and data packets of different protocols and formats.
  • generating the tokens includes applying tokenization logic to the field values of the additional data record, which logic includes a full-text indexer that is configured to break down texts into words and to output the words as tokens.
  • Embodiments can have the advantage that effective tokenization of texts or text files can be implemented.
  • the corresponding text files can be any text.
  • the corresponding text files can be measured value files or algorithms for controlling computers and / or technical systems.
  • the field values of the additional data record include full texts, the full texts including words and / or one or more numbers formed from letters of one or more alphabets.
  • Full-text indexing involves breaking down texts into individual words, with the individual words of a text field then being stored in an index assigned to this field.
  • Full text indexing is only supported if the corresponding field is used to selectively store a certain data type, e.g. CFIAR, VARCFIAR or TEXT, is configured.
  • CFIAR CFIAR
  • VARCFIAR VARCFIAR
  • TEXT TEXT
  • natural language text in JSON format can be stored in a field.
  • generating the tokens includes applying tokenization logic to the field values of the additional data record, which logic includes a generic tokenizer that is configured to recognize data of different data types in the field values and to generate from these tokens in different data types .
  • tokenization logic includes a generic tokenizer that is configured to recognize data of different data types in the field values and to generate from these tokens in different data types .
  • Embodiments can have the advantage that effective tokenization can be implemented for different types of data, such as text data, image data, audio data and / or video data.
  • the method further comprises: • Receiving a search query, the search query containing a search value,
  • Embodiments can have the advantage that the index can be used for effective searches in the data records even though they are stored in their original form.
  • the learning module can search for patterns and / or regularities within the data sets using appropriate search queries.
  • the index stores all tokens generated from the field values of the data records of a database in such a way that the index contains each token only once.
  • Each token contains pointers to one or more of the data records from whose field values it was generated. If an index generated according to the invention is searched for a specific search value and a token stored in the index is identified as the result of the search, which is identical to the search value, this token uses pointers to refer to all data records that contain this token at least once contained in at least one of their field values and which were used to create the index.
  • the data records, which represent a “hit” with regard to the search value can be identified and returned very quickly using the references, without the need for a sequential search across all data records.
  • the search value further comprises a class assignment and the identification of the token within the index further requires that the identified token has the same class assignment.
  • Embodiments can have the advantage that class assignments and thus meta or context information indexed with the class assignments can be taken into account in the search queries.
  • triggers are identified in the index with a flag.
  • the search value furthermore includes an assignment to a trigger definition and / or a flag identifying a trigger, and the identification of the token within the index further requires that the identified token is assigned to the same trigger definition and / or has the same flag .
  • tokens which are assigned to the trap class are excluded from the search.
  • Embodiments can have the advantage that the resulting search results have a high degree of reliability, since unknown data are excluded from the search.
  • the method further comprises pre-training the learning module.
  • the pre-training includes:
  • the predetermined trigger definitions define initial triggers that are used to structure or classify received data. According to embodiments, before data is loaded into the database, the initial triggers are specifically defined, ie predetermined trigger definitions are given. If data is loaded, these initial triggers enable an initial classification according to known classes as well as unknown data which are assigned to the collection class. According to embodiments, the pre-training further comprises:
  • data e.g. Text data, audio data, image data, video data or N-dimensional data from N sensors are loaded into the database and the triggers are used to automatically classify the data.
  • Embodiments can have the advantage that the learning module can be effectively pretrained in this way on the basis of the predetermined trigger definitions.
  • These predetermined trigger definitions can serve as a basis for obtaining further trigger definitions, for example by combining trigger definitions.
  • trigger definitions For example, there is an automatic learning phase of the database system or the learning module, which includes a combination of the initial triggers.
  • the initially loaded triggers can be combined, as described above, based on the data comprised by the data records, and thus the number of trigger definitions available can be increased.
  • token combinations that have already been classified can be identified. The purpose of this is to ensure that identical data that is later loaded into the database do not have to be reclassified, but are already marked as "known" in the system.
  • generating one of the additional tokens comprises using one of the field values of the additional data set in its entirety as the corresponding additional token. It is entirely possible that the index also contains tokens from fields to which no tokenization is applied or whose content simply cannot be divided into individual tokens. According to embodiments, generating one of the additional tokens comprises dividing one of the additional field values of the additional data record into a plurality of subfield values and using one of the subfield values as the corresponding additional token. Embodiments can have the advantage that the granularity of the data used or the tokenization can be adapted independently of the granularity of the fields.
  • the index stores all tokens generated from the field values of the stored data records in such a way that the index contains each token exactly once for each of the token assignments of the corresponding token.
  • the further data model is structured in such a way that the tokens and token assignments of the index stored in the further data model meet the fifth and / or sixth normal form.
  • Embodiments can have the advantage that redundancies can be avoided.
  • the tokens, the class assignments and the assignment to the trigger definitions can be stored in the form of relations or equivalent structures.
  • a relation is a set of tuples.
  • a tuple is a set of attribute values.
  • An attribute designates a data type or a property that is assigned to one or more data.
  • the number of attributes determines the degree, the number of tuples the cardinality of a relation.
  • Normalization in particular normalization of a relational data model, is understood to mean a division of attributes into a plurality of relationships according to a normalization rule, so that redundancies are reduced or minimized.
  • a relational data model can be implemented, for example, in table-like data structures in which the relations are implemented in the form of tables, the attributes in the form of table columns and the tuples in the form of table rows.
  • a relational data model can be brought into a normal form, for example, in that the relations of the data schema are progressively broken down into simpler relations based on the functional dependencies applicable to the corresponding normal form.
  • Normal form (1 NF), 2nd normal form (2NF), 3rd normal form (3NF), Boyce-Codd normal form (BCNF), 4th normal form (4NF), 5th normal form (5NF), 6th normal form ( 6NF).
  • a relation is in the first normal form if each attribute of the relation has an atomic range of values and the relation is free of repeating groups.
  • atomic is understood to mean an exclusion of composite, quantity-valued or nested value ranges for the attributes, ie relation-valued attribute value ranges. Freedom from repeating groups requires that attributes that contain the same or similar information are relocated to different relations.
  • a relation is in the second normal form if it meets the requirements of the first normal form and no non-primary attribute is functionally dependent on a real subset of a key candidate.
  • a non-primary attribute is an attribute that is not part of a key candidate. This means that each non-primary attribute is dependent on all whole keys and not just on part of a key.
  • Relations in the first normal form, the key candidate data of which is not composed but consist of a single attribute therefore automatically satisfy the second normal form.
  • a key candidate is understood here to be a minimal set of attributes which uniquely identify the tuples of a relation.
  • a relation is in the third normal form if it fulfills the requirements of the second normal form and no non-key attribute depends transitively on a key candidate.
  • An attribute is transitively dependent on a key candidate if the corresponding attribute is dependent on the corresponding key candidate via a further attribute.
  • a relation is in Boyce-Codd normal form if it meets the requirements of the third normal form and every determinant is a super key.
  • a determinant is understood here to be a set of attributes on which other attributes are functionally dependent. A determinant thus describes the dependency between the attributes of a relation and defines which attribute sets determine the value of the other attributes.
  • a super key is a set of attributes in a relation which uniquely identify the tuples in this relation. The attributes of this set therefore always include different values for tuples selected in pairs. The key candidate is therefore a minimal subset of the attributes of a super key, which enables the tuple to be identified.
  • a relation is in the fourth normal form if it fulfills the requirements of the Boyce-Codd normal form and does not include any nontrivial multivalued dependencies.
  • a relation is in the fifth normal form if it fulfills the requirements of the fourth normal form and does not include any multi-valued dependencies that are dependent on one another.
  • the fifth normal form is thus present if every nontrivial composite dependency is implied by the key candidates.
  • a link dependency is implied by the key candidates of the output relation if each relation of the set of relations is a super key of the output relation.
  • a relation is in the sixth normal form if it fulfills the requirements of the fifth normal form and does not include any nontrivial join dependencies.
  • a relation is sufficient for a join dependency of a plurality of relations if the relation as the starting relation can be broken down into the corresponding set of relations without loss.
  • the compound dependency is trivial if one of the relations of the set of relations has all the attributes of the starting relation.
  • At least the document-based data model used by the multi-model database management system to store the data sets is a NoSQL data model.
  • the DBMS is a NoSQL DBMS. This can be advantageous because it has been found that NoSQL DBMS in particular, which often have a more flexible structure than classic SQL-based DBMSs. Due to the flexibility of their structure, NoSQL DBMSs are therefore particularly suitable for the management and storage of data records from which an index can be created according to embodiments of the invention.
  • the index has the structure of a tree, in particular a B + tree.
  • Embodiments can have the advantage that a tree structure, in particular the structure of a B + tree, enables a particularly efficient and fast search for the tokens stored in the index.
  • a B + tree is a data and / or index structure that is an extension of a B-tree. With a B + tree, the actual Data elements are only stored in the leaf nodes, while the inner nodes only contain keys.
  • several of the data records stored in a document-oriented data model each comprise a different number of fields.
  • Embodiments can have the advantage that data sets of different sizes and structures or granularity can be processed.
  • the fields each have a common, generic data format.
  • Embodiments can have the advantage that since a large number of different data types can be stored in a specific field. A user or an application program who wants to save data records in the database does not have to worry about the consistency and fit of data types. A high degree of flexibility with regard to the structure and the scope of the data records that can be managed and stored by the multi-model database management system is therefore offered.
  • the learning module or the machine learning implemented by it is configured for data extraction, consistency checking, image recognition, speech recognition, voice control, device monitoring and / or autonomous device control.
  • This can, for example, already consist in the classification of the tokens, with tokens assigned to the collection class as unknown data being viewed, for example, as an indication of a potential malfunction. For example, this can be based on the index with the tokens and their meta or context information, which serve as the basis for an additional algorithm for machine learning applied thereon.
  • the collection class is emptied by adding additional trigger definitions, so that meta or context information is provided for all tokens of the database system.
  • Data extraction can include, for example, recognizing and extracting a pattern in a text, image, audio or video file. This pattern can, for example, be defined by a trigger definition or it can be recorded in the classified data.
  • a corresponding pattern can be, for example, a predetermined event recorded in the form of sensor values, for example a person in an effective range of a device.
  • a consistency check can include, for example, a consistency check in a text, image, audio or video file. In this case, it is checked, for example, whether the corresponding data include unknown and thus inconsistent data, include data that deviate significantly from the remaining data or include data that is explicitly predefined as inconsistent.
  • a corresponding consistency check can be used, for example, to check for errors in control algorithms of devices, to detect malfunctions using measurement data from a function of a device, or to detect errors in text files, for example in the form of a spell check.
  • Image recognition can be used to recognize objects, events or features in image or video files. For example, context information on what is visually represented is recorded and / or displayed. This can include, for example, a visual display of information, that is, the addition of images or videos with computer-generated additional information or virtual objects by means of fading in / overlaying. Such a method is generally referred to as augmented reality. Furthermore, image recognition can be based on annotated image or video files.
  • Speech recognition can be used to recognize speech in audio files or video files, for example for voice control or for converting speech into text.
  • Pattern recognition in text, image, audio or video files can be used for device monitoring. In particular, it can be occurring or threatening Malfunctions are detected. This can be used for safety and enables predictive maintenance of the corresponding device, since potential problems can be identified early on.
  • a corresponding text file is, for example, a data record with sensor measured values.
  • an autonomous device control can also be implemented, for example an autonomous control of vehicles, robots or industrial plants.
  • a “device” is generally understood here to mean a technical device with sensors for capturing status data of the device and a device computer system for logging the captured status data.
  • the device can also exist in the corresponding computer system with sensors.
  • the received data sets are data sets acquired by a device computer system using the sensors.
  • Computer system for machine learning A device comprises, for example, a vehicle, a system such as a production system, a processing system, a conveyor system, an energy generation system, a heat generation system, a control system, a monitoring system, etc.
  • a “vehicle” is understood here to mean a mobile means of transport. Such a means of transport can be used, for example, to transport goods (goods traffic), tools (machines or auxiliary equipment) or people (passenger traffic). Vehicles in particular also include motorized means of transport.
  • a vehicle can, for example, be a land vehicle, a water vehicle and / or an aircraft.
  • a land vehicle can be, for example: an automobile such as a passenger car, bus or truck, a motor-driven two-wheeler such as a motorcycle, moped, motor scooter or motorcycle, an agricultural tractor, forklift, golf cart, truck crane.
  • a land vehicle can also be a rail-bound vehicle.
  • watercraft can be: a ship or boat.
  • an aircraft can be, for example: an airplane or helicopter.
  • a vehicle is also understood to mean, in particular, a motor vehicle.
  • the device comprises at least one sensor for detecting status data of the device.
  • the status data of the device are received by the device computer system from the at least one sensor.
  • the device comprises a plurality of sensors for acquiring status data of the device.
  • Embodiments can have the advantage that the device's own sensor system can be used to detect the state of the device.
  • the state of the device can, for example, be described by information on parameters of the current performance of the device, such as the mileage of a vehicle, consumption values, performance values, error messages, results of predefined test protocols and / or identifiers of components of the device.
  • Parameters of the current performance of a vehicle can be, for example, engine speed, speed, fuel consumption, exhaust gas values, and transmission gear.
  • a “sensor” is understood here to mean an element for recording measurement data.
  • Measurement data are data which physical or chemical properties of a measurement object, such as amount of heat, temperature, humidity, pressure, flow rate, sound field sizes, brightness, acceleration, pH value, ion strength, electrochemical potential, and / or its material properties qualitatively or quantitatively reproduce. Measurement data are recorded using physical or chemical effects and converted into an electronic signal that can be further processed electronically. Furthermore, measurement data can reflect states and / or changes in state of electronic devices due to external influences and / or as a result of use by a user.
  • Sensors for capturing status data in a vehicle can include, for example: crankshaft sensors, camshaft sensors, air mass sensors, Air temperature sensor, cooling water temperature sensor, throttle valve sensor, knock sensor, transmission sensor, distance sensor, transmission sensor, level sensor, brake wear sensor, axle load sensor, steering angle sensor. These sensors record and monitor the driving behavior of the vehicle. Malfunctions can be recognized and identified from deviations from target values and / or the occurrence of certain patterns. In some cases, specific causes of errors, such as failed vehicle components, can also be identified. Sensors can also query the identifiers of electronic components that are installed in the vehicle in order to check their identity.
  • Embodiments include a computer system for machine learning, where the computer system has one or more processors, a database provided by one or more data storage media, a multi-model database management system which manages the database and is configured to do so of data records in a document-oriented data model in the data storage media, the stored data records each comprising one or more field values, the individual field values of the stored data records being stored in a field, the field values of the stored data records each having one or more field types A plurality of different field types are assigned, a pre-trained learning module for machine learning and a program logic comprises.
  • the database further comprises a searchable index which is stored in a further data model, the index comprising a plurality of tokens generated from the field values of the stored data records, with tokens in the index each having one or more pointers to one or more of the data records stored in the document-oriented data model are linked, from whose field values the corresponding token was generated.
  • First tokens in the index which are included in one of the trigger definitions as triggers, are each assigned to the corresponding trigger definition, with second tokens in the index each being one or more classes of the first group of Classes are assigned, and the remaining tokens in the index for identifying the corresponding remaining tokens as unknown data are assigned to a trap class, the assignment to the trap class being an assignment to one of the trigger definitions as well as an assignment to one of the classes of the first Excludes group of classes.
  • the program logic is configured to carry out a method for machine learning.
  • the procedure includes:
  • the learning module identifies the corresponding token as a trigger
  • the learning module Using the identified triggers for assigning one or more second additional tokens to one or more classes of the first group of classes by the learning module, if the corresponding second additional tokens from the additional data set in a combination with one or more of the identified triggers according to one of the trigger definitions are included, the corresponding triggers triggering a corresponding class assignment according to the corresponding trigger definition,
  • the computer system is configured to execute one or more of the aforementioned embodiments of the method for machine learning.
  • Embodiments can have the advantage that they create a self-learning system which works on all data in the database, does not use any randomness in the decision-making or classification process, and uses initial defined triggers to classify received data. According to embodiments, the system also allows at any time, i.e. Even after an initial learning phase, external interventions in the decision-making processes.
  • the computer system provides a system for machine learning on the basis of a database, which divides any data into known classes and unknowns using initially defined trigger definitions. Due to the assignment to known classes, meta or context information is identified in the data records.
  • the index provided enables search processes and / or machine learning processes to run efficiently on the data comprised by the data records. This can be done without explicit access to the data records, i.e. exclusively on the index, or with explicit access to relevant data records based on pointers which are linked to tokens identified in the index.
  • Figure 1 is a schematic block diagram of an embodiment of a
  • Figure 2 is a schematic block diagram of an exemplary data processing by the multi-model database management system
  • FIG. 3 is a schematic block diagram of an exemplary data processing by the multi-model database management system
  • FIG. 4 is a schematic block diagram of an embodiment of an exemplary computer system
  • FIG. 5 shows a flow chart of an embodiment of an exemplary method
  • FIG. 6 shows a flow chart of an embodiment of an exemplary method
  • FIG. 7 shows a flow chart of an embodiment of an exemplary method
  • FIG. 8 shows a flow chart of an embodiment of an exemplary method
  • FIG. 9 shows a flow diagram of an embodiment of an exemplary method.
  • FIG. 1 shows a block diagram of an embodiment of an exemplary computer system 100 for machine learning.
  • the computer system 100 comprises at least one database 104 and a multi-model database management system (MM-DBMS) 1 18.
  • the MM-DBMS 1 18 manages the, possibly structured, storage of the data in the at least one database 104 and controls all transmission and writing access to the database 104.
  • the MM-DBMS 1 18 supports at least two data models 106, 1 10, in which the data in the database 104 is stored.
  • the database model defines the form in which the relevant data is organized, saved and processed.
  • One or both data models 106, 1 10 are NoSQL data models.
  • the MM-DBMS 1 18 is a NoSQL DBMS.
  • the first data model 106 is a document-based data model in which a plurality of data records DS1, DS2, DS3 are stored. Each data record DS1, DS2, DS3 is saved in a document or a data container. No specific structure is specified for the data records DS1, DS2, DS3 themselves when they are stored by the document-based data model 106.
  • the data records DS1, DS2, DS3 can therefore be stored with the internal structure with which the data records DS1, DS2, DS3 are received from the database 104.
  • the data records DS1, DS2, DS3 stored in the document-based data model 106 are raw data.
  • the data records DS1, DS2, DS3 can include, for example, text data, image data, audio data and / or video data.
  • the data records DS1, DS2, DS3 each include at least one field F 1, ..., F8, with field values.
  • the DS3 can already have an internal structure with a plurality of fields F1, ..., F8 when they are stored. Then the corresponding data records DS1, DS2, DS3 each include a plurality of fields F1, ..., F8. If the data records DS1, DS2, DS3 do not have any fields even when they are received, they each include, for example, exactly one field in stored form which includes the entire data volume of the corresponding data record DS1, DS2, DS3.
  • the fields F 1, ..., F8 each comprise one or more field values. Each of the field values of a data record DS1, DS2, DS3 is stored in a corresponding field, a type of data container. Each field F1, ..., F8 can be assigned to a field type.
  • the fields F 1, ..., F8 can be assigned different or all of the same field type.
  • the composition of the field values of the individual data records DS1, DS2, DS3 can differ in terms of their field types. It is also possible that individual data records do not contain any fields of a certain field type.
  • mandatory field types can also be defined, i.e. that each document includes a field for each mandatory field type and optionally includes one or more additional fields for optional field types.
  • the data of the data records are then stored in fields of the field type intended for them, i.e. e.g. Text data in one or more text fields, image data in one or more image fields, audio data in one or more audio fields and / or video data in one or more video fields.
  • the computer system 100 further comprises a learning module for processing the data stored in the database 104.
  • the learning module 120 comprises, for example, at least one tokenizer 120 for generating tokens 109, trigger Definitions 123, which define triggers for a classification of tokens 109, and / or a classifier 124, which classifies the tokens 109 using the trigger definitions 123.
  • the learning module 120 further comprises a statistical model 125.
  • the statistical model 125 can be configured to detect trigger combinations and to create combined trigger definitions, to create additional trigger definitions and / or to assign corrected trigger definitions create.
  • the MM-DBMS 118 can also include the tokenizer 122 and / or access a tokenizer 122 provided by the learning module 120.
  • the trigger definitions 123 can also be stored in the database 104.
  • the MM-DBMS 118 and / or the learning module 120 have a built-in program logic that is configured to generate an index 112.
  • the corresponding index 112 is provided in a further data model 110 in which the complete data of the data records DS1, DS2, DS3 are stored in a restructured, redundancy-free form.
  • the tokenizer 122 is accessed, which is configured to tokenize the field values of the fields F1,..., F8 of the data records 106 stored in the document-based data model 106.
  • the resulting tokens 109 can also be identical to a field value of a field or a data record if no further breakdown into tokens 109 is possible or useful.
  • the tokenization can also take place in stages, so that an ever finer breakdown takes place. In this case, the resulting index 112 therefore includes tokens 109 which are composed of other tokens 109.
  • all or at least most of the field values of all data records DS1, DS2, DS3 of the database 104 are tokenized, so that an extensive amount of tokens 109 is produced.
  • the tokens 109 can be a mixture of numbers, letters, images or image segments, audio files or audio elements or other data structures, in particular sensor data from one or more Sensors include.
  • Each of the tokens 109 generated is stored in the index 112 linked to a pointer, the pointer pointing to the data record or the field from which the token 109 originates.
  • tokens 109 in index 112 which are included as triggers in one of the trigger definitions 123, are each assigned to the corresponding trigger definition 123. Furthermore, tokens 109 in the index 112, which are comprised by one of the data records DS1, DS2, DS3 in a combination with one or more of the identified triggers according to one of the trigger definitions 123, are each assigned to one or more classes. The corresponding class assignments provide meta or context information for the corresponding tokens 109. Finally, the remaining tokens 109 in the index 112, which using the trigger definitions 123 neither identify themselves as triggers nor assign a class as, are assigned to a collection class for identification as unknown data. An assignment to the collection class excludes an assignment to one of the trigger definitions 123 as well as an assignment to one of the classes according to the trigger definitions 123. The assignments described above take place, for example, using the classifier 124 of the learning module 120.
  • a non-redundant, unique token set is formed from the set of tokens 109, in which each of the tokens 109 occurs only once. Even if a token 109 with a certain value and a certain class assignment occurs several times in the database 104 or in the data model 106, it is only saved once with this class assignment in the non-redundant token set and in the index 112. All tokens 109 of the non-redundant token set are preferably stored in the index 112 in such a way that the tokens 109 are sorted according to a sorting criterion and are stored in sorted form in the index structure.
  • the sorting can take place, for example, using the alphabet for alphanumeric data or other sorting criteria adapted to the data. Since the tokens 109 are preferably stored in the index 1 12 in sorted form, and are also preferably stored in a tree structure, it is possible very quickly to identify a specific token 109 within the index 112 and then to use the references of this identified token 109 to one or more data records DS1, DS2, DS3 in order to very quickly identify those data records which contain a specific token 109 searched for. It is therefore not necessary to search through all data records DS1, DS2, DS3 of the database 104 sequentially.
  • FIG. 2 shows a schematic block diagram of an exemplary data processing by the multi-model database management system and the learning module.
  • This trigger definition 123 defines two triggers, i.e. a first trigger “lives in” and a second trigger “in”.
  • the trigger definition defines that a token immediately preceding the first trigger is a surname, while a token immediately preceding the surname is a first name.
  • Trigger definition also defines that a token arranged between the two triggers is a street and that a token immediately following the second trigger is a city.
  • Two documents 108 are stored in a document-based data model 106 of a database.
  • Each document 108 each comprises a data record DS1, DS2.
  • the data records DS1, DS2 are each a text file.
  • the first data record DS1 includes, for example, the sentence: "Muster firstname_1, MusterstadM_1 lives in Musterstrasse_1 in MusterstadM". This sentence is broken down into tokens 109 by means of a tokenizer: "Muster first name_1", “Muster lastname_1", “lives in the", “Musterstr._1", “in”, “MusterstadM”.
  • the two tokens “lives in” and “in” are identified as triggers according to the trigger definition 123.
  • the remaining tokens 109 are each assigned the Definition assigned to defined classes 1 1 1.
  • the tokens identified as triggers like the tokens classified using these triggers, are stored in an index in a second data model 110.
  • the triggers are each assigned to the trigger definition 123 in the form of a trigger assignment 1 17.
  • the remaining tokens 109 are each stored in the form of a class assignment 1 13 assigned to one of the classes defined by the trigger definition 123.
  • the two tokens “lives in” and “in” are identified as triggers according to the trigger definition 123. Since these two triggers of the trigger definition 123 are already included in the index, they are not stored again in the second data model 110. Only a pointer to the second data record DS2 is added. Using the identified triggers and the trigger definition 123, the remaining tokens 109 of the data record DS2 are each assigned to the classes 1 1 1 defined by the trigger definition.
  • the classified tokens 109 of the data set DS2 are each stored in the form of a class assignment 1 13 assigned to one of the classes defined by the trigger definition 123 and linked with a pointer 1 15 to their storage location in the first data model, ie DS2. All tokens of the second data record DS2 are therefore also stored in a redundancy-free form, each with their class assignments in the second data model 110 linked with a pointer to their storage location in the first data model.
  • FIG. 3 shows a schematic block diagram of an exemplary data processing by the multi-model database management system and the learning module.
  • This trigger definition 123 is used to classify tokens generated from an image file, the image file being broken down into tokens in the form of pixel groups.
  • the trigger definition 123 defines two triggers, i.e. a first trigger in the form of a pixel group with the content “+” and a second trigger in the form of a pixel group with the content “x”.
  • the trigger definition defines that a pixel group which is arranged within a first radius of N pixels around the first trigger and at the same time within a second radius of N pixels around the second trigger is a token of the Class "class" acts.
  • a document 108 is stored in a document-based data model 106 of a database.
  • This document 108 comprises a data record DS.
  • the data record DS is a two-dimensional image file.
  • This image file is broken down into tokens by means of a tokenizer, the tokens each being pixel groups 150.
  • the pixel groups of equal size in Z by Z are broken down.
  • the tokens include, for example, a first token in the form of a pixel group with the content “x”, a second token in the form of a pixel group with the content “+”, a third token in the form of a pixel group with the content “#” and a fourth token in Shape of a group of pixels with the content
  • the two tokens “+” and x ” are identified as trigger 121 according to trigger definition 123.
  • the third token “#” becomes the one defined by the trigger definition Assigned to class 111, since it is arranged in the two-dimensional image file within a first radius 152 of N pixels around the first trigger “+” and at the same time within a second radius 154 of N pixels around the second trigger “x”. Since the fourth token does not fall under the trigger definition 123, it is assigned to the trap class as an unknown date.
  • the tokens “+” and “x” identified as triggers 121 are stored in an index in a second data model 110, as are the token “#” classified using these triggers and the token assigned to the collection class.
  • the triggers “+” and “x” are each assigned to the trigger definition 123 in the form of a trigger assignment 117.
  • the token “#” is stored in the form of a class assignment 113 assigned to the classes defined by the trigger definition 123.
  • the token is stored in the form of an assignment 119 assigned to the collection classes.
  • all triggers and classified tokens in the second data model 110 are identified with a pointer 115 to their storage location in the first data model, i. E. DS, linked.
  • FIG. 4 shows a schematic block diagram of an embodiment of an exemplary computer system 110.
  • the computer system 100 comprises a processor 114, which program instructions 116, whereby the computer system is caused to carry out the above-described method for machine learning.
  • the processor 114 executes a multi-model database management system 118 and a learning module 120 for machine learning with a tokenizer 122 and a classifier 124.
  • the learning module 120 includes trigger definitions 123.
  • the learning module 120 also includes a statistical model 125.
  • the computer system 110 also includes a database 104 in a memory 102, which is managed by the multi-model database management system 118.
  • the database comprises a first data model 106, for example a document-oriented data model, in which data records 108 are stored.
  • the database also includes a second data model 110 with an index 112 of all the data stored in the data records 108.
  • FIG. 5 shows a flow chart of an embodiment of an exemplary method for machine learning.
  • a pre-trained learning module for machine learning is provided, which comprises a plurality of predetermined trigger definitions. These predetermined trigger definitions define triggers for assigning tokens to classes of a group of classes.
  • a database is provided.
  • the database is managed by a multi-model database management system and comprises a plurality of data sets which are stored in a document-oriented data model.
  • the database provided includes a searchable index of all the data comprised by the stored data records.
  • This index is stored redundancy-free in a further data model managed by the multi-model database management system.
  • the index comprises a plurality of tokens generated from the field values of the stored data records, each of which is linked in the index with one or more pointers to one or more of the data records and / or fields stored in the document-oriented data model, from whose field values the corresponding Token was generated.
  • the first tokens in the index which are included as triggers by one of the trigger definitions, are each assigned to the corresponding trigger definition.
  • Second tokens in the index are each assigned to one or more classes of the group of classes.
  • the remaining tokens in the index are finally assigned to a collection class to identify the corresponding remaining tokens as unknown data.
  • the assignment to the collecting class excludes an assignment to one of the trigger definitions as well as an assignment to one of the classes of the first group of classes.
  • an additional data record is received and in block 206 it is stored by the multi-model database management system in the document-oriented data model of the database. The storage takes place in a document or data container.
  • one or more additional Token generated from additional field values which the additional data record includes.
  • one or more first additional tokens are identified as triggers if these are included as triggers by one of the trigger definitions.
  • the remaining additional tokens are classified.
  • the triggers identified in block 210 are used to assign one or more second additional tokens to one or more classes of the group of classes if the corresponding second additional tokens from the additional data set are in combination with one or more of the identified triggers are included in accordance with one of the trigger definitions and the corresponding triggers trigger a corresponding class assignment in accordance with the corresponding trigger definition.
  • the remaining additional tokens, for which no assignment to one of the trigger definitions and no class assignment based on one of the trigger definitions has been made, are assigned to the collection class in the course of the classification in block 212.
  • the index is supplemented by the multi-model database management system using the data record stored in the document-oriented data model using the additional tokens from block 208, the class assignments of the additional tokens from block 212 and a pointer to the additional tokens. If pointers indicate individual fields of the additional data record, a plurality of pointers is used for a plurality of fields.
  • the addition in block 214 can include comparing the additional tokens with the index. If one of the additional tokens is not included in the index, the corresponding additional token is added to its class assignments in the index and linked to the pointer to the additional data record stored in the document-oriented data model. If one of the class assignments of an additional token included in the index is not included in the index, the corresponding class assignment is supplemented with the corresponding additional token in the index and the corresponding additional token in the index with the pointer to the additional token in the documents - oriented data model linked to stored data set. If one of the additional tokens with all of their class assignments is included in the index, then the corresponding additional token in the index is linked with the pointer to the additional data record stored in the document-oriented data model.
  • the addition in block 214 can include marking combinations of second additional tokens with one or more of the identified triggers, which have triggered a class assignment in accordance with one of the trigger definitions, in the index as classified combinations.
  • Class assignments are only carried out for combinations of second additional tokens and one or more identified triggers that are not marked as classified combinations.
  • token combinations are compared with the index. If the index already includes the corresponding token combination and this is marked as classified, there is no new classification for this token combination. Only the corresponding token combination and / or the partial combinations and individual tokens comprised by the corresponding token combination are linked in the index with the pointer to the additional data record stored in the document-oriented data model.
  • FIG. 6 shows a flow diagram of an embodiment of an exemplary method for generating combined trigger definitions.
  • one or more trigger combinations are identified by the learning module, which are each comprised by at least one of the data sets and which meet a combination criterion.
  • the trigger definitions of the triggers of the corresponding trigger combinations are combined into one or more additional combined trigger definitions.
  • the plurality of predetermined trigger definitions of the learning module is supplemented by the one or more additional combined trigger definitions.
  • FIG. 7 shows a flow chart of an embodiment of an exemplary method for adding additional trigger definitions.
  • the previously trained learning module is supplemented by one or more additional trigger definitions.
  • the additional trigger definitions define additional triggers for a replacement of assignments of tokens in the index to the receiving class by assignments to one or more classes of a further group of classes in the course of a reclassification.
  • the additional trigger definitions can be received by the learning module, for example.
  • the corresponding additional trigger definitions are provided by an administrator.
  • the additional trigger definitions to be supplemented are created by the learning module.
  • the learning module comprises a statistical model which is used for a statistical analysis of the tokens included in the collection classes and their occurrence in the data records. The result of the statistical analysis is used to create the additional trigger definitions to be added.
  • one or more tokens assigned to the trap class are reclassified in the index, which tokens define the additional trigger definitions as additional triggers.
  • the reclassification by the learning module includes a replacement of the assignment to the collection class by an assignment to the corresponding additional trigger definition, which includes the corresponding token as an additional trigger.
  • the additional triggers for reclassifying one or more tokens assigned to the collection class in the index to one or more classes of the further group of classes are used by the learning module if the corresponding tokens assigned to the collection class are from one of the data records in a Combination with one or more of the additional triggers are included and the corresponding additional triggers trigger a corresponding assignment to the one or more classes of the further group of classes in accordance with the corresponding additional trigger definition.
  • the method for adding additional trigger definitions can be executed repeatedly following a recursive scheme.
  • the too Supplementary trigger definitions of each recursion level each include additions to a trigger definition of a preceding recursion level, so that the recursive additions form tree structures which each include one of the predetermined trigger definition as a root node.
  • FIG. 8 shows a flow diagram of an embodiment of an exemplary method for correcting trigger definitions in blocks.
  • a corrected trigger definition for replacing one of the stored trigger definitions of the learning module is received.
  • This corrected trigger definition is provided, for example, by an administrator.
  • the corrected trigger definition is created by the learning module using a statistical model.
  • the corresponding stored trigger definition is replaced by the corrected trigger definition.
  • the tokens classified using the corresponding stored trigger definition are reclassified, the reclassification taking place using the corrected trigger definition.
  • FIG. 9 shows a flowchart of an embodiment of an exemplary method for performing a search on the database.
  • a search query is received that includes a search value.
  • the index is searched for the search value, and in block 604 a token is identified within the index which is identical to the search value.
  • the search value can also specify a class assignment in addition to a token value.
  • the identification of the token within the index further requires that the identified token has the class assignment specified in the search query.
  • tokens which are assigned to the collection class are excluded from the search.
  • pointers are analyzed with which the identified token is linked. This determines one or more of the data records which contain one or more field values from which the indexed token was generated.
  • the particular records or one or more references to the particular records are returned in response to the search query. List of reference symbols

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Business, Economics & Management (AREA)
  • Artificial Intelligence (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

L'invention concerne un procédé informatisé d'apprentissage par machine. Un module d'apprentissage (120) préentraîné et une banque de données (104) sont préparés. Le module d'apprentissage comprend une pluralité de définitions de déclencheurs (123) préétablies, lesquelles définissent des déclencheurs (121) destinés à associer des jetons (109) à des classes (111) d'un groupe de classes. Un ensemble de données additionnel (108) sauvegardé dans un premier modèle de données (106) de la banque de données est reçu, un ou plusieurs jetons sont produits. Des déclencheurs sont identifiés parmi les jetons et associés respectivement à la définition de déclencheurs fondamentale. Les déclencheurs identifiés sont utilisés pour associer d'autres jetons produits à une ou plusieurs classes du groupe de classes. Des jetons produits restants, pour lesquels aucune association à une des définitions de déclencheurs ni aucune association de classe sur la base d'une des définitions de déclencheurs n'a eu lieu, sont associés à une classe de collecte. Dans un deuxième modèle de données (110), un index (112) est complété à l'aide des jetons produits, des associations de classe des jetons produits et d'un pointeur (115) sur l'ensemble de données additionnel sauvegardé.
PCT/EP2020/059042 2019-04-04 2020-03-31 Apprentissage par machine sur la base de définitions de déclencheurs Ceased WO2020201249A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
EP20719125.5A EP3948578A1 (fr) 2019-04-04 2020-03-31 Apprentissage par machine sur la base de définitions de déclencheurs

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
DE102019108858.5 2019-04-04
DE102019108858.5A DE102019108858A1 (de) 2019-04-04 2019-04-04 Maschinelles Lernen auf Basis von Trigger-Definitionen

Publications (1)

Publication Number Publication Date
WO2020201249A1 true WO2020201249A1 (fr) 2020-10-08

Family

ID=70289372

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2020/059042 Ceased WO2020201249A1 (fr) 2019-04-04 2020-03-31 Apprentissage par machine sur la base de définitions de déclencheurs

Country Status (3)

Country Link
EP (1) EP3948578A1 (fr)
DE (1) DE102019108858A1 (fr)
WO (1) WO2020201249A1 (fr)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE19627472A1 (de) * 1996-07-08 1998-01-15 Ser Systeme Ag Datenbanksystem
DE102010043265A1 (de) * 2009-11-06 2011-05-12 Symantec Corporation, Mountain View Systeme und Verfahren zum Verarbeiten und Verwalten von objektbezogenen Daten zur Verwendung durch mehrere Anwendungen
DE102016226338A1 (de) * 2016-12-30 2018-07-05 Bundesdruckerei Gmbh Bitsequenzbasiertes Datenklassifikationssystem

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102017208084A1 (de) * 2017-05-12 2018-11-15 Bundesdruckerei Gmbh Datenbank mit feldbezogenen Zeitstempeln

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE19627472A1 (de) * 1996-07-08 1998-01-15 Ser Systeme Ag Datenbanksystem
DE102010043265A1 (de) * 2009-11-06 2011-05-12 Symantec Corporation, Mountain View Systeme und Verfahren zum Verarbeiten und Verwalten von objektbezogenen Daten zur Verwendung durch mehrere Anwendungen
DE102016226338A1 (de) * 2016-12-30 2018-07-05 Bundesdruckerei Gmbh Bitsequenzbasiertes Datenklassifikationssystem

Also Published As

Publication number Publication date
EP3948578A1 (fr) 2022-02-09
DE102019108858A1 (de) 2020-10-08

Similar Documents

Publication Publication Date Title
DE112020002600T5 (de) Entdecken einer semantischen bedeutung von datenfeldern anhand von profildaten der datenfelder
DE60121231T2 (de) Datenverarbeitungsverfahren
EP3948577B1 (fr) Apprentissage automatisé par machine sur la base de données stockées
EP1779271B1 (fr) Dispositif d'analyse vocale et textuelle et procede correspondant
EP2100234B1 (fr) Systeme et procede pour la navigation multidimensionnelle commandee par l'utilisateur et/ou agregation thematique et/ou surveillance de donnees multimedia
DE102008027605B4 (de) System und Verfahren zur rechnerbasierten Analyse großer Datenmengen
DE102015009910A1 (de) Automatische Aggregation von Online-Benutzerprofilen
EP4010766A1 (fr) Procédé pour générer un jumeau numérique d'une installation ou d'un dispositif
WO2021104608A1 (fr) Procédé de génération d'une proposition d'ingénierie pour un dispositif ou une installation
EP3563261B1 (fr) Système de classification de données à base de séquence de bits
EP4002145B1 (fr) Enregistrement de données basé sur des listes destiné à la recherche de données
WO2018206304A1 (fr) Banque de données à estampilles temporelles associées au champ
EP4123517A1 (fr) Intégration des modèles d'apprentissage automatique distribué
EP3948578A1 (fr) Apprentissage par machine sur la base de définitions de déclencheurs
WO2025003524A1 (fr) Procédé mis en œuvre par ordinateur destiné à fournir une séquence de sortie, en particulier basée au moins en partie sur du texte
WO2020201248A1 (fr) Index s'étendant à plusieurs banques de données sur un système réparti de banques de données
Hui et al. Analysis of decision tree classification algorithm based on attribute reduction and application in criminal behavior
Li et al. Research on abnormal monitoring of vehicle traffic network data based on support vector machine
EP2423830A1 (fr) Procédé de recherche dans une multitude d'ensembles de données et machine de recherche
Liu et al. Automotive prospective technology mining method based on big data content analysis
DE102022125396A1 (de) Vorhersage eines wiederholten Auftretens einer Funktionsstörung
WO2026057878A1 (fr) Procédé de recherche d'objets dans une base de données
CN110879829A (zh) 一种知识产权大数据业务智能系统
DE202009018584U1 (de) Technisches Klassifikationssystem
DE102008007035A1 (de) Verfahren und Vorrichtung zur Bildähnlichkeitsanalyse

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 20719125

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2020719125

Country of ref document: EP

Effective date: 20211104