Method for repeatedly splitting text with precision

By leveraging the collaborative work of CPUs and GPUs within an AI infrastructure, combined with generative large language models and rule analysis, the problem of information loss in text segmentation is solved, achieving efficient and accurate text segmentation and feature combination comparison.

CN122228502APending Publication Date: 2026-06-16ABP SERVICES LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202480073609.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2023-11-20
Filing Date
2024-11-18
Publication Date
2026-06-16

AI Technical Summary

Technical Problem

Existing technologies struggle to repeatedly and accurately split the interrelated feature combinations in complex texts and reliably compare them with other text feature combinations, leading to incorrect splitting or information loss.

Method used

By leveraging the collaborative work of CPU and GPU servers within an AI infrastructure, combining generative AI large language models and rule-based analytics, text is split while preserving contextual information, parallel processing is performed using GPUs, and model accuracy is refined by incorporating user input.

Benefits of technology

It achieves accurate segmentation of repeated text, preserves contextual information, reduces hardware resource investment, and improves processing efficiency and result accuracy.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122228502A_ABST
    Figure CN122228502A_ABST
Patent Text Reader

Abstract

The invention relates to a method and a computer-implemented method for splitting text accurately into semantically correctly separated parts in an infrastructure or hardware for artificial intelligence, the splitting of the text being performed by a trained large language model on one or more graphics processors or graphics processor servers understanding natural language understanding and the processing of the natural language by stored rules being used for improving the splitting of the individual text parts and the individual text parts being transmitted in a machine-readable format to a management program for storing and / or displaying the split text parts and / or the relationship to the assigned text source. Here, the steps of high computing power are optimally distributed between different hardware components or servers in order to be able to guarantee fast processing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a method for accurately splitting repetitive text into semantically correct parts in infrastructure or hardware used for artificial intelligence. Background Technology

[0002] Analyzing complex texts (especially patent specifications, patent claims, technical specifications for machinery or equipment, or invention applications) requires a significant amount of time. Extensive research exists in the prior art on classifying and keyword indexing long and complex texts, both for human readers and through computer-based methods to make them comparable. The main problem here is that the disclosure in a text can contain different embodiments. These embodiments do not necessarily need to be related and can each be composed of different combinations of features. While some of these features may overlap, they still possess individual, distinct essential features and therefore do not constitute a common feature combination with other embodiments in the disclosure.

[0003] Features and their relation to the text are crucial in determining the scope of patent protection or the disclosed combination of features and comparing them with other disclosures or combinations of patent features. For example, features in patent claims must be clearly and accurately drafted to ensure demarcation relative to other technologies.

[0004] A feature combination is understood as an arrangement of different features or characteristics that collectively describe a particular technical or unique technical aspect of the invention. Feature combinations are typically included in the patent claims that define the legally binding scope of protection of the patent.

[0005] Feature combinations play a decisive role in defining the boundaries of a licensed invention relative to prior art. Clear and unambiguous drafting of feature combinations is important in order to define the scope of patent protection and ensure that third parties cannot use similar techniques or methods with the defined features without the patentee's permission.

[0006] US11714839B2 discloses a method for automating patent claim analysis. Here, the claims are analyzed word-for-word to identify word sequences as scope concept phrases. Additional scope concept phrases are generated using a thesaurus in a thesaurus and then automatically categorized into the product specification to create a graphical representation of the textual scope of each scope concept phrase.

[0007] Even newer large language models with natural language understanding (such as ChatGPT) cannot reliably, repeatedly, accurately, and reproducibly extract interrelated feature combinations from the analysis of longer texts and compare these feature combinations with other feature combinations from other longer texts.

[0008] Here, the particular problem lies in the fact that rule-based splitting, such as at commas or conjunctions, or connecting words (e.g., the word "and"), leads to the incorrect splitting of text portions, or the split text portions losing their original meaning in themselves. Similarly, there is the problem of illusion or the fabrication of text that does not exist in the text source. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method by which users can easily and accurately split text into semantically correct parts and optimally distribute the computationally intensive steps among different hardware components or servers to ensure fast execution.

[0010] The objective is achieved by a method according to the claim.

[0011] The method according to the invention is used to accurately segment text repetitions into semantically correct parts within an infrastructure for artificial intelligence, said infrastructure comprising one or more storage systems and one or more central processing unit servers (CPU servers) and one or more graphics processing unit servers (GPU servers), said method comprising the following: - In the management program, the text to be split is retrieved from the storage system via a central processing unit server (CPU server). - The text to be split is transmitted in a machine-readable format to a large language model with generative artificial intelligence running on one or more graphics processing unit (GPU) servers. - Text is segmented by understanding natural language through one or more large language models trained on said graphics processor servers (GPU servers). - Store the text source about the split text portions on one or more of the aforementioned storage systems. The text portions of the text, segmented by the large language model, and the original text source are transmitted to the program in a machine-readable format. The program, in addition to processing natural language, also uses rules stored on the storage system to improve the segmentation of the individual text portions. - The individual text portions are transmitted in a machine-readable format to a management program running on a central processing unit (CPU server), and - Store and / or display the split text portions and / or their relationship to the associated text source.

[0012] In this approach of the first embodiment, it is advantageous to distribute the high-computing-power steps among different servers or server types in such a way that, on the one hand, the advantages of different servers (i.e., CPU servers and GPU servers) can be utilized (e.g., parallel processing), and on the other hand, the contextual information about the text source is not lost and illusions are suppressed.

[0013] In a second embodiment, the computer-implemented method according to the present invention is used to accurately segment text repetitions into semantically correct parts in hardware for artificial intelligence, the method comprising the following steps: - In the management program (6), a central processing unit (22) retrieves the stored text (5) to be split from a storage system (2). - The text to be split (5) is transmitted in a machine-readable format to a large language model (11) with generative artificial intelligence running on one or more graphics processors (23). - Text (5) is split by understanding natural language through a large language model (11) trained on one or more graphics processors (23). - The text source (12) of the split text portion (13) is stored on one or more of the storage systems (2). - The text portions (13) of the text (5) split by the large language model (11) and the text source (12) are transmitted to the program (15) in a machine-readable format. The program, in addition to processing natural language, also uses rules (16) stored on a storage system (2) to improve the splitting of the individual text portions (13). - The individual text portions (13) are transmitted in a machine-readable format to the management program (6) running on the central processing unit (22), and - Store and / or display the split text portions and / or their relationship to the associated text source (19).

[0014] In this alternative implementation, the advantage is that the high-computational steps are distributed among different hardware components in such a way that, on the one hand, the advantages of different processor types (i.e., CPU and GPU) can be utilized, and on the other hand, contextual information about the text source is not lost and illusions are suppressed. Additionally, less hardware investment is required.

[0015] Furthermore, the present invention relates to a system for data processing, the system comprising a plurality of processors, all processors being adapted or configured such that the processors perform the steps described above in the method. Advantageously, the interaction between the various programs and hardware components enables the efficient and precise separation of text portions.

[0016] Additionally, the present invention relates to a computer program product comprising instructions that, when executed by a computer, cause the computer to perform the method (as shown above). Advantageously, the various instructions for controlling the method steps are controlled by the computer program product, or management program, on different hardware components, enabling further processing.

[0017] Furthermore, it is advantageous to transmit the split text portions and the text source in a machine-readable format to a program for rule-based analysis in order to improve the splitting. The program then performs rule-based analysis on a central processing unit server according to rules stored on the storage system, thereby further improving the splitting effect while conserving hardware resources. Here, for example, rules can be used that specify at least the separation from subsequent text portions before, after, or after a combination of words, or that certain characters such as formulas, symbols, variables, figure tags, bracket expressions, or formatting information should not be considered when splitting using a large language model.

[0018] Another advantageous implementation is to consider the stored text source as a relationship between text parts and text sources, so as to ensure the same display of the split text parts during further retrieval. This allows for repeated and precise queries without requiring high-intensity computational power on hardware components.

[0019] Another alternative implementation is characterized by further processing of the segmented text portions using a large language model running on one or more graphics processing units (GPUs) to generate concepts, keywords, or superordinate terms, and storing and / or displaying these concepts, keywords, or superordinate terms in the management program. This improves the relevance of subsequent automated queries. Additionally, these concepts, keywords, and superordinate terms can assist the user when composing queries, but also when further using the large language model to create text (as claimed, specifications, or examination opinion responses).

[0020] Similarly, it is advantageous to utilize artificial intelligence for natural language processing (NLP) running on one or more graphics processor servers to further process the segmented text portions for generating concepts, keywords, or hypernyms, and to store and / or display these concepts, keywords, or hypernyms in the management program. Using a specific trained NLP system leads to similar but different results, resulting in improved quality of the results and potentially forming the basis for automating the improvement of prompts and instructions.

[0021] An advantageous improvement to the method is to train the large language model using data accessible on one or more of the storage systems via the management program. This allows for automated improvement of prompts and instructions using information already associated with the context by human users within the management program, without additional time expenditure for training.

[0022] Particularly effective is the training of the large language model using patent publications, especially patent claims, that are accessible on storage via the Internet.

[0023] Another advantageous implementation is to refine (also known as fine-tuning) the accuracy of a large language model with generative artificial intelligence using prompts, instructions in natural language, or scripts and example text, when using one or more GPU servers. This allows for more efficient and faster work with a smaller language model. Training a language model for a specific application domain (e.g., claims) on a GPU server requires fewer lexical units.

[0024] A particularly advantageous aspect is that, when using a large language model with generative artificial intelligence, the segmented text portions are stored and / or displayed in a graph within the management program in a manner that compares them with other text portions accessible via the management program on one or more of the storage systems. Thus, while a machine-readable database exists to improve the method, feature analysis and feature comparison can also be performed.

[0025] Therefore, the method can be further improved by calculating the consistency of two or more text portions to be compared using one or more graphics processing units (GPUs) or GPUs, and the consistency is used as a basis for assessing the degree of association with respect to storage or representation. Parallel processing on the GPU can significantly improve training and inference speed.

[0026] An effective improvement to the method could advantageously be that the separation points between the split text segments can be altered via user input within the management program, and these alterations can be used to further refine the accuracy of the large language model with generative artificial intelligence, particularly via prompts, instructions, or scripts. Thus, the large language model can be further trained to some extent "incidentally" through input within the management program.

[0027] It is also advantageous that the comparison results between the segmented text parts can be modified or evaluated via user input within the management program, and these user inputs can be used to further refine the accuracy of the large language model with generative artificial intelligence, particularly via prompts, instructions, or scripts. Thus, the large language model can be further trained within the management program, for example, through decision boxes, evaluation functions, or selection lists available to the user.

[0028] Large Language Models (LLMs) refer to large-scale language models that are based on machine learning and capable of understanding, generating, and responding to natural language. Models such as GPT-3 (Generative Pretrained Transformer 3, or LLaMA) are examples of LLMs.

[0029] The term (NLP) is understood in this application as: Natural Language Processing (NLP) is a branch of information technology, artificial intelligence, and linguistics that focuses on the interaction between computers and human (natural) language. In this sense, NLP is related to the field of human-computer interaction, particularly in the understanding of natural language, enabling computers to derive meaning from human or natural language input.

[0030] Many NLP systems use ontology to support the execution of NLP tasks. An ontology is a representation of knowledge. In the context of NLP, a semantic ontology is a representation of knowledge about relationships between semantic concepts. Ontologies are created by humans, typically by professionals, and are not a perfect representation of all available knowledge. Ontologies are usually highly targeted to a specific branch of a domain and often reflect the knowledge level or level of detail of the compiler. Ontologies are generally task-oriented, meaning they serve a practical purpose in managing information or physical entities, and their design reflects the tasks required by their terminology system.

[0031] In the context of artificial intelligence (AI) (and so in this application), the term "lexicon" often refers to the basic unit of text processed by AI models. Here, we have two common meanings related to AI: A lexical unit refers to a unit such as a word or a subunit of a word. For example, the sentence "Die Katze läuftschnell" can be divided into lexical units, each of which is a single word: "Die", "Katze", "läuft", and "schnell". This lexicalization enables AI to understand and analyze text. Different large language models can only work with a limited number of lexical units. For example, the Llama2 language model has a maximum of 4096 lexical units, or the CodeLlama language model has a maximum of 16384 lexical units.

[0032] In the field of artificial intelligence (AI) (and here as well), the term "cue word" describes instructions or text snippets given to an AI model in order to obtain a desired response or output. Cue words are like task requirements that drive the model to generate or understand certain information.

[0033] Features are, for example, technically and / or functionally related parts of a larger combination of features in a patent claim, invention, technical specification, or product. These parts may also consist of single words, sentence components, conjunctions, or connecting words, but they cannot be further divided in the absence of technical relevance or when semantics are lost (in themselves).

[0034] This application understands feature analysis or feature comparison as the study and identification of specific features or characteristics, or combinations of features, in a granted invention, implementation of a product, invention application, or technical disclosure. Feature analysis plays a decisive role in determining the scope of patent protection and in demarcating it relative to other technologies. The analysis aims to identify key features or combinations of features that embody uniqueness and inventiveness. The identified features or combinations of features can be compared with prior art to confirm whether similar features have been disclosed or whether the invention possesses innovative points. Here, feature analysis helps determine the scope of patent protection. Feature analysis is an important step for patent examiners, patent agents, inventors, and technicians to gain a deeper understanding of granted inventions and clarify their legal implications. Furthermore, feature comparison provides a concise overview of separately disclosed combinations of features in two or more documents to be studied.

[0035] GPU stands for "Graphics Processing Unit" and is a specialized hardware component optimized for processing graphics and parallel computing, particularly for use in the fields of machine learning and artificial intelligence.

[0036] In more detail, this can be understood as fine-tuning, which is a method for learning transfer in pre-trained artificial neural networks. The base model is a machine learning model pre-trained on a large amount of data, which is then fine-tuned to adapt to specific tasks or applications and optimize the model. Attached Figure Description

[0037] To better understand this invention, the accompanying drawings will be used to explain it in detail.

[0038] In the highly simplified diagrams: Figure 1 A schematic diagram of the infrastructure for performing the method according to embodiment 1 is shown; Figure 2 A schematic diagram is shown for performing the method according to embodiment 2; Figure 3 A diagrammatic representation of the split text, featuring comparative features and concepts, is presented. Detailed Implementation

[0039] First, it should be noted that in different described embodiments, the same components are given the same reference numerals or the same component names, and the disclosure contained throughout the specification can be applied semantically to the same components having the same reference numerals or the same component names. The location descriptions selected in the specification, such as upper, lower, lateral, etc., also refer to the directly described and illustrated figures, and these location descriptions are applied semantically to the new location when the location changes.

[0040] exist Figure 1 The diagram shows an infrastructure for performing the method according to Embodiment 1, and the method steps are explained with the aid of this diagram.

[0041] To accurately break down repetitive text into semantically correct parts, an infrastructure 1 for artificial intelligence exists, which has one or more centralized or distributed storage systems 2, such as long-term storage (e.g., hard drives, cloud storage, storage systems, or the like) and / or short-term storage (e.g., random access memory). Additionally, one or more central processing unit servers 3 (CPU servers) and one or more graphics processing unit servers 4 (GPU servers) are provided.

[0042] The text 5 to be split, stored in storage system 2, is retrieved from a storage system via a management program 6, which has a database 7 with diagram 8 and access to a central processing unit server 3. Data and relationships 9 regarding files (such as text 5) saved on storage system 2 are entered into diagram 8.

[0043] The text 5 to be split is transmitted 10 in a machine-readable format to a large language model 11 with generative artificial intelligence running on one or more graphics processor servers 4. Optionally, the graphics processor server 4 may be equipped with a central or decentralized storage system 2.

[0044] The text 5 is then segmented by understanding natural language through a large language model 11 trained on one or more graphics processor servers 4. The text portions 13 of the text source 12, or text 5, with respect to the segmented text portions 13 are stored on one or more of the storage systems 2.

[0045] The text portion 13, split by the large language model 11, and the text source 12 of text 5 are further transmitted 14 to program 15 in a machine-readable format. Program 15, in addition to processing natural language, also uses rules 16 stored on storage system 2 to improve the splitting of the individual text portions. These rules 16 may include, for example, certain sentence sequences such as “characterized by” or certain other separating conditions or operators such as semicolons.

[0046] The individual text portions are transmitted in a machine-readable format to the management program 6 running on the central processing unit server 3.

[0047] Store the split text portions 13 on the storage system 2 and / or display them on the user interface 18 of the management program 6 and / or their relationship to the associated text source.

[0048] The large language model 11 can be based on models such as LLaMA, Granite, or other commercially available models that are partially personalized or specifically trained for the method according to the invention. The large language model 11 runs on a GPU server. The graphics processing unit (GPU) is optimized for parallel processing, meaning that it can perform multiple computations simultaneously. The large language model 11 has parallel processing capabilities because a large number of computations must be performed when processing large amounts of text. This significantly speeds up the training and inference processes. The large language model 11 performs large-scale matrix operations. Graphics processing units (GPUs) are particularly efficient at performing matrix operations, resulting in faster computation.

[0049] Large language models can be trained on massive datasets, and access to this data requires rapid data processing capabilities. The high storage bandwidth and high-power computing capabilities of graphics processing units (GPUs) enable efficient processing of large data volumes.

[0050] Graphics processors (GPUs) can be easily integrated into server clusters or cluster architectures such as GPU servers, which enables high scalability for using large language models 11 in large infrastructures 1.

[0051] The learning model associated with the large language model 11 benefits from the customized computing power of the GPU. GPUs are designed to perform the computations necessary for training and inference neural networks more efficiently than traditional central processing units (CPUs). Overall, GPU servers enable more efficient and faster processing of larger language models such as LLMs, which is crucial to their performance.

[0052] Rule-based text analysis can also run on a central processing unit server (CPU server) because rule-based systems based on predefined rules typically work serially, i.e., sequentially, unlike a large number of machine learning models that operate in parallel.

[0053] Rule-based systems, such as Rule 16, involve complex decision structures and require precise control. CPUs can efficiently handle such logical operations. Compared to demanding machine learning models, rule-based text analysis requires less parallel processing and specific computations. CPU performance is sufficient to accomplish these tasks.

[0054] Rule 16 can be easily adapted to specific requirements by modifying or adding rules. The CPU benefits from its flexibility in efficiently executing different rules 16 without relying on pre-training.

[0055] Infrastructure 1 can be run on a local area network independently of the Internet in order to meet security requirements specifically in the areas of patent application and patent search.

[0056] To improve the transmission of the split text portion 13 of text 5 and the text source 12, the split text portion 13 can be transmitted to the rule-based program 15 in a machine-readable format, and the program 15 performs rule-based analysis on the central processing unit server 3 according to the rules 16 stored on the storage system 2.

[0057] The stored text source 12 can be considered in the form of a text portion and text source relationship 19 to ensure that the split text portion 13 is displayed in the same way during further retrieval.

[0058] Further processing of the split text portion 13 can be performed using a large language model 11 that can run on one or more graphics processor servers 4, in order to generate concepts 20, keywords, or hypernyms. Their storage and / or display can be performed on the user interface 18 in the management program 6.

[0059] Alternatively, further processing of the split text portions 13 can be performed using artificial intelligence for natural language processing (NLP) running on one or more of the graphics processor servers 4 to generate concepts 20, keywords, or hypernyms. Their storage and / or display can then be performed on the user interface 18 of the management program 16.

[0060] The large language model 11 can be trained using data 21 accessible on one or more of the storage systems 2 via management program 6 or database 7.

[0061] However, the large language model 11 can also be trained using data from patent publications (especially patent claims) that are accessible on storage via the Internet.

[0062] To refine the accuracy of the large language model 11 with generative artificial intelligence, or to fine-tune the large language model, prompts, instructions or scripts in natural language and example text can be used when using one or more graphics processors 23 or graphics processor servers 4.

[0063] Multiple graphics processors 23 can be integrated into the graphics processor server 4.

[0064] Multiple central processing units 22 can be integrated into the central processing unit server 3.

[0065] When using a large language model 11 with generative artificial intelligence, the split text portions 13 can be stored and / or displayed in a graph 8 in the management program 6 in a manner that compares them with other text portions 13 accessible on one or more of the storage systems 2 via the management program 6.

[0066] The consistency of two or more text portions 13 to be compared is calculated using one or more of the graphics processors 23 or graphics processor servers 4, and the consistency is used as a basis for the degree of relevance of the representation of the text stored or on the user interface 18 of the management program 6.

[0067] In the management program 6, the separation points 25 between the split text portions 13 can be changed via user input, and these changes can be used to further refine the accuracy of the large language model 11 with generative artificial intelligence, particularly via prompts, instructions, or scripts.

[0068] In the management program 6, the comparison results 26 between text segments 13 split from different texts 5 can be changed or evaluated via user input, and these user inputs can be used to further refine the accuracy of the large language model 11 with generative artificial intelligence, particularly via prompts, instructions or scripts.

[0069] The system 27 for data processing may include multiple processors adapted or configured to perform the steps of the method as explained above.

[0070] A computer program product 28 that can run on infrastructure 1 or hardware 24 includes instructions that cause the computer to perform the method (as shown above) when the program is executed by the computer.

[0071] exist Figure 2 The middle shows Figure 1 Another, and if necessary, independent implementation method, which uses the same method as described above for the same part. Figure 1 The same reference numerals or component names are used in the accompanying drawings. To avoid unnecessary repetition, refer to or refer to the preceding text. Figure 1See the detailed explanation in the document. Figure 2 The form is the implementation method of the computer-implemented method and Figure 1 The difference in the implementation method is that all the components and programs used to accurately break down text repetitions into semantically correct parts run in hardware 24 for artificial intelligence, which has at least one central processing unit 22 and at least one graphics processing unit 23. Here, [the following will be discussed]... Figure 1 The described running steps on the central processing server 3 in the method correspond to the following on at least one central processing unit 22: Figure 1 The described steps are performed meaningfully.

[0072] Here, will Figure 1 The described steps in the method, which are run on the graphics processor server 4, correspond to the following on at least one graphics processor 23: Figure 1 The described steps are performed meaningfully.

[0073] exist Figure 3 Another, and if necessary, independent, implementation is shown, with the same parts used in the preceding examples. Figure 1 The same reference numerals or component names are used in the accompanying drawings.

[0074] Figure 3 A diagram of the split text 13 is shown in graphical form, which has feature comparisons and concepts 20, as can be displayed on the user interface 18 of the management program 6.

[0075] Of course, other forms of representation, such as lists or graphs, can be used instead of charts.

[0076] Here, the split section 25 is symbolically represented by a line break in the diagram. Users can move or manipulate the smaller portions of the split text section 13 in such a way that they can be incorporated into subsequent or previous text sections 13 or features.

[0077] Display concept 20 for individual text portions 13 or features, for example, in the same chart row but in an adjacent chart column, or as a floating pop-up or annotation box for text portion 13.

[0078] It can integrate other functions, such as evaluation and annotation of comparison result 26.

[0079] Optionally, the text source 12 about the text section 13 can be displayed.

[0080] Furthermore, in a separate embodiment where necessary, an input box 29 may be provided on the user interface 18 of the management program 6 for entering prompts and / or other instructions or commands, or for inputting examples used to train the large language model 11. Thus, the comparison results 26, the separation of text portions 13, concepts 20, or rules 16 can be improved through interaction with the user.

[0081] Furthermore, in standalone embodiments where necessary, the large language model 11 can be used to automatically summarize text, text portions, selected files, or data 21.

[0082] Furthermore, in standalone embodiments where necessary, the large language model 11 can be used to automatically create text, such as in a patent application, based on multiple selected split text segments 13 and / or concepts 20.

[0083] Furthermore, in a separate embodiment where necessary, management procedure 6 may display or compare the relationship 26 between the split text portion 13 and products, invention filings, documents, foreign intellectual property, patent publications, or Internet search results already entered in pipeline procedure 6, and mark overlaps or similarities.

[0084] The embodiments illustrate possible variations of implementation. It should be noted that the invention is not limited to the embodiments specifically shown. Rather, different combinations of the various embodiments are possible, and such variations are based on the teachings of the specific invention and are within the capabilities of those skilled in the art.

[0085] The scope of protection is defined by the claims. However, the description and drawings are considered in order to interpret the claims. Individual features or combinations of features in the different embodiments shown and described can be independent, inventive solutions in themselves. The task of finding a solution based on the independent invention can be derived from the description.

[0086] The full description of the numerical range in the specific specification should be understood as follows: the numerical range together includes any and all sub-ranges that arise therefrom. For example, the description of 1 to 10 should be understood as including all sub-ranges starting from the lower limit of 1 and the upper limit of 10. That is, all sub-ranges that start with a lower limit of 1 or greater and end with an upper limit of 10 or less, such as 1 to 1.7, or 3.2 to 8.1, or 5.5 to 10.

[0087] The regulations stipulate that, in order to better understand the structure, the components are shown partially out of proportion and / or enlarged and / or reduced.

[0088] List of reference numerals

[0089] 1. Infrastructure

[0090] 2 Storage System

[0091] 3 Central Processing Unit Servers

[0092] 4 Graphics Processor Servers

[0093] 5 texts

[0094] 6 Management Procedures

[0095] 7 databases

[0096] 8 charts

[0097] 9 Contact

[0098] 10 transmission

[0099] 11 major language models

[0100] 12 text sources

[0101] 13 Text Section

[0102] 14 Transmission

[0103] 15 Programs

[0104] 16 rules

[0105] 17 Transmissions

[0106] 18 User Interfaces

[0107] 19 Relationship

[0108] 20 concepts

[0109] 21 Data

[0110] 22 Central Processing Units

[0111] 23 graphics processor

[0112] 24 Hardware

[0113] 25-point part

[0114] 26 Comparison Results

[0115] 27 Systems for Data Processing

[0116] 28 Computer Program Products

[0117] 29 Input boxes

Claims

1. A method for precisely splitting text repetitions into semantically correct segments within an infrastructure (1) for artificial intelligence, said infrastructure (1) comprising one or more storage systems (2) and one or more central processing unit servers (3) and one or more graphics processing unit servers (4), said method comprising: - In the management program (6), the stored text (5) to be split is retrieved from the storage system (2) via a central processing unit server (3). - The text to be split (5) is transmitted in a machine-readable format to a large language model (11) with generative artificial intelligence running on one or more graphics processor servers (4). - Text is segmented by understanding natural language through a large language model (11) trained on one or more graphics processor servers (4). - The text source (12) of the split text portion (13) is stored on one or more of the storage systems (2). - The text portions (13) of the text (5) split by the large language model (11) and the text source (12) are transmitted to the program (15) in a machine-readable format. The program, in addition to processing natural language, also uses rules (16) stored on a storage system (2) to improve the splitting of the individual text portions (13). - The individual text portions (13) are transmitted in a machine-readable format to the management program (6) running on the central processing unit server (3), and - Store and / or display the split text portions and / or their relationship to the associated text source (19).

2. A computer-implemented method for accurately segmenting text repetitions into semantically correct parts in hardware (24) for artificial intelligence, the method comprising the following steps: - In the management program (6), the stored text (5) to be split is retrieved from the storage system (2) via a central processing unit (22). - The text to be split (5) is transmitted in a machine-readable format to a large language model (11) with generative artificial intelligence running on one or more graphics processors (23). - Text (5) is split by understanding natural language through a large language model (11) trained on one or more graphics processors (23). - The text source (12) of the split text portion (13) is stored on one or more of the storage systems (2). - The text portions (13) of the text (5) split by the large language model (11) and the text source (12) are transmitted to the program (15) in a machine-readable format. The program, in addition to processing natural language, also uses rules (16) stored on a storage system (2) to improve the splitting of the individual text portions (13). - The individual text portions (13) are transmitted in a machine-readable format to the management program (6) running on the central processing unit (22), and - Store and / or display the split text portions and / or their relationship to the associated text source (19).

3. The method according to claim 1 or 2, wherein, The split text portion (13) of the text (5) and the text source (12) are transmitted in a machine-readable format to a program (15) for rule-based analysis in order to improve the splitting, and the program (15) performs rule-based analysis on a central processing unit server (3) or a central processing unit (22) according to rules (16) stored on a storage system (2).

4. The method according to any one of claims 1 to 3, wherein, The stored text source (12) is considered in the form of a text part and text source relationship (19) to ensure the same display of the split text part (13) when further retrieved by the management program (6).

5. The method according to any one of claims 1 to 4, characterized in that, Further processing of the split text portions (13) will be carried out using a large language model (11) running on one or more graphics processor servers (4) or graphics processors (23) to generate concepts (20), keywords or superordinate terms, and to store and / or display the concepts, keywords or superordinate terms in the management program (6).

6. The method according to any one of claims 1 to 4, wherein, Further processing of the split text segments (13) will be carried out using artificial intelligence for natural language processing (NLP) running on one or more graphics processor servers (4) or graphics processors (23) to generate concepts (20), keywords or superordinate terms, and to store and / or display the concepts, keywords or superordinate terms in the management program (6).

7. The method according to any one of claims 1 to 6, wherein, The large language model (11) is trained using data (21) accessible on one or more of the storage systems (2) via the management program (6).

8. The method according to any one of claims 1 to 7, wherein, The large language model (11) is trained using patent publications, particularly patent claims, that are accessible on Internet storage via the Internet.

9. The method according to any one of claims 1 to 8, wherein, The precision of a large language model (11) with generative artificial intelligence is refined via prompts, instructions in natural language or scripts and example text when using one or more graphics processor servers.

10. The method according to any one of claims 1 to 9, wherein, In the case of using a large language model (11) with generative artificial intelligence, the split text portions are stored and / or displayed in a graph (8) in the management program (6) in a manner that compares them with other text portions (13) accessible on one or more of the storage systems (2) via the management program (6).

11. The method according to any one of claims 1 to 10, wherein, The consistency of two or more text portions (13) to be compared is calculated by using one or more graphics processor servers (4) or graphics processors (23), and the consistency is used as a basis for the degree of relevance with respect to storage or representation.

12. The method according to any one of claims 1 to 11, wherein, The management program (6) can change the separation part (25) between the split text parts (13) via user input, and use these changes to further refine the accuracy of the large language model (11) with generative artificial intelligence, especially via prompts, instructions or scripts.

13. The method according to any one of claims 1 to 12, wherein, The management program (6) can change or evaluate the comparison results (26) between the split text segments (13) via user input, and use these user inputs to further refine the accuracy of the large language model (11) with generative artificial intelligence, particularly via prompts, instructions or scripts.

14. A system for data processing, the system comprising a plurality of processors adapted or configured such that the processors perform the steps of the method according to any one of claims 1 to 13.

15. A computer program product comprising instructions that, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 13.

Citation Information

Patent Citations

  • Apparatus and method for automated and assisted patent claim mapping and expense planning

    US11714839B2