Method for generating a patent application body

Abstract

The invention relates to a method for generating a patent application body executed by a computer, having the following steps: a. providing an input interface for a user, which input interface is designed to input features of an invention; b. preparing the features input via the input interface in the form of a computer-readable graph comprising a structure consisting of nodes, edges and weightings; c. reading the graph into an artificial intelligence, the artificial intelligence being configured to process the structure of the graph and those features stored in the graph; d. by means of artificial intelligence, which is designed to generate human speech, and on the basis of step c., generating text, the text being generated in individual text sections associated with corresponding structure sections of the structure of the graph, and the respective generated text section being embedded in the structure of the graph as additional nodes, edges and weightings, and when generating a further text section according to step d. the newly embedded structure is taken into account as a function of the newly produced weightings by the artificial intelligence. The method proposed here ensures that patent applications are drafted efficiently and to a high standard.

Description

[0001] Method for generating a patent application body

[0002] The invention relates to a method for generating a patent application body.

[0003] With the introduction of impressively capable artificial intelligence for performing text-based tasks, a demand has arisen, particularly in the field of so-called legal tech, for an interface that allows users to utilize this technology even without extensive knowledge of artificial intelligence itself, experience in its use, or simply with minimal time investment. Patent attorneys face enormous challenges when drafting a patent application based on an invention: they must present the content technically correctly and comprehensibly while simultaneously creating a legal text that offers the applicant a suitable and clearly understandable scope of protection. In such cases, the relevant prior art, from which the invention must be clearly distinguished for patentability, may be unknown or at least not sufficiently known.Adding to this tension is the fact that patent applications are often pursued internationally, and each country of filing has its own, sometimes conflicting, requirements. Furthermore, a translation of the original text is often necessary, carrying the risk of legally relevant deviations.

[0004] Current approaches assume that an invention disclosure, such as the one often prepared by the inventor and submitted to the applicant, is sufficient as a basis for generating patent claims using artificial intelligence. Other approaches assume that a user, for example a patent attorney, drafts a set of patent claims, and that the description—that is, a continuous text including explanations, definitions, and further variations of the invention—is generated by artificial intelligence based on this set of claims. Increasingly, such approaches also attempt to have artificial intelligence recognize drawings or even create them from text, in accordance with the requirements of the patent offices.

[0005] A particular challenge, especially in connection with an LLM (Large Language Model), is striking a balance between a degree of creativity in the writing and sufficient precision and accuracy of the information. A high degree of creativity is desirable in order to find further explanations and definitions in addition to the information already provided, for example, in the invention disclosure and / or the patent claims. Artificial intelligences are designed to present the solution they deem most probable in a probability space as the irrefutably correct answer—in this respect, no different from a human approach. The (machine) answers of an artificial intelligence are not, and by their very nature cannot be, edited before being presented to the user. Rather, this must be verified by the user themselves.In some use cases, this machine response therefore represents an additional time expenditure, because information has to be checked which otherwise would not have been part of the text or could have been woven in more precisely and quickly when formulating the text with a quick search in relevant reference works or through the user's knowledge.

[0006] Based on this, the present invention aims to overcome, at least partially, the disadvantages known from the prior art. The features of the invention are defined in the independent claims, for which advantageous embodiments are shown in the dependent claims. The features of the claims can be combined in any technically meaningful way, whereby the explanations in the following description and features from the figures, which comprise supplementary embodiments of the invention, can also be used.

[0007] The invention relates to a method for generating a patent application body performed by a computer, comprising the following steps:

[0008] a. Providing an input interface for a user, which is set up for entering features of an invention;

[0009] b. Processing the features entered via the input interface into the form of a computer-readable graph comprising a structure consisting of nodes, edges and weights;

[0010] c. Reading the graph into an artificial intelligence, wherein the artificial intelligence is configured to process the structure of the graph and those features stored in the graph;

[0011] d. by means of an artificial intelligence designed to generate human speech, and based on step c. generating text,

[0012] where the text is generated in individual text sections assigned to corresponding structural sections of the graph's structure, and

[0013] wherein the respective generated text section is embedded in the structure of the graph as additional nodes, edges and weights, and when generating another text section according to step d. the newly embedded structure is taken into account by the artificial intelligence depending on the newly resulting weights.

[0014] Unless explicitly stated otherwise, ordinal numbers used in the preceding and subsequent descriptions serve solely for unambiguous differentiation and do not indicate any order or ranking of the components described. An ordinal number greater than one does not necessarily imply the presence of another such component. Furthermore, the text often refers to a fictitious invention, fictitious features, and similar terms in this context, which are used to explain this method, so that this meta-level is clearly distinguishable from the present invention.

[0015] The approach proposed here aims to leverage artificial intelligence's text generation capabilities to save time and improve quality, while simultaneously reducing the review effort compared to other approaches and, most importantly, integrating the review process into a conventional workflow. This not only fosters high user acceptance but also allows for straightforward implementation into the established processes of those who have previously prepared patent applications, including users who are unfamiliar with artificial intelligence and therefore have limited experience with its strengths and weaknesses.

[0016] It should be noted that patent applications may not be published in advance, thus ruling out open processes that require data extraction for the improvement of a software service. Furthermore, applicants often fear industrial espionage. It is therefore preferable to use a service that does not store data for any period beyond the mere processing, for example, by excluding so-called data retention. Regional legal requirements also frequently exist, stipulating that the location of the executing computers must be within the respective region. One approach therefore utilizes local AI-optimized computers, for example, as local servers. However, the performance of such local computers is currently limited. Data centers therefore represent a preferred solution at present.

[0017] The present patent application was generated with the support of such secure access to a currently well-known LLM (Large Learning Management) system. However, this requires a very complex API (Application Programming Interface) architecture with encrypted communication, ensuring that no third party can intercept this communication and misuse this secure environment for their own purposes. Furthermore, the information transmitted to the operator is only available briefly, namely exclusively during the interaction with the artificial intelligence.This means that, in the present case, the domain-specific training of the LLM (to improve its context and knowledge of domain-specific tasks) is not achieved through so-called fine-tuning, but rather through so-called few-shot prompting. This involves presenting a small number of interconnected questions (so-called prompts) for contextual information, background information, and style preferences (for example, using text examples) alongside the actual question (for example: define artificial intelligence). It should be noted that this approach is easily replaceable, not only by other ways of interacting with an LLM, but also by other types of artificial intelligence. This is not essential for the present invention, just as the artificial intelligence used, while performing essential steps, is not considered an essential element of the method.However, it should be noted that a person cannot solve the respective tasks (posed during the course of the procedure) because the requests are too extensive and complex, and the time requirement is also impossible to meet. The procedure can therefore only be carried out by or with the aid of a computer.

[0018] For a (human) user, an input device used in step a., through which the user's input interface is made available, comprises a (for example, conventional) screen, a keyboard, a (computer) mouse and / or a microphone, as well as a processing unit and software with which these are operated for human use. Thus, it is possible to input data, which is typed, for example, via the keyboard and / or mouse and / or spoken via the microphone. Preferably, input is followed by playback, which is processed into text form, for example, via a text editor on the screen, preferably using artificial intelligence in the case of input in speech form, and preferably using NLP (Natural Language Processing) in the case of speech input. Speech form thus includes both text form and the spoken word, or alternatively, speech input converted into text form.A (fictitious) invention is not precisely defined and encompasses anything, possibly even things outside the scope of patent law. A patent application must be drafted in such a way that it claims only that part of a (fictitious) invention which is permissible under patent law. However, the boundaries of this permissibility are sometimes not fully clarified in case law, and individual cases cannot simply be subsumed under national regulations and patent office practice. Linguistic understanding and human judgment also play a role. If, for example, a process is claimed that is to be executed by a computer, then this is inherently a technical process. Its technical nature is then beyond question.

[0019] In this context, it should be noted that while the present method can be used to create a patent application body for a (fictitious) patent application and is thus geared towards (fictitious) technical inventions, it can also be used for other purposes. For example, it can be used for so-called bench marking in the reengineering of competitor products (e.g., to detect patent infringements) to create a protocol for this, or an instruction manual for a device and / or software.

[0020] Features are the fundamental building blocks of a hypothetical invention. For example, such (hypothetical) features can be physical, as is often the case in mechanical-physical inventions, or they can be concepts, interactions, and other relationships, although these are not always clearly distinguishable. In a simple, but not exclusively used here, form of a feature, it is a word feature, as is listed, for example (and usually), in the list of reference numerals of the present patent application. In a somewhat more complex form, a feature is a sentence or clause, for example, a relative clause or two main clauses joined by a conjunction.

[0021] It is important to note that the user does not need to be aware of the nature of the attribute; rather, assistance is provided if the attribute or its components should be entered separately, as described below, or can be entered entirely in the usual way via the input interface. In the latter case, the input is preferably broken down into separate attributes within the process, as described further below.

[0022] It should be noted that in one embodiment, an input interface is an API (Application Programming Interface), whereby the features do not necessarily have to originate from a human source, for example, they may have been generated by an artificial intelligence based on a specification (such as a fictitious invention disclosure). For the method, it is only important that such features are in language form. These can be traced back to a human, at least indirectly (for example, with an artificial intelligence as a fictitious inventor that has solved a task posed by a human). For example, there is a fictitious inventor and a fictitious human patent attorney, the latter providing the input in language form when using the method described herein, as in the present patent application.

[0023] In step b., these provided (fictitious) features are incorporated into a graph, or a graph is created from them. According to the general definition in information technology, a graph is a structure consisting of nodes and edges, where the edges and / or nodes are assigned a weight (in an exceptional case, all existing weights are set to 1 [one], thus effectively eliminating any weighting). The nodes represent objects, and the edges indicate connections between these objects. Graphs are used to represent relationships and connections in networks, such as social networks or road networks. A graph can be directed in one embodiment and undirected in another; that is, the connections either have a direction or no direction.

[0024] The fundamental principle here is that the graph represents a (simply) algorithmically processable, i.e., computer-readable, form of data. This means that the content, i.e., the meaning, is not necessarily understandable to a computer (as is the case, for example, with source code). Rather, the fictitious features and their meanings are handled in containers. Such a container is a node or an edge in the graph. In a simple embodiment, the fictitious features are embedded algorithmically, preferably deterministically, into a graph. In one embodiment, this is done using NLP (Natural Language Processing). Preferably, the user receives assistance via the input interface or is forced to enter the fictitious features in such a structured manner that they can be integrated into a graph using a simple algorithm.

[0025] For example, a single text input field is provided for each of the or a small number of fictitious features.

[0026] For example, two fictitious word features (e.g., screw and nut) are represented as nodes and their semantic relationship (e.g., an assembly) as edges in the graph. In one embodiment, the assembly is extended by a semantic property, for example, by specifying that the two are screwed together, which in one embodiment of the graph is represented as an additional node or, alternatively, as a property of the edge.

[0027] A graphical user interface (GUI) provides text fields that can be filled with text. Multiple, and potentially any number, of text fields can be created (for example, via user input or automatically). Users are free to create a separate text field for each of the listed steps of a hypothetical process, or for smaller sub-units thereof. It is also possible to include two, several, or all steps of a hypothetical process in a single text field.

[0028] The text fields already have a fixed relationship defined in the background, for example, as the title of the patent application body to be generated, transitions between text sections, dependencies, claim types (independent, dependent, subordinate), and several others. It is important that the created text field has a fixed relationship to one or more other text fields, which form part of the general description and / or figure description in the patent application body to be generated. This text field is thus already algorithmically linked in the graph (via an edge, possibly with one or more nodes), for example, to the fictitious description section, in order to precisely define the content of the text field, provide an alternative formulation for it, and / or offer an example of the often more abstract representation in the created text field, which is part of a fictitious patent claim.It should be noted that while such a linked text field can preferably be edited by a user or filled with language without further support, it is preferably filled by artificial intelligence in step d.

[0029] The graph prepared in step c. is now provided to an artificial intelligence (AI) designed to process, and especially in this application to generate, human language. This means that in step c., the AI ​​is not exclusively, and preferably not, tasked with extracting the fictional features from a fictional text entered by the user, nor with recognizing syntax and semantics. Rather, this work is (at least primarily) already completed through the interplay of the user's input and the support provided by the input interface, which pre-structures this input into (fictional) features. This leverages the fact that the user can input this information without major difficulty, thanks to their intrinsic knowledge of the order of the fictional features.While it may require a brief learning curve, some familiarization, and a little practice from the user to create this structure simultaneously with the input, this is intuitive and corresponds to the structured, cognitive approach of a patent attorney. Moreover, this cognitive effort is facilitated by visual feedback (via an input interface) and suggestions for shaping the input, combined with simplified modification options and great flexibility in the input.

[0030] This is shown in an exemplary representation of an input mask in Fig. 2 and will be explained in more detail later using this example. This input interface allows both inexperienced and experienced users to convert features of a (fictitious) invention into a structured claim form in a shorter time compared to pure text input.

[0031] To save time, further aspects are helpful, but fundamentally optional and independent of each other: automatic generation or computer-aided creation of a reference list; dynamic autocomplete, which is automatically created based on previously entered letters, words, and phrases, possibly based on loaded (fictitious) documents (prior art, invention disclosure, previous patent applications) and is maintained in a global or topic-specific glossary (possibly created manually or by artificial intelligence); and interactive spell check and syntax check. The artificial intelligence thus receives highly accurate information from the user almost immediately about the relationships of the entered fictitious features, and the ambiguity in the interpretation of the question posed to the artificial intelligence is significantly reduced.Furthermore, a step is eliminated for the user; namely, the question to the artificial intelligence does not need to be created first. Rather, a further explicit (fictitious) context is created, in that the location of the input (for example, the input mask for a patent claim) makes clear how it relates to the other sections of the patent application body that are yet to be generated or filled with text.In short, a (fictitious) feature of a patent claim generally requires a definition, an embodiment and / or an alternative formulation in the description and a drawing in which (preferably) all fictitious features (at least those provided with a reference numeral) are shown and also specifically explained, namely in such a way that the feasibility of the specific fictitious invention described, i.e., its implementation by a person skilled in the art, is possible.

[0032] It should be noted that other approaches to structured input are also possible and even highly beneficial, such as starting with a fictitious drawing and its fictitious description. In this approach, the fictitious features shown are labeled with reference symbols, thus creating a list of reference symbols. Such a fictitious list of reference symbols is already a simple graph or a part of a more complex graph. Furthermore, by placing reference symbols in a fictitious drawing within the input interface, a relationship and / or location of the corresponding fictitious features is established. Preferably using artificial intelligence configured for image recognition, and preferably with additional information such as a label for the fictitious drawing (for example, for use as a list of figures in the patent application), a relationship between the fictitious features is recognized.This already provides information for fictitious patent claims and / or descriptions of the drawings and / or general descriptions. Again, this is based on an implicit connection provided by the input interface, as explained above.

[0033] In step d., text generation takes place. This process is now well-known, and its quality is widely discussed in the public sphere, particularly in relation to established products. However, user experiences are often not comparable because the requirements vary considerably. Even in professional life, there are many different types of texts, such as emails, responses to official notices, formal and informal communication, ticket formulations, and notes from conversations. Furthermore, the amount of information a user intuitively accesses and the context are often underestimated. For example, it is often difficult to clearly distinguish between information and context; a state-of-the-art technology might have a different focus, meaning that while the terms and functions within it are relevant, they may not be applicable to use in a text.A simple example is the question of defining a joint. The user, a mechanical engineer, might not even consider that the human joints of the shoulder, pelvis, knee, and hand are quite interesting examples, but rather wants to know which joints are commonly used in machines, such as ball joints, hinge joints, pivot joints, sliding joints, or universal joints. This is the context. However, in a (fictitious) invention mentioned here as an example, a connection between joint nodes is described, for instance, the novel use of a rivet connection. It could be inferred from the fictitious invention that a ball joint or sliding joint are not relevant examples in this context.The challenge here is to find a term that describes the overarching concept and thus, in this example, encompasses a desired selection of joints while excluding, as far as possible, only those joints that are not relevant here, for example, those that wouldn't function. Because the hypothetical task is very specific and also rare, there are often no pre-existing texts or rules for how such a word or a suitable expression can be composed or generated. An artificial intelligence is then often unable to provide an answer. However, it is capable of suggesting counter-questions or examples, as well as reviewing a formulation and making suggestions for improvement.

[0034] The complexity of a question cannot always be predicted or may even be misjudged, due to the asymmetry between human intuition and machine precision, as well as limitations. For example, a humanoid robot can easily perform a somersault but struggles with simply walking. Using the graph, such questions are implicitly posed and can be addressed appropriately by an artificial intelligence.For example, in one case, the artificial intelligence finds a treatise on precisely the relevant aspect, such as a word, phrase, or paragraph from a textbook made available to it (for example, in the case of a neural network, through pre-training). Minor adjustments may still be necessary, such as replacing a word, translating into the desired language, and / or rewriting in the required style. In another case, however, the artificial intelligence recognizes that the specific graph, or a section thereof, presents a problem it cannot solve and subsequently expresses its lack of knowledge or makes specific or general suggestions on how this hypothetical problem could be solved.The user does not need any prior knowledge or have to think about it, since the task of creating a patent application body is already sufficiently complex.

[0035] It should be noted that the text is not necessarily, and indeed preferably not, the entire text of the remaining part of the (to be generated) patent application body. For example, initially only (fictitious) claim 1 is provided with text. Preferably, a (fictitious) first feature is provided with text first, whereby the user particularly prefers to choose where to begin, for example, with a fundamental aspect such as the clarification of the fictitious problem in the prior art or a fundamental function of a fictitious feature that is implemented and / or used in a conventional manner. It should be noted that, for the introduction in the patent application body, an independent fictitious patent claim is particularly relevant, and especially its structure.In a straightforward example, the so-called two-part formulation is chosen, whereby the user has already specified what is to be classified as prior art (general term) and what is to be the novel element (identifier). Due to this structure, the fictitious problem and the fictitious solution are already implicitly defined. A supplementary or alternative source is an attached (fictitious) invention disclosure or a (fictitious) prior art document, such as an earlier patent application. While it is possible to have artificial intelligence identify the differences, for example, between the two aforementioned sources, the preferred approach here is that this information originates from the user, who has understood the fictitious invention, mentally categorized it within the known fictitious prior art, and drafted a patent claim based on this understanding.In one embodiment, the user can be supported by an artificial intelligence based on the aforementioned comparison of the sources and the resulting differences, i.e., the classification of the fictitious features as fictitious prior art and novelty.

[0036] An important, indeed fundamental, aspect of generating text in structured sections, assigned to specific features via structural segments, is that these sections should be relatively short, for example, a sentence, a paragraph, or several paragraphs. Such a volume of text is manageable for a (human) user, meaning that interrelations can be grasped and corrections can be made within a manageable timeframe and within the scope of average human concentration. It is clear to both the artificial intelligence and the user what the fictitious text is intended to address. This sounds obvious, but it is not, as will be explained below. For example, the user might want to explain or supplement a fictitious invention or the fictitious subject matter of a patent claim in general terms, without any predefined sequence.When creating text using a conventional, free-form text editor, there are no instructions or aids. Therefore, it can happen that certain aspects (such as features, but also their function and / or interrelationships) are not addressed or not adequately supplemented. Artificial intelligence (as discussed above) already faces the challenge of separating the context and content of the fictional invention from stylistic requirements. It is (at least currently) not uncommon for an aspect to be overlooked; for example, simply because it has been specified that the fictional text should not exceed three paragraphs, and this requirement is weighted so heavily that it is considered more important than actually addressing all aspects of the fictional subtask.

[0037] One positive aspect regarding time and effort is this: the user did not write the fictitious text themselves; rather, the artificial intelligence generated it. Given suitable specifications, the text is correctly executed in the desired style and positioned within the patent application body (for example, general definitions placed in the fictitious description and not in the fictitious patent claims or fictitious figure description). Such a fictitious text or text passage generally meets with the user's approval, often more so than input from a human who has their own writing style and potentially a different understanding of the fictitious invention, or who simply lacks sufficient training. The fictitious text passage can therefore be read quickly and easily supplemented and / or corrected.On the other hand, a text based on one's own draft is more likely to be seen as a naturally good end product compared to a text generated by artificial intelligence, requiring no further review or at least making the review of the text tedious and tolerable.

[0038] In principle, there is a risk that the review will only be superficial. However, according to the fundamental concept of using the proposed method for generating a patent application body, the user has already completed the required preparation by directly or indirectly creating the fictitious patent claims. The conceptual framework is therefore already in place and now only needs to be aligned with the generated text.

[0039] In one embodiment, the division into separate text fields and the corresponding mask structure of the GUI input interface are advantageous because they require less concentration time, make the input more comprehensible and verifiable, and facilitate the recognition of relationships. Furthermore, this approach allows for specific and concise assignments, automated and structured for prompts, yet still directly generated by the user. This, in turn, creates a clear understanding of how the artificial intelligence arrives at its result. Moreover, it significantly simplifies the process for users to create individually tailored prompts or to optimize and / or supplement pre-defined ones.

[0040] It should be noted that this includes the possibility that, as a result of the user's review of the (fictitious) text, the features may need to be modified and / or further features added. The time required for this task is generally less than for writing or dictating independently and is comparable to the time spent proofreading a dictation, the latter being a conventional and therefore well-practiced activity for the user.

[0041] It is further proposed that a newly generated text segment also be added to a (supplementary) source for the graph, i.e., corresponding to the input from step a., in step b., to the computer-readable graph. It should be noted that generating a text segment already includes an intermediate step, which may be performed internally by the program, but preferably externally (i.e., readable by a user, for example via a GUI), namely creating a basic structure upon which the respective text segment is built. Thus, the artificial intelligence already provides a structural template that can easily be represented in a graph.In one embodiment, the generation of additional structural sections in the graph precedes the generation of a text section and is made available to the user for review, as is explained below with regard to the graphic (based on the graph).

[0042] The graph on which the text generation is based is thus built up step by step, under the (not necessarily, but preferably continuous) control of a user. In an application, the sequence is created: (fictitious) patent claim (one or all), introduction, general description (feature by feature), and figure description. The amount of information for the artificial intelligence therefore increases in the usual order of a Japanese patent application. However, the order is not prescribed by the input interface or in any other way, but is fundamentally arbitrary, as can be seen from the approach described above, which begins with fictitious drawings.

[0043] To illustrate with the joint example above: If the generated text passage refers to a human joint instead of a mechanical engineering component, initially only one sentence or a few paragraphs will be generated. This can therefore be corrected directly with minimal effort. This error will not propagate to every detail of the patent application body to be generated because the user has the opportunity to intervene before the subsequent text passages are created.

[0044] If such an intervention is not performed, an erroneous text section can be corrected retrospectively, resulting in a change to the graph. This makes it clearly possible to define which other text sections are or could be affected by such a change. In one embodiment, the system automatically triggers the generation of new or modified (further) text sections. These are then marked accordingly, enabling targeted review. In a trivial example, which is therefore better solved deterministically and algorithmically, a copy of the first patent claim serves as the summary of the patent application body. If the (fictitious) first patent claim is amended, the (fictitious) summary is automatically amended accordingly or simply marked to indicate that a change may be necessary.In the above (also simplified) example with the joint, another text section containing a medical or biological treatise on human joints is deleted and replaced by a treatise on mechanical joints.

[0045] Weightings arise from the context, the number of occurrences of a feature, the number of edges to a node, prior knowledge, and / or general (e.g., programmatic) specifications, such as a distinction from another text field in the input interface, which clarifies that this is a different feature than in the previous text field. This distinction can be supplementary, additional, or alternative. In a simple example, this relationship is at least partially defined by a chosen junction, such as AND, OR, or AND / OR. This junction can be either free text in one implementation or, in an input interface, selectable as needed as a computer-readable element, for example, from a list.

[0046] In a further advantageous embodiment of the method, it is proposed that when generating text in step d., an information sequence in the course of the text is taken into account.

[0047] Due to the difficulties artificial intelligence faces in distinguishing between context and content, it struggles to determine whether a text passage or processed input should be repeated or whether the topic has been sufficiently addressed. Therefore, the approach here is to provide the AI ​​with the sequence of information, thus specifying that the subsequent text passage to be generated should not be repeated, but rather built upon and / or referenced. It should be noted that this problem is not solved if an AI is asked to generate the entire patent application. In particular, repetitions and overlaps are sometimes desirable, for example, when specifying a hypothetical feature design.

[0048] In one embodiment, the potential need to modify an already created text section is also shown, for example, if a contradiction or redundancy arises.

[0049] It should be noted that the effort and sometimes the cost of answering a question to an artificial intelligence increases with the length of the text, while the quality of the answer tends to decrease because more complex weighting is required, and information can be overlooked for various reasons. A graph, on the other hand, generates data that is already prepared in a computer-readable format, or at least significantly reduces the amount of text. Furthermore, the graph already contains weightings that correspond to the user's preferences or, at least due to its programmatic design and structure, are closer to these preferences than an artificial intelligence can (reliably and transparently) achieve.

[0050] In a further advantageous embodiment of the method, it is proposed that the method be executed repeatedly and progressively inline upon input from a user.

[0051] where step d. is preferably executed after a user has completed an input of a single or multiple features.

[0052] Artificial intelligence (AI) operations are currently still relatively slow. For example, it takes between one second and half a minute to generate a response. Depending on the application, it is therefore advantageous if the AI ​​starts generating text while the user is still typing, ensuring that the text is displayed within a reasonable timeframe. The text is then preferably displayed as a suggestion, which can be ignored if it is not desired or if it does not match the intended result. Conversely, a user can accept such a suggestion, thereby speeding up the input process, or, conversely, retract an input already entered because the suggestion indicates that the input is incorrect or insufficiently precise.

[0053] In a preferred embodiment, a question is not continuously output to the artificial intelligence, but rather certain events (for example, input of the tab key or mouse to change text fields) are awaited, which correspond to the completion of an input or are sufficiently likely to correspond to such a completion. For example, it is determined whether a space or punctuation mark has been entered, from which it follows that a word is highly likely to have been completed. In the case of spoken input, this can be determined accordingly by recognizing pauses in speech or via the text automatically recognized from the spoken language. In another embodiment or in another application, preferably additionally, a fictitious feature, which is an expression comprising several words, is first, for example:

[0054] - after a clearly completed input action, such as inserting from memory,

[0055] - after a longer pause in input,

[0056] - in the event of a significant delay in input speed,

[0057] - after entering a junction, and / or

[0058] - after leaving an input field

[0059] used to generate text. A question to the artificial intelligence is preferably aborted, the output of text suppressed, or the display of a text suggestion even reversed if an input is at least partially deleted or highlighted.

[0060] In a further advantageous embodiment of the method, it is proposed that in step a. the individual features can be entered into the input interface by a user in a data-technically separated manner, wherein at least one of the associated nodes, preferably a group of superior nodes in the graph, is generated for the graph to be created in step b. by means of the user-side separation.

[0061] In one embodiment, the input interface displays multiple input fields, for example, in a number as required, with each input into such a field being understood as a data-separated feature. It is possible that this data-separated feature may contain at least one other feature overlapping within it. For example, the data-separated feature could be a multi-word expression, and at least one of the words could be a (separate) feature (word feature), such as one already known from a previous input, for example, from the data-separated input itself, such as from a previously filled input field. It should be noted that the recognition or classification of an input as a feature is performed based on various programmatically defined specifications.In the patent application presented here, nouns are automatically recognized during creation, and a selection of them are added to a reference numeral list based on predefined rules, thereby generating a word feature, for example, with an assigned reference numeral. Furthermore, the language entered in an input field is recognized as a feature (expression feature) during this process. Additionally, the arrangement within the patent application, specifically in a patent claim or text passage, is recognized as a feature (structural feature) and incorporated into the graph's structure.

[0062] The fundamental problem with using human or human-simulating support is that the subject matter must be understood as thoroughly as possible to recognize whether similar terms and expressions belong together, for example, describing the same thing, or, conversely, are intended to describe different things. For instance, when drafting patent application documents, it is common practice to thematically separate terms through relatively minor modifications, such as a first tab and a second tab. Depending on the context, these tabs can represent mutually exclusive alternatives, complementary configurations, or similar configurations with the same or different function and / or location. Reality is complex and offers a multitude of scenarios.By using an input interface to digitally separate the user's current intrinsic knowledge or concrete (currently purely mental) design, this knowledge can be directly represented in the graph and therefore no longer requires additional processing. This ensures a high degree of reliability in further processing, not only through the artificial intelligence used, but also for the user themselves, who may not have the underlying mental design readily available at a later time.

[0063] It should be noted that the data separation is preferably implemented such that the separated feature (for example, the aforementioned first tab) contains further, uniquely identical or different conditional information within this data-separated unit. For example, a first text field in the input interface specifies that the first tab is located at a first location, and a second text field specifies that the second tab is located at a second location. This can thus be implemented deterministically algorithmically in a graph with, for example, two nodes and two edges, one for each of the tabs and their location.

[0064] In a further advantageous embodiment of the method, it is proposed that the feature separation be visually represented in the input interface.

[0065] In one embodiment, the data separations are visualized using test fields, for example, in the form of a box. Within this box, for instance, the first tab with its (first) properties and the second tab with its (second) properties are entered or displayed. Alternatively or additionally, color coding is used. In another embodiment, an indentation and / or connecting line is displayed, preferably entered via the input interface, which includes further information such as a ranking, for example, a general type of tab that comprises two (for example, mutually alternative) embodiments of tabs.In one embodiment, data-technically separated features are displayed in a separate list, wherein the list contains, for example, only one main feature, such as a word feature, with a single distinguishing property, such as first tab, second tab, where the distinguishing property is the respective ordinal number.

[0066] In a further advantageous embodiment of the method, it is proposed that a feature can be entered in language form in combination with a comment.

[0067] The commentary is also processed as nodes, edges, and weights of the graph and embedded in its structure, and is taken into account by the artificial intelligence in step d. When drafting a patent application body, especially a patent claim, a high degree of abstraction is often necessary. Some (fictitious) patent claims are so abstract that they are hardly understandable on their own. However, these features (as mentioned previously) originated from the user's conceptual design, meaning the reasons for how each individual feature is defined are intrinsically known. Many patent application drafters therefore add comments to their designs, even for their own information, in case work can only be continued after a longer break or a mental rethink.

[0068] For example, such a (fictitious) comment could specify a (fictitious) feature in a (fictitious) patent claim, which is referred to as a tab, as a bent projection from a sheet metal part, as a punched-out lever, and / or as an insertion element for a positive-locking connection. However, indications of dissatisfaction with the chosen terminology and / or expression, as well as references to questions that still need clarification, can also be the subject of such a comment or multiple comments.

[0069] It is proposed here that such a (fictitious) comment be included in the graph to be generated, clearly marked as a comment and therefore not (at least not unchanged) as part of the text of the fictitious patent application body to be generated. Rather, such a comment should be treated as metadata or as a basis for information. The artificial intelligence, which cannot readily distinguish between a part of the patent application body and a comment, thus receives a classification based on the structure of the graph itself. This classification is ultimately based on the user's input behavior and therefore enables low-loss information transfer from the user to the artificial intelligence via the graph and the input interface.In this process, the user's input behavior does not need to be adapted to this method, and the user behaves (at least in this aspect) in the same or very similar way as when conventionally creating a (fictitious) patent application body. Furthermore, in an advantageous embodiment of the method, it is proposed that a (fictitious) feature and / or a (fictitious) comment can be entered into the input interface by a user as an audio recording.

[0070] This proposal proposes enabling user input via audio recording. In this application, (fictitious) features are typed or dictated using a keyboard, as is currently common practice, with (fictitious) comments dictated simultaneously, either concurrently with the typing of the features or sequentially. A preferred approach involves the data separation of these features and comments, represented by an ad-hoc or time-delayed visual display on the input interface, such as text generation using NLP (Natural Language Processing) or LLM (Large Language Model). In one embodiment, the original audio recording is retained and / or only the audio recording is stored and can be played back as needed, while still extracting (fictitious) features and / or comments from the audio recording for the graph to be created.

[0071] In a further advantageous embodiment of the method, it is proposed that an artificial intelligence estimates and thus generates edges, nodes and / or weights based on the syntax and / or semantics of a feature and / or a comment in step b2. before step c.

[0072] In many cases, depending on the specific (fictitious) invention, it can be challenging for a user to mentally break down the subject matter of a patent application into its smallest components. Larger features, containing a greater amount of information, offer a good compromise. This allows part of the graph, such as its basic structure, to be generated deterministically using an algorithm, while the finer subdivision is supported by artificial intelligence. This works for two reasons. First, the complexity is already reduced by the user to a level manageable for artificial intelligence. It's important to note that artificial intelligence needs more than just the currently entered information; it also requires details about the context, style, and potentially other factors.The second point is that the result output by the artificial intelligence is limited in scope, especially when presented graphically. Its scope is so limited that it doesn't cause frustration for the user, nor does it elicit any satisfaction at the reduced workload. It should be noted that artificial intelligence can also be supported deterministically or recursively by comparing the results to determine if they are sufficiently similar.

[0073] In a further advantageous embodiment of the method, it is proposed that an artificial intelligence decomposes a complex feature and / or comment into sub-features in step b1 before step c.

[0074] As previously described, subdividing a (fictitious) feature and / or comment into subunits is often a significant challenge, particularly in a conventional workflow without computer support, where this is either unnecessary or not strictly required. Here, it is proposed that artificial intelligence decompose the fictitious features or comments into further sub-features, if necessary. These fictitious sub-features do not necessarily have to be broken down into separate words, but rather exhibit partial overlap in an application or implementation.

[0075] For example, in the previously mentioned example, there is a feature with a first (and second) tab, from which the sub-feature "tab" is derived as a term for a higher-level concept, and the sub-feature "first" and "second tab" twice, respectively, with a more precise description of their respective properties. It should be noted that such a higher-level concept is not a necessary or meaningful feature in every situation, for example, with a large number of components with the same name and / or similar functions, which, however, do not have a conceptual relationship, such as screws in a construction.

[0076] In a further advantageous embodiment of the method, it is proposed that a feature comprises at least one word feature, wherein a word feature is used multiple times in the multiple features, and in step b. exactly one single node is generated for each of the word features.

[0077] preferably at least one of the word features has a reference sign.

[0078] A word feature is provided as the smallest unit of a (fictitious) feature or as the smallest embodiment of a feature. In one embodiment, the words of a list of reference numerals are set out as word features, but preferably also other words which, while not having a physical equivalent in a representation, for example as a concept, property and / or object of an invention, but which are of great importance and / or require multiple repetitions in a patent application body.

[0079] It should be noted that in one embodiment, a (fictitious) feature comprises a word feature and, in addition, further, preferably overlapping features, for example, syntactically or semantically determined features. Above all, a classification into word features is particularly useful because these smallest units are repeatedly referred to in a patent application.

[0080] Such a (fictitious) word feature is not repeated in the graph, i.e., it is not redundantly created. Rather, it is ensured that the word feature is always recognized as the same word feature despite the many repetitions and is not created anew. This can be easily implemented, for example, with a (preferably deterministic) algorithm, preferably by creating a (for example, grammatical) word stem and recognizing it upon repeated occurrences. In the case of words with the same word stem, as in the previous example with the first and second tabs, one embodiment includes the corresponding (distinguishing) adjective or plural of adjectives. Alternatively or additionally, the user is required to mark the entered word accordingly, for example, by appending a number or other identifier.For example, a label could be implemented as follows: (first) "Laschei" (or "Lasche A") and (second) Lasche2 (or "Lasche B"). Once these "Lasche12" (or "Lasche AB") are defined, i.e., created in the graph, the adjectives can be omitted as needed. Preferably, the adjectives are suggested via optional autocomplete during input, and / or further information about the context is displayed when the word attribute is to be entered again.

[0081] In a preferred embodiment, a word feature is linked to a reference numeral, with such a reference numeral being used mandatorily or optionally at various locations within a patent application body, depending on the embodiment and user requirements. For example, in a classic German or European, Chinese, Korean, and Japanese patent application body, or in the text issued for filing, these reference numerals are inserted in the patent claims, an (optional) list of reference numerals, and in the description of the figures following the respective word feature. The graph, together with the user's targeted input via the input interface, ensures that the relationship between a word feature and the reference numeral is always unambiguously correct.For example, in an embodiment with the example above, the user's input "Lap 2" in a single sentence of a patent claim would result in the following text: "at least one tab (10, 10a, 10b)", where reference numeral 10 denotes the superordinate term and reference numerals 10a, 10b denote the two more precise and / or alternative embodiments, which may then be addressed individually in a dependent patent claim with the input "first tab" with the text in the final version: "first tab (10a)". While this is so simple and trivial, the added value for the user lies primarily in the fact that consistency in the chosen terminology is supported and thus facilitated, eliminating not only a tedious task for a human but also resulting in significantly increased reliability.Neither a human nor an artificial intelligence addressed without such a graph can offer such support and high reliability at the same time.

[0082] In a further advantageous embodiment of the method, it is proposed that the edges and weights of a word feature are generated based on at least one, and preferably exclusively, of the following information:

[0083] - Relations arising from syntax and / or semantics;

[0084] - an adjective;

[0085] - a pronoun; and

[0086] - a junction,

[0087] preferably recognized and / or evaluated using artificial intelligence.

[0088] It should be noted that the information is either provided to the artificial intelligence during or prior to the transmission of the (fictitious) task, or has already been learned by the artificial intelligence. In both cases, it is referred to here as having been taught to the artificial intelligence. The above list is not exhaustive.

[0089] In one embodiment, at least part of the design (for example, a basic structure) is generated deterministically and algorithmically based on this information. The rules for this are simple, transparent, and easily understood by a user. This is often possible because the patent application is a legal text that attempts to translate a lived reality into logic and then express it in words. It helps to limit oneself to a few (preferably legally sound) established terms and avoid resorting to flowery prose.

[0090] In one embodiment, where an artificial intelligence recognizes and evaluates the information for generating or supplementing the graph, the AI ​​has been taught, for example, that the OR junction leads to alternatives when creating a graph. In a simple or simplified case, in a tree structure within the graph, two adjacent nodes are connected to each other without an edge, but with an edge leading to a common node. For example, the AND junction leads to a concatenation; that is, in a simple or simplified case, in a tree structure within the graph, two nodes have a directly connecting edge and are each connected to a common node via an edge. It is further proposed in an advantageous embodiment of the method that in step a.An artificial intelligence supports the input of a feature with word suggestions based on prior information, whereby at least one, preferably exclusively, of the following prior information is used by the artificial intelligence:

[0091] - an explanation and / or specification entered by a user, preferably in written form or as an image;

[0092] - an explanation and / or specification found automatically by the artificial intelligence;

[0093] - a separate pre-generated patent application body, preferably with a graph generated according to step b.;

[0094] - a graph already generated according to step b.; and

[0095] - a conceptual proximity of a feature in a graph already generated according to step b., whereby the conceptual proximity is determined by the artificial intelligence from the edges and weights.

[0096] The aim of the support provided by the proposed method is not solely to reduce working time by relieving the user of time-consuming, error-prone, and / or tedious tasks. Rather, it also preferably provides support in the process of creating the patent application body. The goal is for the artificial intelligence to appear as if it could read or even predict the user's thoughts. This is possible to a limited extent by providing contextual information, preferably the context that a text for a patent application body is to be generated, which is already taught to the artificial intelligence.

[0097] A word suggestion, for example, is displayed as so-called ghost text, which continues the already entered phrase. However, it is visually clear that this has not yet been entered and that it is not mandatory, but merely a suggestion. For example, when entering text, the ghost text is displayed in a subtle color and / or italics behind the cursor line in the direction of writing. Features already contained in the graph are preferentially highlighted within the ghost text.

[0098] For example, when entering spoken text as an inline dictation (i.e., ad-hoc conversion to text), the ghost text is displayed as described previously or similarly. For example, when entering spoken text as a conventional dictation (i.e., creating an audio file), word suggestions are displayed legibly on the input interface.

[0099] When using artificial intelligence, it is difficult to define the boundaries of what information should be transmitted to the AI, what it must receive and what it should not, and also where the information density or amount becomes too high to be adequately processed. It should be noted that a high amount of information is often less critical than a high information density. Furthermore, it is difficult to assess whether additional information facilitates the AI's problem-solving process or dilutes the task definition. It is therefore suggested to use the listed preliminary information and preferably to rely exclusively on it.Which of the relevant preliminary information is relevant depends on the specific (fictitious) task posed by the user, for example, explanations of the state of the art, lists of alternative approaches, formulations of bullet points or rough drafts, and many other scenarios. It should be explicitly noted that the aforementioned (or other) preliminary information can be combined.

[0100] An entered explanation and / or specification is, for example, entered into the input interface as a comment, as previously explained. In one embodiment, an image, for example, for use as a figure, is entered, preferably with reference symbols and associated arrows. Because images have a high information density, such preliminary information is very powerful and in many cases highly useful, for example, for generating a text description of a figure or for creating a patent-compliant drawing. It is particularly convenient for a user to enter preliminary information into the input interface in spoken form. Speech recognition tools are preferably used here, so that rapid intervention is also possible, for example, for necessary corrections and / or to terminate an output from an artificial intelligence.In one embodiment, the artificial intelligence already possesses prior knowledge, which is then automatically determined and appropriately generated when faced with a task. In another embodiment, the artificial intelligence is configured to search a database and present and / or utilize relevant search results. For example, to distinguish between (fictitious) prior art and other known concepts, it is advantageous to search for the meaning and usage of (fictitious) terms so that a word suggestion is not misleading, for instance, because a commonly used term included in the (fictitious) invention disclosure might have other meanings for an inventor or even be incorrect, and / or require substantiation by a definition.

[0101] In one embodiment, a pre-generated patent application body is used to derive stylistic specifications, content requirements and / or further rules, for example rules hidden from average human perception.

[0102] In an advantageous embodiment, patent application bodies that already possess a graph are used, with particular preference being given to additional or exclusive access to this graph.

[0103] In one embodiment, prior work by the user itself or patent applications with relevance, for example, the same classification and / or similar titles, or parts of their content, are used as prior information. For example, when creating descriptions of the prior art, the same parts can be used and adapted to the present task.

[0104] Artificial intelligence often excels at recognizing similarities in (fictitious) terminology, structure, and relevance to the task at hand. While this task is not inherently difficult for a human, it requires time to locate the corresponding (fictitious, earlier) patent applications and to identify the relevant sections and integrate them into the current (fictitious) patent application. Ideally, with each use of the method, the (fictitious, earlier) patent applications are fully indexed and thus made available for full-text searches, and preferably also, or even exclusively, for image searches.

[0105] In one embodiment, other information (in addition to or as an alternative to a completed patent application) is already available as a graph, such as general information on the prior art, terminology (of the user, the inventor, and / or the company where the inventor is employed), and / or stylistic guidelines. For example, this information may be explicitly presented as a graph or entered in text and / or image form and then converted into a graph. One application is, for instance, the identification of the (fictitious) inventor or their company and, based on this, the provision of a glossary of terms commonly used or mandatory within that context.

[0106] In a further advantageous embodiment of the method, it is proposed that in step a. an artificial intelligence supports the input of a feature with teaching suggestions based on prior information, wherein at least one, preferably exclusively, of the following prior information is used by the artificial intelligence:

[0107] - an explanation and / or specification entered by a user, preferably in written form or as an image;

[0108] - a specification automatically discovered by the artificial intelligence;

[0109] - a separate pre-generated patent application body, preferably with a graph generated according to step b.;

[0110] - a graph already generated according to step b.; and

[0111] - a conceptual proximity of a feature in a graph already generated according to step b., whereby the conceptual proximity is determined by the artificial intelligence from the edges and weights.

[0112] As previously stated, the aim of the support provided by the proposed method is to assist in the process of drafting the patent application body. This will leverage the inherent limitations of artificial intelligence, namely its ability to be either unavailable or subject to varying levels of performance and distraction. While artificial intelligence may be limited by computational resources, it is generally not distracted by other tasks.

[0113] A teaching suggestion is not necessarily a lecture, which can sometimes be perceived as unpleasant, but rather serves as a reminder, offering suggestions for more in-depth explanations and / or pointing out potential problems with legal validity and / or translatability into other languages, preferably providing alternative suggestions. A reason is always given for each teaching suggestion, allowing the user to decide whether to heed it or whether it is even appropriate in this context.

[0114] This is also possible to a limited extent by providing contextual information, preferably the context that a text for a patent application is to be generated, which is already taught to artificial intelligence. Reference is made to the preceding explanations and examples regarding usable prior information for generating word suggestions.

[0115] In a further advantageous embodiment of the method, it is proposed that in step a1. an artificial intelligence checks a user's input,

[0116] where at least one, preferably exclusively, of the following prior information is used by the artificial intelligence:

[0117] - an explanation and / or specification entered by a user, preferably in written form or as an image;

[0118] - an explanation and / or specification found automatically by the artificial intelligence;

[0119] - a separate pre-generated patent application body, preferably with a graph generated according to step b.;

[0120] - a graph already generated according to step b.; and

[0121] - a conceptual proximity of a feature in a graph already generated according to step b., where the conceptual proximity is determined by the artificial intelligence from the edges and weights. As previously explained, the aim of the support is to provide assistance in the process of creating the patent application body using the method proposed here. This will utilize the ability of artificial intelligence to quickly process large amounts of information and thereby detect deviations and errors (for example, through pattern recognition).It should be noted that the graph, which was largely created almost directly by the user, will not contain the usual (human) flaws in a patent application, such as incorrectly or inconsistently assigned reference numerals or identical terms used multiple times in different ways, but also the omission of definitions and the suppression of examples.

[0122] The review of the application described here focuses on aspects that can only be evaluated by a person with considerable effort or solely with knowledge of the conceptual framework of the patent application. Examples include quality, consistency, legally sound wording, and / or translatability.

[0123] Reference is made analogously to the preceding explanations and examples regarding usable prior information for generating word suggestions and / or teaching suggestions.

[0124] Furthermore, the substantive proximity of a feature or potential prior information is relevant here, because in this state the new feature is already in a finished state, i.e., already defined and possibly supported by an example. Thus, the (fictitious) feature in question already exhibits a level of complexity that allows for a similarity analysis or match analysis performed by artificial intelligence, which is either impossible or not feasible for a human within a reasonable timeframe. For example, a definition touches upon an aspect that appears in another graph (e.g., a previously created patent application body), thus exhibiting an overlap. In the example of the tab, a materiality or a manufacturing process could be mentioned.For example, the inventory graph includes a comment stating that the inventor has specific ideas or the user has (unavailable) knowledge regarding terminology or distinction from other terms, which could easily lead to confusion and / or difficulties in the licensing process and / or translations, such as the term "Löten" (soldering), which in German is considered a general term for both hard and soft soldering, but in an English translation could be translated with only one of the two terms (i.e., English: brazing or English: soldering).

[0125] Alternatively or additionally, such knowledge is used as a teaching suggestion, but in this case, the soldering process cannot be a mere side issue in the current work, so that this teaching suggestion is predictably useful. Alternatively or additionally, such knowledge is used as a word suggestion, in which case the soldering process, which is more of a side issue in prior knowledge (e.g., graphs), results from the similarity of the device, for example, the tab, so that this word suggestion is predictably useful.

[0126] In a further advantageous embodiment of the method, it is proposed that the graph be output as a human-readable graphic, preferably the graphic being generated by an artificial intelligence.

[0127] Once a graph is available, it's easy to generate a graphic using known (deterministic) algorithms. A graphic has the advantage of being much faster and more structured for human understanding than text. Training may be necessary, for example, on how to represent a junction in a graphic, but this only requires brief explanations, which can be provided with a legend or help text, and minimal practice to quickly grasp even complex graphics.

[0128] Such a graphic is not only useful to the user. The inventor or other client (as the user of the graphic but not the creator of the patent application) can also quickly check whether their invention is presented correctly and comprehensively without having to work through the legally required wording. Above all, alleged combinations are quickly recognizable and easily comprehensible. Currently, combinations that arise from the wording in a strictly logical way, but may not be intended or technically meaningful, are accepted due to time constraints, a lack of overview, and / or the knowledge that a nonsensical combination can be argued to be ineffectively disclosed.The latter, however, requires an implementation that is flawless at least in this aspect, including in the translation, and potentially knowledge of all possible scenarios. Furthermore, an invention may inadvertently be anticipated, while at the same time, protection may not be granted, or may not be legally sound, due to a lack of feasibility.

[0129] In many scenarios, the graph is complex with a multitude of dimensions, which limits its graphical representation, for example, on a single sheet. In such cases, it is advantageous to reduce the complexity, primarily through artificial intelligence, and / or to choose a representation that allows for the most accurate possible depiction of the graph in its specific dimensionality.

[0130] In a simplified example (easily and reliably solvable with a deterministic algorithm), it is recognized that a subset of examples has been mentioned. This subset is, at least logically, also referenced in one or more other places. This subset is then given a title or symbol, referred to here as an icon, that is as intuitively meaningful as possible. The icon is used multiple times, possibly with the same color scheme each time. If a user wants to know more about this aspect, a further display or the opening of a display space can show the details behind this icon, for example, including all links to other attributes. In another example, many links exist, for example, from the attribute of the title of a main claim to each of the other attributes.In one embodiment, such a feature is not integrated into the rest of the graphic, but rather displayed separately, for example, as the graphic's title, within a box framing several features, or as a legend, perhaps with a color coding. The risk of the artificial intelligence omitting features or presenting an incorrect relationship is eliminated because the graph is such a precise data format that the AI's intervention can be limited solely to the presentation. The only remaining question is whether a different presentation format could be better designed. Given the need to minimize time expenditure, this is acceptable in this application compared to the alternative of a fixed presentation format, which might not adequately address the complexity involved, or a representation developed or revised by a human.

[0131] In a further advantageous embodiment of the method, it is proposed that the graphic is represented by a graphical modeling language and is designed to be interactive for user modifications.

[0132] In this embodiment, the displayed graphic is equally well-suited for human intervention and / or for presenting change suggestions from an artificial intelligence. Access via a computer mouse, touchscreen, and / or voice control allows for much faster adjustment of the graphic and a much quicker assessment of the effect of any change compared to text. The graphic is then in a fixed, deterministic relationship with the graph via the modeling language, meaning that a change to the graphic leads to a corresponding change in the graph. This provides a fundamental basis for text generation, ensuring that all dependent aspects are necessarily modified. Simultaneously, the changes are precise and implemented only where absolutely necessary.

[0133] In the aforementioned example with the tabs, both tabs were assigned the same properties regarding manufacturing process and materials because no particular attention was paid to this. In a conventional patent application text, this relationship might not be immediately apparent upon reading, even if it is explicitly stated through cross-referencing or identical terminology. The reviewing user now notices this relationship through the graphical representation of the graph underlying the patent application, and can directly modify and, if necessary, annotate the graphic interactively. For example, an icon (see previous explanation) is opened, a link is broken, and a reduced copy with the correctly matching features is created, or the non-matching part is removed from the icon.A simple cross-reference resulting from this icon in the text is then precisely adjusted by the artificial intelligence, exclusively and exclusively where necessary, for example, in a (fictitious) patent claim, a general description, a character description, and / or a summary. These changes are usually so minor that they can be verified within a reasonable timeframe.

[0134] In one aspect, such a graph, represented as a graphic, is used by a patent examiner, whereby the graph is generated by the examiner's own input or provided by the applicant in a processable file format. This significantly facilitates and speeds up the comparison with prior art. At the same time, control over the process always remains with the patent examiner and is not, or at least not solely, carried out by artificial intelligence. This compromises traceability and / or reliability, and in particular, a level of examination quality with sufficient legal soundness is not necessarily achieved.

[0135] In a further advantageous embodiment of the method, it is proposed that the graphic be generated by an artificial intelligence and designed to be interactive for user modifications.

[0136] In one embodiment, a selection of (fictitious) features is displayed in a graphic so that it is legible for a user or has a manageable size on a limited screen or printout. For example, information from the graph can be omitted from the graphic, or a section can be windowed. In the embodiment proposed here, preferably no modeling language is used. Instead, the graphic is generated by an artificial intelligence, but it is designed to be interactive, allowing a user to modify the graphic via the input interface (preferably a GUI). This modification is then read by the artificial intelligence and implemented as an adjustment to the graph, for example, changing a weight and / or adding, deleting, or relocating an edge.Such an interaction (in this embodiment, as well as possibly in the one previously explained using the modeling language) is, for example, grasping and moving, and / or adding and / or deleting one or more nodes.

[0137] In a further advantageous embodiment of the method, it is proposed that at least one of the features of the graph be provided with reference numerals, wherein the reference numerals are assigned to a word feature and originate from the graph.

[0138] where it is preferably entered by a user which word features receive a reference mark,

[0139] In step a., suggestions are made to the user as to which word feature should receive a reference mark, with particular preference given to the user.

[0140] As mentioned previously, many patent applications require reference numerals. These must meet the highest standards of consistency. This places a significant challenge on the human user, whereas a computer, or more precisely a (preferably deterministic) algorithm, can perform this task with complete reliability without requiring significant processing power.

[0141] At the same time, it is often impossible for an artificial intelligence presented with a text to predict which words need to be tagged with a reference mark and whether homophones should always receive the same reference mark. However, this knowledge is intrinsically present in the user during input and can be quickly implemented with simple instructions. In one implementation, words with certain properties (for example, nouns or those with specific indices) are automatically tagged with a reference mark. The user has the option to override this automatic assignment. Subsequently, such word features with reference marks, once entered, are then deterministically recognized via the graph and sorted accordingly, and are always assigned the same reference mark.In this case, autocomplete is preferably used to provide support, which facilitates the correct repetition of the word feature when typing.

[0142] In one embodiment, the word features, which are provided with a reference mark, are selected solely by the user.

[0143] In one embodiment, such possible, preferably sufficiently probable, word features are recognized by an (preferably deterministic) algorithm, wherein this selection is preferably overridden (i.e., deselected) and / or supplemented by a user. It should be noted that such recognition always applies globally to the entire patent application body. In one embodiment, knowledge from previous graphs and / or patent application bodies is used to determine which word features should be assigned a reference mark.

[0144] It should be noted that in one embodiment, word features are dynamically defined according to certain properties in a glossary of the patent application body. Based on prior experience or specific rules, the word features then either receive a reference mark or no reference mark.

[0145] It should be noted again that in the entire (fictitious) patent application body, a word feature is recognized at each repetition and is accordingly represented in the graph, in an easily understandable example as a node. With each further mention, additional edges to other nodes are created, and possibly separate nodes (i.e., with a property relating exclusively to itself).

[0146] In a further advantageous embodiment of the method, it is proposed that when generating the graphic, the information density is prioritized and, if necessary, reduced, and that at least one, preferably exclusively, of the following prior information is used by the artificial intelligence: - an explanation and / or specification entered by a user, preferably in language form or as an image;

[0147] - an explanation and / or specification independently discovered by the artificial intelligence; and

[0148] - a separate pre-generated patent application body, preferably with a graph generated according to step b.

[0149] In a further advantageous embodiment of the method, it is proposed that the text be output as a text body of a patent application body.

[0150] preferably the output of the text is performed algorithmically using a template, and / or

[0151] The preferred method is to output the text using artificial intelligence in a free-form format, incorporating stylistic specifications from a user.

[0152] For the conventional use of the patent application body, a text body must be created, which can be, for example, an editable text file, an uneditable text file (such as a so-called PDF / A [Portable Document Format for Archiving]), and / or printed pages. Certain formatting requirements must be observed for this. If text is entered directly into the form of an (editable) text body in the conventional manner, there is a high risk that the form of the text body will be altered, thus compromising compliance with the formatting requirements. Examples include changing the font size or hiding text sections due to the unintentional pressing of a key combination.

[0153] Using a template avoids such potential errors and significantly reduces the need for user training compared to a general tool for creating various types of text. The template thus eliminates the need for further checking and potential rework.

[0154] In an application, a rigid template is not the desired way to output the text body. Rather, it is desirable to design it differently for different purposes, for example, for one and the same patent application body differently for examination by the inventor (e.g., without redundancies and with comments) and for the applicant (e.g., with notes for the pending examination procedure and international prosecution, especially information on translation) and for the patent office, or in a format for the individual requirements of the respective patent office (e.g., with a corresponding order of patent claims, description and figures different in Europe, USA and Japan).For example, in a template, the number of royalty-free patent claims (e.g., 15 patent claims for the European Patent Office) is also displayed accordingly, whereby the remaining text components created as patent claims (possibly redundant with the patent claims) are included in the description and / or conversely, a general description is set out as patent claims as far as possible and a remaining part of the description is executed as a figure description (e.g., for the USPTO).

[0155] The following is a simplified representation of an input interface adapted to the standard requirements for a patent application body, in which a user has entered the features in text form, using the present application body as an example (deviations possible).

[0156] One aspect of the invention described above will be explained in detail below against the relevant technical background and with reference to the accompanying drawing, which shows an exemplary embodiment. The invention is in no way limited by the purely schematic drawings, although it should be noted that the drawings are not dimensionally accurate and are not suitable for defining size relationships. This will be illustrated in

[0157] Fig. 1 : a graphic of a feature of claim 1 and claim 10;

[0158] Fig. 2: Exemplary representation of an input mask; and

[0159] Fig. 3: Simplified overview of the (automatically) recognized word features

[0160] Features. Figure 1 shows a graphic G of a feature of claim 1 and claim 10, as well as the associated description (deviations are possible). This illustrates the relationship between the features, which already result structurally from the separation during user input.

[0161] The following is a possible (simplified) source code of a modeling language, which can be used to generate the graphic G using a computer.

[0162] graph TD

[0163] A [ "input interfaces" ] > D ["language form" ] & E ["image" ] & F [ "video" ] & G [ "hand sketch" ]

[0164] D > I [ "merkmall " ] & J [ "merkmal2 " ] &

[0165] K [ "merkmalall " ] & L [ "merkmal4 " ]

[0166] IM [ "word feature " ]

[0167] MN [ "autocomplete" ] & J & K & LI > P ["BZ2" ]

[0168] K > R ["no BZ"]

[0169] 0 S [ "automatic assignment" ]

[0170]

[0171] & &

[0172] T [ "overswappable " ] --> S

[0173] Figure 2 shows a file number and work title in the upper left. To the right of this is a search mask, which is also configured for replacement. The search term is underlined in the text fields of claim 1. Below this, on the left, is the list of (possible) claims, which in this case are reduced to 10 patent claims. Their order can be changed, as indicated by the letter numbering. On the right is an (interactive) reference numeral list, in a version generated automatically without further user input. In this view, only those features that are word features and that are also to be assigned a reference numeral are shown. In the center is the editing area, which displays several text fields and selection fields. Text fields here are "Title," "Transition," and several (expandable by the plus sign at the bottom) "Superordinate term / single-part claim" fields.

[0174] Selection fields include "Type of dependency" with the options "independent", "dependent" and "subordinate", which changes the displayed fields accordingly, and "Claim type" with the options "one-part" or "two-part".

[0175] To simplify the differentiation between user input and the input interface specifications, the user input (here in text form) is displayed in a serif font. Word features included in the reference character list are shown in italics. Features that also have a reference character are shown in bold. Further HTML-like features are also shown: a pair of vertical bars "|" indicates text to be written in bold in the generated text, and a tilde indicates...

[0176]

[0177] The `` character represents a tab stop, and a pair of two hyphens suppresses the enclosed text from the output containing the text to be generated. User comments, which can be understood and read by artificial intelligence, are located between these double dashes.

[0178] The following is a simplified overview of the features automatically recognized as word features, listing their occurrence in the description and their location, specifically their relationship to the respective claim. It should be noted that the word features shown with strikethrough are without reference marks, while the others (in the column with "123" in the header) list the reference marks, using the present application as an example (deviations are possible).

[0179] It should be noted that Figure 3 also shows an optional part of the (interactive) reference numeral list, which allows users to search, filter, assign reference numerals to, and add to word features. For example, by clicking on one of the framed numbers, which represent a descriptive section assigned to the correspondingly numbered claim, the word feature is directly located and displayed. This is also possible for the claims, figure descriptions, and figures. Furthermore, an automated check is included to determine whether listed word features are very similar to one another, thus preventing errors (e.g., due to typos). It should also be noted that all word features listed here are automatically generated from an autocomplete function specific to the patent application body, which is suggested by the user as they type, thereby significantly speeding up the input process.

[0180] The procedure proposed here makes the preparation of patent applications efficient and of high quality. (Reference list)

[0181] 1 Patent application body Input interface

[0182] Features

[0183] Input form

[0184] Language form

[0185] spoken language

[0186] Text form

[0187] Picture

[0188] video

[0189] 10 Hand sketch

[0190] 11 knots

[0191] 12 edges

[0192] 13 Word feature

[0193] 14 automatic assignment

[0194] 15 overridable

[0195] 16 typed in

[0196] 17 Autocomplete

[0197] BZ1 Reference mark

[0198] BZ2 Reference mark

[0199] BZ3 Reference mark

[0200] no BZ

[0201] M1 first characteristic

[0202] M2 second characteristic

[0203] M3 third characteristic

[0204] M4 fourth feature

[0205] G graphic

Claims

Patent claims 1. Method for generating a patent application body (1 ) performed by a computer, comprising the following steps: a. Providing an input interface (2) for a user, which is set up for entering features (M1, M2, M3, M4) of an invention; b. Processing the features (M1, M2, M3, M4) entered via the input interface (2) in the form of a computer-readable graph comprising a structure consisting of nodes (11), edges (12) and weights; c. Reading the graph into an artificial intelligence, wherein the artificial intelligence is configured to process the structure of the graph and those features (M1, M2, M3, M4) stored in the graph; d. by means of an artificial intelligence which is set up to generate human language (6), and based on step c. generating text, wherein the text is generated in individual text sections assigned to corresponding structural sections of the structure of the graph, and wherein the respective generated text section is embedded in the structure of the graph as additional nodes (11), edges (12) and weights and when generating another text section according to step d. the newly embedded structure is taken into account by the artificial intelligence depending on the newly resulting weights.

2. The method of claim 1, wherein When generating text in step d., an information sequence in the course of the text is taken into account.

3. Method according to claim 1 or claim 2, wherein the process is repeated and executed intermittently in response to user input, building upon each other wherein preferably step d. is executed after a user has completed the input of a single or multiple features (M1, M2, M3, M4).

4. Method according to one of the preceding claims, wherein in step a. the individual features (M1, M2, M3, M4) can be entered by a user into the input interface (2) in a data-technically separated manner, wherein at least one of the associated nodes (11), preferably a group of nodes (11) that are superior in the graph, is generated for the graph to be generated in step b. by means of user-side separation.

5. Method according to claim 4, wherein The input interface (2) visually represents the separation of the features (M1 ,M2,M3,M4).

6. Method according to any one of the preceding claims, wherein a feature (M1 ,M2,M3,M4) combined with a comment in language form (5) can be entered, wherein the comment is also processed as nodes (11), edges (12) and weights of the graph and embedded in its structure, and is taken into account by the artificial intelligence in step d.

7. Method according to one of the preceding claims, wherein the graph is output as a human-readable graphic (G), where preferably the graphic (G) is generated by an artificial intelligence.

8. Method according to any one of the preceding claims, wherein at least one of the features (M1 ,M2,M3,M4) of the graph is provided with reference symbols, wherein the reference symbols are assigned to a word feature (13) and originate from the graph, wherein it is preferably entered by a user which word features (13) receive a reference symbol, in particular, suggestions are made to the user in step a. as to which word feature (13) should receive a reference mark.

9. Method according to claim 7 or claim 8, wherein When generating the graphic (G), the information density is prioritized and reduced if necessary, and at least one, preferably exclusively, of the following prior information is used by the artificial intelligence: an explanation and / or specification entered by a user, preferably in written form (5) or as an image; - an explanation and / or specification independently discovered by the artificial intelligence; and a separate pre-generated patent application body (1) , preferably with a graph generated according to step b.

10. Method according to any one of the preceding claims, wherein the text is output as a text body of a patent application body (1 ) preferably the output of the text is performed algorithmically using a template, and / or The preferred method is to output the text using artificial intelligence in a free-form format, incorporating stylistic specifications from a user.

Images