Device for the machine-assisted answering of complex questions and method for same

Abstract

The invention relates to a device for the machine-assisted answering of complex questions, in particular questions that are complex in terms of tax law, comprising: at least one large language model, LLM, for detecting contextual and semantic relationships and generating coherent texts based thereon, a database separate from the LLM for storing subject-matter information associated with the question, a vector database separate from the LLM for storing and retrieving general knowledge relating to the question, a user interface for interacting with a user of the device, and a control unit for communicating with the LLM, the database, the vector database and the user interface, in particular for controlling, receiving and Passing on input or output between the LLM, the database, the vector database and the user interface, wherein the control unit is designed to control the LLM, the database, the vector database and the user interface according to a predefined procedure, and to thereby receive and pass on any respective generated outputs.

Description

[0001] 02313-25 La / Se

[0002] PricewaterhouseCoopers GmbH Wirtschaftsprüfungsgesellschaft

[0003] Device for machine-assisted answering of complex questions and methods for this purpose

[0004] The present invention relates to a device for machine-assisted answering of complex questions and a method for this purpose.

[0005] In today's digital age, artificial intelligence (AI) and machine learning are gaining increasing importance, transforming how we process information and make decisions. One of the most remarkable developments in this field is generative AI, which enables precise and efficient answers to questions. By utilizing advanced models such as Large Language Models (LLMs), generative AI systems can understand and leverage contextual and semantic relationships to deliver relevant and informed answers.

[0006] The problem, however, is that for complex issues, such as those with a complex legal or tax background, the currently available generative master's degrees (LLMs) do not provide satisfactory answers. Instead, they produce work that is both regularly flawed and presents results in a highly abbreviated manner. For example, consider a written report from a tax advisor or lawyer to a client, in which the advisor or lawyer details and assesses a complex (tax) legal issue related to a specific set of facts and perhaps also offers a (tax) legal recommendation. Passing this report on to an LLM yields unsatisfactory results that, as mentioned above, are flawed and overly simplistic.Such an answer generated using an LLM offers little or no added value due to the professionally mandated requirement for comprehensive client clarification, meaning that the potential of generative logic or LLMs available from the state of the art for answering a complex (tax) legal question cannot be utilized.

[0007] Thus, creating a satisfactory answer to a complex question remains an activity that can only be performed by a human being, since conventional LLMs do not possess the capacity to generate comparably differentiating answers in the course of creating the answer.

[0008] The objective of the present invention is to provide a device or method with which the problems listed above can be overcome or mitigated.

[0009] This is achieved using a device according to claim 1 and a method according to claim 8.

[0010] According to the invention, a device for machine-assisted answering of complex questions, in particular complex (tax) legal questions, is provided, comprising at least a Large Language Model (LLM) for capturing contextual and semantic relationships and generating coherent texts based thereon, a separate database for storing factual information pertaining to the question and preferably legal texts and / or outline templates, a separate vector database for retrieving general expert knowledge concerning the question, a user interface for interacting with a user of the device, and a control unit for communicating with the LLM, the database, the vector database and the user interface, in particular for controlling, receiving and forwarding input.The output between the LLM, the database, the vector database and the user interface includes the control unit being designed to control the LLM, the database, the vector database and the user interface according to a predefined schedule and thereby receive and forward any respective outputs generated.

[0011] The structure according to the invention, comprising the control unit which is capable of controlling the other components of the device, namely the LLM, the vector database, the database and the user interface, in a specific manner according to a sequence plan and of receiving and forwarding the outputs generated by the respective components, forms the basis for the ability to answer complex questions using LLMs.

[0012] On the one hand, the vector database reflects the background of the complex issue, for example by storing information on legal restrictions or requirements, such as specialist literature, legal texts or specially created specialist information, and using it according to the specifications of the control unit by the LLM.

[0013] Sticking with the example of a complex tax law issue, the vector database stores relevant tax laws, pertinent commentaries on those laws, and procedures or special features already developed for specific questions. This allows for the creation of a vector database with a multitude of text files, so-called "text chunks," whose subject-matter background can correlate with the background of a potentially complex question to be solved. In a traditional database, on the other hand, the information specific to the question is stored in order to describe the situation in which the question arises as comprehensively as possible. This information can, for example, include...in the form of emails, other correspondence, contracts, transcripts or notes of conversations, and other legal documents, all of which describe or supplement the facts that have a direct connection to the facts of the complex issue.

[0014] For example, if the complex question to be answered is whether two companies can merge at book value, the underlying merger agreement between the two companies would be an obvious piece of information relevant to this question. Of course, additional information communicated via email could also supplement the facts, and this email could therefore also be included in the database. Similarly, notes relating to the facts of a meeting can also be stored in the database in a machine-readable format. The database helps to record the facts in a structured manner and to create a comprehensive picture of the situation within which the complex question to be resolved is situated.

[0015] Furthermore, the database is primarily used to store legal texts and / or outline templates.

[0016] Based on the structure according to the invention, the LLM is not left to answer the question alone, as the control unit ensures that the LLM, according to a flowchart contained within the control unit, addresses the complex question to be solved according to a specific scheme, using the specific information stored in the vector database and the database. This surprisingly significantly improves the answer quality. According to the invention, the control unit is designed to control the LLM, the database, the vector database, and the user interface according to a predefined flowchart, thereby receiving and forwarding any respective outputs generated.

[0017] Using a structured workflow, the LLM can be instructed to analyze the complex question successively, e.g. by answering sub-points that subdivide the question.

[0018] The process plan is independent of the specific complex question and should therefore be seen as a scheme that reveals the actual performance of the LLM in conjunction with the vector database and the database.

[0019] According to a further optional modification of the present invention, it can be provided that the predetermined flow plan of the control unit is designed to subdivide a question from the user received via the user interface.

[0020] The control unit either accesses standardized templates stored in the database to categorize specific questions or subject areas. The selection is made by the control unit based on the question submitted via the user interface, additional information selected by the user, and relevant information from the database. The standardized template can be manually adapted by the user or by the LLM (Learning Management Lab) for the specific question.

[0021] Alternatively, the LLM performs the subdivision based on the question received via the user interface and information from the vector database and / or information from the database.

[0022] The LLM creates the subdivision of the complex issue into sub-points by referring to the vector database, for example, by accessing selected legal regulations stored in the vector database or the database itself. If the LLM deduces that further sub-points arise from this, these sub-points are added, thereby achieving the most complete possible subdivision of the complex issue.

[0023] The subdivision can of course also be influenced by further information, e.g. by means of outline templates that are stored in the database for a specific or frequently occurring complex question.

[0024] If the outline has been created by the LLM, it can be manually adjusted by the user after an optional further development of the present invention with regard to the question or the underlying facts, e.g. to set certain priorities in the examination of the question.

[0025] After a further optional modification, it can advantageously be provided that the specified flow plan of the control unit is further designed to summarize the factual information stored in the database with regard to the question or for each of the sub-items using the LLM, whereby the summarization includes restricting to information relevant to the question or the respective sub-item of the factual information by the LLM comparing the semantic and content-related similarity of the sub-item and / or the question with specialist knowledge stored in the vector database and, based on this, categorizing the factual information and restricting it to relevant information for further processing.

[0026] The selection of sources from the vector database is therefore based on their semantic or content-related similarity to the research question and / or the established outline with its sub-points. To determine this similarity, the texts (both the texts stored in the vector database, the "text chunks," and the research question itself) can be converted into embeddings, which, in a highly vector-based space, allow for a simple comparison of semantic or content-related similarities.

[0027] According to a further optional modification of the present invention, it can therefore be provided that the determination of the expert knowledge used for a comparison from the vector database includes converting the question and / or the sub-item into embeddings, word embeddings, preferably in order to determine a similarity with expert knowledge stored in the vector database about determining a distance vector.

[0028] Furthermore, according to an advantageous embodiment of the present invention, it can be provided that the predetermined flow chart of the control unit is designed to allow the LLM to answer the sub-items of the subdivided question by referring to the summarized factual information, preferably wherein suitable specialist knowledge for a respective sub-item is identified by a similarity comparison with the corresponding sub-item to content in the vector database and used by the LLM.

[0029] Once the outline for answering the question has been created, each individual sub-point of the outline is answered separately by the LLM, drawing on the relevant specialist knowledge stored in the vector database.

[0030] By providing a subdivision with individual sub-points and answering the sub-points separately, it is possible to consider in detail the individual aspects required to answer the question.

[0031] Furthermore, according to the present invention, it can be provided that the answers to the individual sub-questions, which have been generated separately, are used as the basis for creating the answer to the question. Thus, it is provided that the LLM, with knowledge of the answers to the individual sub-questions, generates a response to the question in a further step. This step ensures that the answer to the question takes into account all aspects deemed relevant in the assessment of the question and delivers an answer of high substantive quality.

[0032] According to a further advantageous embodiment of the present invention, a second LLM may be provided which is independent of the first LLM and communicates with the control unit via a data connection in order to check and assess the quality of an answer to the question or part of the question generated with the help of the first LLM, in particular to find any inconsistencies in an answer of the first LLM.

[0033] To check the quality of the generated answer to the question, the first LLM (Large Learning Model) is used, as this often identifies inconsistencies. Additionally, a different LLM, which was not used to generate the answer, can be used to check the delivered answer for inconsistencies or contradictions and to further improve its quality.

[0034] Furthermore, it may be provided that any inconsistencies or contradictions found in the answer generated by the first LLM to the question can be presented to the user via the user interface. For example, this can be done by displaying the discovered inconsistencies in text form.

[0035] According to a further advantageous embodiment of the present invention, the control unit can have a separate data connection to each of the at least one LLM, the database, the vector database, and the user interface, and the LLM, the database, the vector database, and the user interface can only communicate with each other after authorization by the control unit or only with the knowledge of the control unit. This ensures that the control unit is always aware of the state of the LLM, the database, the vector database, and the user interface, and which data or information is being exchanged between the individual units.

[0036] The invention further relates to a method for machine-assisted answering of complex questions, such as complex tax law questions, in particular with a device according to one of the aspects set out above, wherein the method comprises the steps:

[0037] Answering the question by executing a predefined sequence of actions: a control unit for controlling, receiving and forwarding outputs from an LLM for capturing contextual and semantic relationships and generating coherent texts based on them; a database for storing factual information related to the question; a vector database for retrieving general expert knowledge concerning the question; and a user interface for interacting with a user of the device.

[0038] By implementing a workflow, the LLM enables a structured processing of the questions to be answered, leading to significantly improved answer quality. By controlling the LLM, the vector database containing the subject-matter information, and the database containing the information describing the situation, the control unit can structure the question-answering process according to the steps outlined in the workflow. This structure allows for monitoring the path to answering the question and enabling early verification of whether the necessary foundation for answering the question is sufficient or whether further information is required.The method according to the invention, just like the device according to the invention described above, makes it possible to significantly increase the quality of the answer to the question and at the same time to ensure a structure in answering the question that is typically only achieved by experienced professionals familiar with the subject matter, for example a certified tax advisor with several years of professional experience in the field of tax law issues.

[0039] According to an optional further development of the inventive method, it can be provided that, according to the flow plan, a question from the user received via the user interface is subdivided into sub-items by the LLM on the basis of information from the vector database for retrieving general expert knowledge concerning the question and / or information from the database for storing factual information belonging to the question.

[0040] To subdivide the research question, the general specialist knowledge stored in the vector database is used. This database may contain, for example, theoretical treatises on a complex legal issue, legal texts, commentaries, or even existing subdivisions for specific questions. The LLM searches the vector database for content that is thematically related to the research question, analyzes it, and uses this analysis to create a structured outline of the research question with sub-points.

[0041] The question, which is subdivided into sub-points, is used in the further course of answering the question, for example by assigning the relevant facts from the database containing the factual information to a respective sub-point.

[0042] Furthermore, according to the present invention, it can be provided that the factual information stored in the database is summarized into a factual representation with regard to the question or for each of the sub-points by means of the LLM, wherein, during the summarization, the factual information is limited to information relevant to the question or to the respective sub-point by the LLM comparing the semantic and content-related similarity of the sub-point and / or the question with expert knowledge stored in the vector database and, on the basis of this, restricting it to information relevant to the sub-point.

[0043] The generated sub-items are thus linked to the relevant factual information by summarizing the factual information stored in the database for each sub-item. The LLM then checks the thematic similarity of the factual information to the respective sub-item or to the relevant specialist knowledge from the vector database.

[0044] As a result, the question is subdivided into numerous sub-points, with the relevant factual information known for each individual sub-point. For each sub-point, the totality of factual information is thus limited to the components relevant to that specific sub-point, presented in a factual description assigned to that sub-point.

[0045] According to a further optional modification of the present invention, it can be provided that, in accordance with the flowchart, the LLM checks, after or during the summarization of the factual information for each sub-item, whether any information in the case file stored in the database is contradictory or missing. If this is the case, the LLM outputs a report containing the contradictory or missing information in the case file to enable the factual information stored in the database to be supplemented or the case file to be revised to produce a consistent and complete case file. The user can be notified via the user interface that further factual information is required to answer the question and can request this information from the user. If this information is provided by the user via the user interface, the LLM then...Once the information is stored in the relevant database, the LLM checks whether any contradictory statements of facts remain or whether any factual information is missing. If this is not the case, the process of answering the question can continue.

[0046] An optional modification of the present invention allows the LLM to analyze the question to determine whether it is complete with respect to the generated factual representation, or whether typical questions related to the situation are missing. This is preferably done by converting the question into embeddings and comparing it with the information stored in the vector database. If missing questions are detected, these are suggested to the user as an extension of the existing question. The user then has the option of accepting the question suggested by the LLM or proceeding with the unchanged question. If the user accepts the suggested question, the workflow involves repeating the subdivision of the modified question and subsequently processing the facts.

[0047] Furthermore, it can be checked whether the question correlates with the provided factual information or the generated factual representation. If a discrepancy arises with the question to be answered, and if the factual information or the generated factual representation suggests a further question, the LLM can propose this further question to the user, allowing them to decide whether to adjust the question accordingly. Thus, based on the provided factual information or the generated factual representation, the LLM recognizes the questions typically asked based on this information and compares them with the question to be answered. If there are differences or aspects that the question to be answered does not cover, these can be suggested to the user, giving them the opportunity to modify the question.

[0048] According to a further optional modification of the present invention, it can be provided that the LLM answers the question by answering each of the individual sub-items of the subdivided question separately, using only the information in the vector database that exceeds a threshold value with respect to similarity to the respective sub-item, preferably wherein a similarity of the information in the vector database and the sub-item is determined by means of embeddings, and the LLM generates the answer to the question from the answers to the individual sub-items.

[0049] The question is therefore primarily answered by addressing the sub-questions of the structured question. Based on these answers, the LLM then generates the final answer. Using the facts summarized for each sub-question and the specialist knowledge stored in the vector database, the LLM answers each sub-question. Once all sub-questions, which together form the subdivision of the question, have been processed, an answer to the question is provided based on all the answers to the sub-questions. The actual answer to the question is thus based on the answers to the individual sub-questions that subdivide the question and are subsequently used to generate the answer to the question.

[0050] According to an advantageous embodiment of the present invention, it can be provided that the answer to the question generated with the aid of the LLM is checked by the first LLM or a second LLM independent of the first LLM in order to obtain an assessment of the quality of the answer to the question generated with the aid of the first LLM and to find any inconsistencies in an answer of the first LLM.

[0051] To assess the extent to which the answer generated by the first LLM contains inconsistencies, the first LLM performs a review to increase the reliability of its answers, which were generated using the vector database. The additional use of a second LLM, separate from the first, to review the first LLM's answers further enhances the quality of the responses.

[0052] To ensure this, one can, for example, work with the first LLM or with a completely different LLM.

[0053] According to a further advantageous embodiment of the present invention, it can be provided that, after the user has submitted the question via the user interface, a list of required factual information is generated that is necessary to answer the question, preferably wherein the list is generated by processing information stored in the vector database by the LLM, which has a similarity value to the question exceeding a threshold, in particular by converting the question into embeddings and performing a similarity comparison with the information stored in the vector database with regard to content and / or semantics.

[0054] Thus, after a question is submitted via the user interface, the LLM can generate a list, drawing on the specialist knowledge stored in the vector database, outlining the factual information required to answer the question. This requirements list is particularly helpful when requesting and obtaining the relevant factual information. According to a further modification of the present invention, the LLM can be configured to compare the generated requirements list with the factual information submitted via the user interface to determine whether further factual information is required to answer the question.

[0055] This ensures, through the LLM, that the information base for answering the question is based on sufficiently sound factual information.

[0056] According to a further modification of the present invention, it can be provided that the control unit's sequence of actions performs at least one, any combination or all of the following steps:

[0057] Creating a requirements list regarding the necessary factual information to answer the given question,

[0058] Receiving factual information provided via the user interface and storing it in the database,

[0059] Receiving and narrowing down specialist information identified by the user as relevant to the present question via the user interface, in particular selecting specific legal or tax regulations,

[0060] Receiving a question entered via the user interface, which must be answered in relation to the provided factual information,

[0061] Creating an outline based on the relevant specialist information or legal regulations, whereby the outline can be created with sub-points as described above and can be based on standardized templates or created entirely by the LLM; adapting the created outline via the user interface based on user input to set different priorities in the preparation of the answer.

[0062] The LLM processes the provided factual information to obtain a presentation of the facts that is appropriate to the question and the structure, whereby the LLM presents only the information relevant to answering the question in an organized and concise manner, whereby the relevant factual information can be assigned to the individual sub-points as described above.

[0063] The LLM checks whether the processed description of the facts is complete and contains all the information requested in the requirements list, or whether it still has gaps that prevent the generation of an answer to a sub-item or to the question, or that reduce the answer quality, and creates a requirements list containing the missing factual information, preferably whereby this check is carried out until the missing factual information has been provided by the user via the user interface.

[0064] The LLM will review the processed presentation of the facts against all (supplemented or originally stored in the database) factual information for inconsistencies, whereby this review can be carried out according to the design outlined above.

[0065] The LLM checks whether the question to be answered corresponds to the generated factual description, or whether further additional questions relating to the factual description are obvious but not yet part of the question to be answered. If so, the LLM presents the possible extensions of the question to the user via the user interface so that the user can decide whether the additions suggested by the LLM should become part of the question to be answered. The review of any additions to the question is carried out by referring to the vector database and the information stored therein that exhibits a semantic and content-related similarity to the factual description. The LLM then answers the question based on the generated factual description.wherein each of the sub-points subdividing the question is answered individually using the vector database, and the question is then answered based on the individual answers to the sub-points in order to obtain the most comprehensive and detailed answer possible, preferably wherein the answering of the question can be carried out as described above,

[0066] The LLM will check whether there is any missing factual information that would be useful for a more complete answer to the question, and if so, issue a corresponding message to the user via the user interface and receive the missing factual information via the user interface in order to then revise the generated answer to the question based on the newly submitted factual information.

[0067] Checking the answer prepared by the first LLM for inconsistencies by a second LLM or by the first LLM who prepared the answer to the question, whereby the check can be carried out as described above.

[0068] Summarize the answer given to the question, focusing on the key statements and necessary actions, preferably in the style of an executive summary, where the output type may include, for example, Word format or PPT format.

[0069] Further features, details, and advantages of the invention will become apparent from the following description of the figures. These show:

[0070] Figure 1: a schematic representation of the structure of the device according to the invention, and

[0071] Figure 2: A sequence of steps in carrying out a possible implementation of the method according to the invention. Figure 1 shows a schematic representation of the structure of the device 1 according to the invention, in which a generative computer, implemented here as a Large Language Model, LLM 2, is used to answer a complex question, for example one with a complex tax law background.

[0072] The LLM 2 is controlled by a control unit 7, which is also connected to a user interface 5, a database 3, and a vector database 4. The user interface 5 can be a standard terminal device, such as a laptop, desktop PC, mobile device, or the like, which can be operated by a user 6 and transmit output to a user 6.

[0073] Database 3 serves to store factual information describing the circumstances in which the question arises. This factual information could include, for example, emails, contracts, machine-readable correspondence, transcripts of conversations, legal documents, or meeting notes that define the underlying facts relevant to the question. Database 3 is designed to record and store this individual factual information so that it can be processed later using LLM 2.

[0074] In addition, reference number 4 specifies a vector database 4 in which the specialist knowledge required to answer the question is stored. This vector database can contain and store various specialist texts from different sources, such as specialist literature, legal texts, court rulings, or specialist texts specifically created for particular application scenarios.

[0075] LLM 2 comprises several units, such as unit 21 for creating speech output, unit 22 for creating embeddings, and unit 23 for accessing external data. Examples of LLM 2 implementations include GPT-3 or GPT-4 from OpenAI, BERT from Google AI, RoBerta from Facebook AI, and similar technologies.

[0076] The control unit 7, which can also be referred to as a memo generator, orchestrates the interaction of the various elements of the device 1 using a sequence of steps, thereby ensuring that the LLM 2 processes the answer to the question submitted by the user 6 via the user interface 5 of the control unit 7 in a specific manner. This results in an increase in the quality of the answer, which far exceeds the expected level of response from the LLM 2.

[0077] Figure 2 shows the steps of device 1 triggered by the flow chart, which are described in detail below.

[0078] Step 1, S1:

[0079] Creating a requirements list regarding the necessary factual information to answer the given question.

[0080] The question that user 6 wants answered is transmitted via the user interface 5 of the control unit 7, whereby the LLM returns a list of required factual information in response to the question, which is necessary to answer the question to a sufficient standard.

[0081] Step 2, S2:

[0082] Receiving factual information provided via the user interface and storing it in the database, and receiving, via the user interface, the narrowing down of specialist information identified by the user as relevant to the present question, in particular the selection of specific legal or tax regulations. The user can then supply the required factual information according to the list of requirements using the user interface of device 1, whereby the individual pieces of factual information are stored in database 3.The factual information may include emails, correspondence, contracts, legal documents, transcripts or notes of conversations, as well as comparable documents or texts, whereby, using the example of a question concerning the tax implications of a merger of two companies, a typical piece of factual information is the underlying merger agreement of the two companies.

[0083] After initial storage of factual information, the LLM is designed to compare the stored factual information with the requirements list to indicate whether further input is needed.

[0084] Furthermore, the user identifies, preferably following a suggestion or specification by the LLM, the components of specialist knowledge identified as relevant to the problem to be solved. For example, this could include specific tax regulations applicable to a merger of two companies.

[0085] It may be intended that the LLM narrows down the components of the specialist knowledge deemed relevant and presents them to the user via the user interface, so that the user can make a final selection or supplement the relevant specialist knowledge.

[0086] Step 3, S3:

[0087] Receiving a question entered via the user interface, which must be answered in relation to the provided factual information.

[0088] Now the user can enter or further specify the question that needs to be answered. Step 4, S4:

[0089] Creating an outline based on the relevant specialist information or legal regulations, whereby the outline can be based on standardized templates or created entirely by the LLM, and adapting the created outline via the user interface based on user input to set different priorities in the preparation of the answer.

[0090] Based on the selected expertise, such as legal regulations or similar, the LLM 2 creates a detailed outline of the question. This outline can either be based on standardized templates stored in the database or created independently by the LLM based on the previous input. The outline includes, for example, the elements of the offense, the legal consequences, and other necessary steps relevant to answering the question. Alternatively, the LLM can use a predefined outline and adapt it to the specific circumstances of the case.

[0091] Furthermore, the created outline, which consists of numerous sub-points, can be presented to user 6 via user interface 5, allowing them to make manual changes. This enables the outline to be reduced or expanded to focus on specific aspects when examining the research question.

[0092] Step 5, S5:

[0093] The LLM processes the provided factual information to create a factual presentation that aligns with the question and structure, presenting only the information relevant to answering the question in a clear and concise manner. The data uploaded in step 2 is summarized in a factual presentation that matches the given questions and the established structure. The factual information is processed in such a way that the LLM assigns only the information deemed relevant to the question to the corresponding sub-points. This leads to better results and reduces the processing time required to create the answer to the question.

[0094] The necessary expertise is obtained from the information stored in the vector database, which contains specialist information from various sources.

[0095] The selection of specialist knowledge from vector database 4 is based on its semantic or content-related similarity to the question received in step 3, as well as on the outline and its sub-points created in step 4. Similarity is determined using embedding model 22 by converting the texts into so-called embeddings, which can also be called word embeddings. These embeddings allow for a comparison of the similarity in terms of text content. When selecting specialist knowledge from vector database 4, the components of the specialist knowledge identified as relevant by the user in step S2 can be added. Furthermore, the selected specialist knowledge from vector database 4 can be displayed to the user 6 via user interface 5, allowing the user to adjust the specialist knowledge deemed relevant as needed.

[0096] Step 6, S6:

[0097] The LLM checks whether the processed description of the facts is complete and contains all the information requested in the requirements list, or whether it still has gaps that prevent the generation of an answer to a sub-item or to the question, or that reduce the answer quality, and creates a requirements list showing the missing factual information, preferably whereby this check is carried out until the missing factual information has been provided by the user via the user interface.

[0098] If missing or contradictory information is discovered in the previous step S5, the LLM generates a document describing the missing information regarding the matter. Advantageously, the content of this document is structured in list form, allowing the generated document to be used directly or with only minor manual changes in communication between user 6 and third parties who can provide the missing information.

[0099] If the user responds to such a request to provide further factual information by making the additional factual information available via user interface 5, step S5 is executed again. If the check in step S6 then shows that all information is presented consistently in the factual description, the procedure proceeds to step S7, described below.

[0100] Step 7, page 7:

[0101] The LLM will review the processed account of the facts against all (supplemented or originally stored in the database) factual information for inconsistencies.

[0102] In this step, the LLM 2 checks the factual description created in steps S5 and S6 for inconsistencies, taking into account all the factual information entered in step S2. A corresponding value regarding inconsistencies between the factual description created in steps S5 and S6 and the totality of factual information can be displayed to user 6 via user interface 5, allowing the user to decide whether to make changes to the factual description.

[0103] Step 8, S8: The LLM checks whether the question to be answered corresponds to the generated factual representation, or whether further additional questions relating to the factual representation are obvious but not yet part of the question to be answered. If so, the LLM presents the possible extensions of the question to the user via the user interface, allowing the user to decide whether the additions suggested by the LLM should become part of the question to be answered. The verification of a question addition is carried out by accessing the vector database and the information stored therein that exhibits semantic and content-related similarity to the factual representation. Determining the similarity in terms of semantics and content is done using so-called embeddings, which are generated by the LLM.These embeddings are high-dimensional vectors that represent text in vector space. High similarity is inferred if the distance between the vectors of two texts in vector space falls below a certain threshold.

[0104] If the user decides that they wish to expand the question as suggested, the procedure goes back to step S4, so that a subdivision adapted to the new question is carried out and the procedure is repeated starting from step S4.

[0105] Step 9, page 9:

[0106] The LLM answers the question based on the generated factual description, whereby each of the sub-points subdividing the question is answered individually using the vector database, and the answer to the question is then based on the individual answers to the sub-points in order to obtain the most comprehensive and detailed answer possible.

[0107] Based on the description of the facts created in steps S5, S6, and S7, each sub-point of the subdivided question is answered by the LLM (Learning Management Module). Only text passages from vector database 4 that directly relate to the sub-point are used during the answer creation process to achieve the most comprehensive and detailed answer possible. This ensures that the relevant sources of expertise from vector database 4 are used for each sub-point, resulting in a particularly high-quality answer for that sub-point.

[0108] Only after the LLM has fully answered all the sub-questions that subdivide the question will an answer to the question be generated based on the answers provided for the sub-questions. Thus, the LLM generates the answer to the question based on the results or answers to the individual sub-questions.

[0109] Step 10, page 10:

[0110] The LLM will check whether there is any missing factual information that would be useful for a more complete answer to the question, and if so, issue a corresponding message to the user via the user interface and receive the missing factual information via the user interface in order to then revise the created answer to the question based on the newly submitted factual information.

[0111] If, after reviewing the various outline points, any information regarding the facts of the case is still missing when answering the question, user 6 is notified and asked to provide the relevant information via user interface 5. Once the additional factual information is received via the user interface, the process returns to step S5 and is repeated from there. Upon reaching step S10 again, an updated answer to the question is obtained based on the newly added factual information.

[0112] Step 11, S11: Checking the answer created to the question for inconsistencies by the first LLM or a second LLM that is different from the first LLM that created the answer to the question.

[0113] In this step, the quality of the answer generated for the question is checked using either the first LLM or a second LLM. The second LLM can be another instance of the first LLM model, but the invention also encompasses the use of a different LLM model from a different provider for quality control. For example, the first LLM, which generated the answer to the question, could be operated by OpenAI, and the second LLM by Google AI.

[0114] The check for inconsistencies in the answer to the question generated by the second LLM can, for example, be a user-friendly metric between 1 and 100, which allows the user to easily assess the quality of the answer to the question.

[0115] Step 12, page 12:

[0116] Summarize the answers given to the question, focusing on the key statements and necessary actions, preferably in the style of an executive summary, where the output type may include, for example, Word format or PPT format, or outputs in both formats.

[0117] The detailed answer to the question is summarized to succinctly present the key findings and necessary actions. Appendix with implementation example.

[0118] The following will demonstrate that a device operating according to the principle of the invention for machine-assisted answering of complex questions, or a device operating according to the method of the invention, is superior to an LLM, OpenAI GPT-40, which was widely used at the time of filing. The specific advantages of the invention are particularly evident in the level of detail of the analysis, the definition of the legal basis, the completeness, and the structure of the answers.

[0119] The invention is characterized by an in-depth, comprehensive analysis and answering of legal and tax-related questions that goes far beyond the superficial answers generated by conventional LLMs. It develops answers characterized by a precise and detailed understanding of the relevant legal principles. Thus, the method according to the invention ensures a complete capture and presentation of all relevant legal aspects and potential consequences.

[0120] Internal tests showed that the answers of the invention contained 53% to 61% of the expected content of a sample solution across eleven different test scenarios, whereas only 28% could be achieved with pure LLM without application of the principle according to the invention.

[0121] Furthermore, the answers generated by the invention are clearly structured and logically organized, which allows for easy comprehension and application of the information.

[0122] Input (identical for invention and OpenAI GPT-4o):

[0123] Müller GmbH is a corporation subject to unlimited tax liability in Germany. Müller GmbH produces and markets furniture. It has a loss carryforward subject to continuation requirements pursuant to Section 8d of the German Corporation Tax Act (KStG). In the 2022 tax year, Müller GmbH acquires a shoe production and marketing business from Meier GmbH. Following the acquisition, Müller GmbH continues to operate this shoe production and marketing business alongside its furniture production and marketing activities. What impact does the acquisition of this business have on Müller GmbH's use of the loss carryforward subject to continuation requirements?

[0124] Output Invention:

[0125] memorandum

[0126] Subject: Tax implications of Müller GmbH's acquisition of an additional business operation on the loss carryforward subject to continuation requirements under Section 8d of the German Corporation Tax Act (KStG).

[0127] 1. Forfeiture of a loss carryforward subject to continuation requirements pursuant to Section 8d Paragraph 2 of the German Corporation Tax Act (KStG) - Principle

[0128] The loss carryforward provision under Section 8d of the German Corporation Tax Act (KStG) allows corporations to continue claiming losses for tax purposes despite a detrimental acquisition of a shareholding as defined in Section 8c KStG. However, this provision is subject to strict conditions, which are detailed in Section 8d, paragraphs 1 and 2 KStG. According to Section 8d, paragraph 1, sentence 1 KStG, at the corporation's request, Section 8c KStG may be disregarded after a detrimental acquisition of a shareholding if the corporation has operated exclusively the same business since its incorporation or at least since the beginning of the third assessment period before the detrimental acquisition and no event as defined in Section 8d, paragraph 2 KStG has occurred during this period.

[0129] The forfeiture of loss carryforwards subject to business continuation requirements is regulated in Section 8d Paragraph 2 of the German Corporation Tax Act (KStG). According to Section 8d Paragraph 2 Sentence 1 KStG, loss carryforwards subject to business continuation requirements are forfeited if the business operations within the meaning of Section 8d Paragraph 1 KStG are discontinued. Furthermore, Section 8d Paragraph 2 Sentence 2 KStG provides for additional detrimental events that also lead to the forfeiture of loss carryforwards subject to business continuation requirements. These are:

[0130] 1. Suspension of business operations,

[0131] 2. Transfer of the business operations to a different purpose,

[0132] 3. Commencement of an additional business operation,

[0133] 4. Participation of the corporation in a co-entrepreneurship,

[0134] 5. Assuming the position of a controlling entity within the meaning of Section 14 Paragraph 1 of the German Corporation Tax Act (KStG),

[0135] 6. Transfer of business assets to the corporation, which values ​​them at a lower than their fair market value. 2. Occurrence of a detrimental event as defined in Section 8d Paragraph 2 of the German Corporation Tax Act (KStG).

[0136] 2.1. Definition of the term "business operation"

[0137] The concept of a business operation in the context of Section 8d of the German Corporation Tax Act (KStG) is of central importance. According to Section 8d Paragraph 1 Sentence 3 KStG, a business operation comprises "the corporation's sustainable, mutually complementary, and mutually reinforcing activities, driven by a unified profit motive, and is determined by qualitative characteristics in an overall assessment." These characteristics include, in particular, the services or products offered, the customer and supplier base, the markets served, and the qualifications of the employees.

[0138] 2.2. Cessation of business operations

[0139] The cessation of business operations is a key detrimental event within the meaning of Section 8d Paragraph 2 Sentence 1 of the German Corporation Tax Act (KStG), which leads to the forfeiture of the loss carryforward subject to continuation requirements. A cessation occurs when the corporation completely and permanently abandons its economic activity and no longer pursues the intention of generating profit.

[0140] 2.3. Suspension of business operations

[0141] The suspension of business operations is, according to Section 8d Paragraph 2 Sentence 2 No. 1 of the German Corporation Tax Act (KStG), a detrimental event that leads to the forfeiture of the loss carryforward subject to continuation requirements. A suspension of operations occurs when the corporation temporarily ceases its economic activity but intends to resume business operations at a later date.

[0142] 2.4. Transfer of business operations to a different purpose

[0143] The transfer of business operations to a different purpose is a detrimental event pursuant to Section 8d Paragraph 2 Sentence 2 No. 2 of the German Corporation Tax Act (KStG). A different purpose exists if the business operations undergo a significant change in their economic activity or actual business purpose.

[0144] 2.5. Commencement of an additional business operation

[0145] The commencement of an additional business operation is a detrimental event according to Section 8d Paragraph 2 Sentence 2 No. 3 of the German Corporation Tax Act (KStG). Such a commencement occurs when the corporation undertakes a new, qualitatively different economic activity that does not fit into its existing business operations.

[0146] 2.6. Participation of the corporation in a co-entrepreneurship

[0147] Participation in a co-entrepreneurship is a detrimental event according to Section 8d Paragraph 2 Sentence 2 No. 4 of the German Corporation Tax Act (KStG). Participation exists if the corporation, as a co-entrepreneur, shares in the income and losses of the partnership and bears co-entrepreneurial risk and initiative.

[0148] 2.7. Assumption of the position of a controlling entity within the meaning of Section 14 Paragraph 1 of the German Corporation Tax Act (KStG) by the corporation

[0149] Assuming the position of a controlling entity within the meaning of Section 14 Paragraph 1 of the German Corporation Tax Act (KStG) is a detrimental event pursuant to Section 8d Paragraph 2 Sentence 2 Number 5 of the KStG. A controlled entity exists if a corporation (controlling entity) and another corporation (controlled entity) are linked by a profit transfer agreement and the other requirements of Section 14 of the KStG are met.

[0150] 2.8. Transfer of assets to the corporation, which values ​​them at a lower than their fair market value.

[0151] The transfer of business assets at a price lower than their fair market value is a detrimental event according to Section 8d Paragraph 2 Sentence 2 No. 6 of the German Corporation Tax Act (KStG). According to Section 9 of the German Valuation Act (BewG), fair market value is the price that could be obtained in the ordinary course of business upon sale.

[0152] 3. Legal consequences of the occurrence of a harmful event pursuant to Section 8d Paragraph 2 of the German Corporation Tax Act (KStG).

[0153] 3.1. Extent of the forfeiture of the loss carryforward tied to business continuation

[0154] The scope of the forfeiture of a loss carryforward subject to continuation requirements following the occurrence of a detrimental event pursuant to Section 8d Paragraph 2 of the German Corporation Tax Act (KStG) is comprehensive and encompasses all losses determined up to the time of the detrimental event. The forfeiture is retroactive to the end of the preceding assessment period in which the detrimental event occurred.

[0155] 3.2. Application of the hidden reserves clause

[0156] The hidden reserves clause pursuant to Section 8d Paragraph 2 Sentence 1 of the German Corporation Tax Act (KStG) in conjunction with Section 8c Paragraph 1 Sentences 5 to 8 of the KStG stipulates that the forfeiture of losses is suspended in whole or in part to the extent that the hidden reserves in the business assets cover the loss carryforward. This means that the hidden reserves contained in the business assets must be taken into account when determining the loss carryforward subject to continuation requirements.

[0157] 3.3. Loss of a continuation-linked interest carryforward

[0158] The forfeiture of the loss carryforward subject to continuation requirements under Section 8d Paragraph 2 of the German Corporation Tax Act (KStG) also includes the forfeiture of the interest carryforward subject to continuation requirements. This means that the previously accumulated non-deductible interest expenses, which were carried forward as interest expenses, can no longer be used to offset future profits.

[0159] Result

[0160] For Müller GmbH, which has a loss carryforward subject to continuation requirements under Section 8d of the German Corporation Tax Act (KStG), the acquisition of the shoe production and marketing business from Meier GmbH has significant tax consequences. The commencement of the additional business operations constitutes a detrimental event within the meaning of Section 8d Paragraph 2 Sentence 2 No. 3 KStG, resulting in the forfeiture of the loss carryforward subject to continuation requirements. This means that Müller GmbH can no longer use the previously accumulated losses to offset future profits.

[0161] To avoid the forfeiture of the loss carryforward tied to business continuation, Müller GmbH could consider spinning off the new business operations into a separate company. However, this requires a thorough tax and legal review to ensure that no other detrimental events within the meaning of Section 8d Paragraph 2 of the German Corporation Tax Act (KStG) occur.

[0162] In summary, the acquisition of the shoe production and marketing business by Müller GmbH will result in the forfeiture of the loss carryforward subject to continuation requirements under Section 8d Paragraph 2 Sentence 2 No. 3 of the German Corporation Tax Act (KStG). Müller GmbH should therefore carefully examine the tax implications and, if necessary, consider alternative structures to preserve the loss carryforward. Output OpenAI GPT-4o

[0163] The acquisition of the shoe production and marketing business by Müller GmbH has potential implications for the use of the loss carryforward linked to business continuation under Section 8d of the German Corporation Tax Act (KStG). Section 8d KStG allows for the continued use of a loss carryforward even after a detrimental acquisition of a shareholding, provided certain conditions are met.

[0164] The key points to consider are:

[0165] 1. **Continuation of business operations (§ 8d para. 1 sentence 1 KStG)**:

[0166] - The loss carryforward can only be continued if the business that caused the loss continues uninterrupted. The acquisition of a new business does not generally affect its continuation, as long as the original business (here, the production and marketing of furniture) continues unchanged.

[0167] 2. **Damaging events (§ 8d para. 2 KStG)**:

[0168] - No harmful event may occur. A harmful event could be, for example, the cessation of the original business operations or the commencement of a new business operation.

[0169] - According to Section 8d Paragraph 2 Sentence 1 No. 1 KStG, a harmful event occurs if the business operations are discontinued or the nature of the business operations is changed.

[0170] - The acquisition and continuation of a new business (here the business for the production and marketing of shoes) could be regarded as the start of a new business operation and thus as a harmful event within the meaning of Section 8d Paragraph 2 Sentence 1 No. 2 KStG.

[0171] 3. **Overall activities of the company**:

[0172] - It depends on the company's overall activities. If Müller GmbH experiences a significant change in its business activities through the acquisition of the new plant for the production and marketing of shoes, this could jeopardize the use of the loss carryforward.

[0173] **Conclusion:** The acquisition of the business for the production and marketing of shoes could be considered a detrimental event within the meaning of Section 8d of the German Corporation Tax Act (KStG), as this can be interpreted as the commencement of a new business operation. This could impair the utilization of the loss carryforward linked to business continuation. It would be advisable to conduct a thorough examination of the circumstances and possibly seek tax advice in order to determine the specific implications and utilize any available structuring options.

[0174] Reference symbol list:

[0175] 1 Device for machine-assisted answering of complex questions

[0176] 2 LLM, Large Language Model as a form of generative learning

[0177] 3. Database for storing factual information

[0178] 4. Vector database for storing and retrieving general specialist knowledge

[0179] 21 Unit for creating the speech output, Language Model

[0180] 22 Unit for creating embeddings, Embeddings Model

[0181] 23 Unit for accessing external data, for example, connection to the internet

Claims

02313-25 La / Se PricewaterhouseCoopers GmbH Wirtschaftsprüfungsgesellschaft Device for machine-assisted answering of complex questions and methods for this purpose Patent claims 1. Device (1) for machine-assisted answering of complex questions, in particular complex tax law questions, comprising: at least one Large Language Model (2), LLM, for capturing contextual and semantic relationships and generating coherent texts based thereon, a separate database (3) for storing factual information pertaining to the question and preferably legal texts and / or outline templates, a separate vector database (4) for storing and retrieving general expert knowledge concerning the question, a user interface (5) for interacting with a user (6) of the device (1), and a control unit for communicating with the LLM (2), the database (3), the vector database (4), and the user interface (5), in particular for controlling, receiving, and forwarding input.Output between the LLM (2), the database (3), the vector database (4) and the user interface (5), wherein. the control unit is designed to control the LLM (2), the database (3), the vector database (4) and the user interface (5) according to a predefined sequence and thereby receive and forward any respective outputs generated.

2. Device (1) according to the preceding claim 1, wherein the predetermined flow plan of the control unit is designed to subdivide a question from the user (6) received via the user interface (5) by the LLM (2) into sub-items based on information from the vector database (4) and / or information from the database (3).

3. Device (1) according to the preceding claim 2, wherein the predetermined sequence of events of the control unit is further designed to summarize the factual information stored in the database (3) with regard to the question or for each of the sub-items by means of the LLM (2) into a factual representation, in particular wherein the summarization includes a restriction to information of the factual information relevant to the question or the respective sub-item, by the LLM (2) comparing the semantic and content-related similarity of the sub-item and / or the question with expert knowledge stored in the vector database (4) and on the basis of this a restriction to information relevant to the sub-item.

4. Device (1) according to claim 3, wherein the determination of the expert knowledge from the vector database (4) used for comparison comprises converting the question and / or the sub-item into embeddings, word embeddings, preferably to determine a similarity with expert knowledge stored in the vector database (4) about determining a distance vector.

5. Device (1) according to one of the preceding claims 3 or 4, wherein the predetermined flow chart of the control unit is further designed to select the sub-items of the subdivided question by referring to the summarized to have factual information answered by the LLM (2), preferably whereby for this purpose suitable specialist knowledge for a respective sub-item is identified by a similarity comparison with the corresponding sub-item to content in the vector database (4) and used by the LLM (2).

6. Device (1) according to one of the preceding claims, further comprising: a second LLM, which is independent of the first LLM (2) and has a The data connection communicates with the control unit to check and assess the quality of an answer to the question or part of the question generated by the first LLM (2), and preferably to detect any inconsistencies in an answer from the first LLM (2).

7. Device (1) according to one of the preceding claims, wherein the control unit has a separate data connection to the at least one LLM (2), the database (3), the vector database (4), and the user interface (5), and the LLM (2), the database (3), the vector database (4), and the user interface (5) perform communication with each other only with the knowledge of the control unit.

8. Method for machine-assisted answering of complex questions, such as complex tax law questions, in particular with a device (1) according to one of the preceding claims, wherein the method comprises the steps: Answering the question by executing a predefined process plan of a control unit for controlling, receiving and forwarding outputs of an LLM (2) for capturing contextual and semantic relationships and generating coherent texts based thereon, a database (3) for storing factual information pertaining to the question and preferably legal texts and / or outline templates, a vector database (4) for storing and retrieving general knowledge relating to the question, and a user interface (5) for interacting with a user (6) of the device (1).

9. Method according to the preceding claim 8, wherein, according to the flowchart, a question from the user (6) received via the user interface (5) is subdivided into sub-items by the LLM (2) on the basis of information from the vector database (4) for retrieving general expert knowledge concerning the question and / or information from the database (3) for storing factual information pertaining to the question, and preferably the factual information stored in the database (3) is categorized and summarized for each of the sub-items by means of the LLM (2), in particular wherein, when summarizing, the factual information is limited to information relevant for the respective sub-item.by comparing the semantic and content-related similarity of the sub-item and / or question with specialist knowledge stored in the vector database (4) and, based on this, restricting the search to relevant information.

10. Method according to the preceding claim 9, wherein, according to the flow chart, the LLM (2) checks after or during the summarization whether information in the facts stored in the database (3) is contradictory or missing, wherein, if this is the case, the LLM (2) outputs a report containing the contradictory or missing information in the facts in order to enable the supplementation of the facts information stored in the database (3) or the presentation of the facts to create a consistent and complete set of facts.

11. Method according to the preceding claim 9 or 10, wherein the question is analyzed by the LLM (2) to determine whether it relates to the factual information stored in the database (3). is complete, or typical questions related to the matter are missing, preferably by converting the question into embeddings and performing a comparison with the information stored in the vector database (4), wherein, if the absence of questions is detected, these are suggested to the user (6) as an extension of the existing question.

12. Method according to any one of the preceding claims 9 to 11, wherein the LLM (2) answers the question by answering each of the individual sub-items of the subdivided question separately, using only the information in the vector database (4) that exceeds a threshold value with respect to similarity to the respective sub-item, preferably wherein a similarity of the information in the vector database (4) and the sub-item is determined by means of embeddings, and the LLM (2) generates the answer to the question from the answers to the individual sub-items.

13. Method according to any one of the preceding claims 8 to 12, wherein the answer to the question generated by the LLM (2) is checked by the first LLM (2) itself and / or a second LLM independent of the first LLM (2) in order to obtain an assessment of the quality of the answer to the question generated by the first LLM (2) and in particular to find any inconsistencies in an answer of the first LLM (2).

14. Method according to one of the preceding claims, wherein, after the user (6) has submitted the question via the user interface (5), a list of required factual information is generated that is necessary to answer the question, preferably wherein the list is generated by processing information stored in the vector database (4) by the LLM (2) which has a similarity value to the question exceeding a threshold, in particular by the question is converted into embeddings and a comparison is performed with the information stored in the vector database (4).

15. Method according to any of the preceding claims, wherein the control unit's sequence of actions effects at least one, several or all of the following steps: (51) Creating a requirements list concerning the necessary factual information to answer the given question, in particular according to the preceding claim 14, (52) Receiving factual information provided via the user interface (5) and storing it in the database (3), (52) Receiving, via the user interface (5), the narrowing down of specialist information identified by the user (6) as relevant to the question at hand, in particular the selection of specific legal or tax provisions to narrow down the information stored in the vector database (4), (53) Receiving a question entered via the user interface (5) which is to be answered in relation to the factual information provided, (54) Creating an outline based on the technical information or legal regulations deemed relevant, preferably wherein the outline can be created with the sub-items according to claim 9 and the outline can be based, for example, on standardized templates or can be created entirely by the LLM (2), (54) Adapting the created outline via the user interface (5) based on user input (6) to set different priorities in the creation of the response, (55) Processing the provided factual information by the LLM (2) in order to obtain a presentation of the facts that is appropriate to the question and the structure, wherein the LLM (2) presents only the information relevant to answering the question in an organized and concise manner, preferably whereby the allocation of the relevant Factual information regarding the individual sub-points according to claim 9 can be provided, (56) Checking by the LLM whether the processed description of the facts is complete and contains all the information requested in the requirements list or whether it still has gaps that prevent the generation of an answer to a sub-item or to the question or reduce the answer quality, in particular carried out according to claim 10, and creating a requirements list containing the missing factual information, preferably wherein this checking step is carried out until the missing factual information has been provided by the user (6) via the user interface (5), (57) Checking the processed factual account against all factual information (supplemented or originally stored in the database) for inconsistencies by the LLM, preferably in accordance with claim 10, (58) The LLM (2) checks whether the question to be answered corresponds to the generated factual presentation, or whether further additional questions relating to the factual presentation are obvious but are not yet part of the question to be answered, in particular in accordance with claim 11, and if so, presents the possible extensions of the question to the user (6) via the user interface (5) so that the user can decide whether the additions proposed by the LLM (2) should become part of the question to be answered or not, wherein the checking of an addition to the question is carried out by referring to the vector database (4) and the information stored therein, which has a semantic and content-related similarity to the factual presentation. (59) Answering the question by the LLM (2) on the basis of the generated factual representation, whereby each of the sub-items subdividing the question is answered separately using the vector database (4), and the answering of the question is then carried out starting from the individual answers to the sub-items in order to obtain the most comprehensive possible and to receive a detailed answer, preferably wherein the answer to the question is given according to claim 12, (S10) The LLM (2) checks whether there is any missing factual information that would be useful for a more complete answer to the question, and if so, issues a corresponding message to the user (6) via the user interface (5) and receives the missing factual information via the user interface (5) in order to then revise the generated answer to the question based on the newly submitted factual information, (S11) the first LLM (2) or a second LLM that is different from the first LLM (2) checks the generated answer to the question for inconsistencies, in particular wherein the check is carried out according to claim 13, and / or (S12) summarizes the answer given to the question on the key statements and necessary actions, preferably in the style of an executive summary, wherein the output type may, for example, include Word format or PPT format.

Images