System And Method for Generating Multi-stage Summary of Electronic Disclosure Document based on Item Unit

Abstract

A multi-stage summary generation system for electronic disclosure documents is provided. A multi-stage summary generation system for electronic disclosure documents according to one embodiment includes a communication unit that transmits and receives data to and from a user terminal and an external electronic disclosure system; a database that stores original disclosure data, item-specific prompt data, and summary result data; a disclosure data collection unit that collects electronic disclosure documents; a data preprocessing unit that separates the collected electronic disclosure documents into item units and converts the format; a multi-stage summary output unit that generates a first summary for each item and integrates the first summaries to generate a second summary for the entire document; a result structuring unit that converts the generated summary results into a structured data format; and a control unit that controls the operation of each component and manages the data flow.

Description

Technology Field

[0001] The present invention relates to the automatic generation of summaries of electronic disclosure documents, and specifically, to a system and method for performing a primary summary on an item-by-item basis for regulatory disclosure documents with a vast amount of information, such as SEC electronic disclosure data of U.S. listed companies and DART electronic disclosure data of Korean listed companies, and systematically generating a secondary summary of the entire document using this. Background Technology

[0003] SEC disclosures by U.S. listed companies and DART disclosures by Korean listed companies are lengthy documents spanning hundreds of pages, condensing diverse information such as financial status, business operations, risks, and management discussions into a single document. While these disclosure documents are essential sources of information for investors, analysts, and researchers, they present challenges: excessive length, complex content structures by category, and a mixture of text, tables, and images make it difficult for users to quickly identify only the information they need.

[0004] Recently, attempts have been made to generate summaries by inputting disclosure documents using Large-Scale Language Models (LLM); however, inputting the entire document verbatim often exceeds token limits due to its excessive length, while inputting only a portion carries the risk of omitting important item content. Furthermore, summarizing the entire document at once blurs the contextual distinction between items, making it prone to diluting or distorting the core message as the content is homogeneously compressed, even if important information is concentrated in specific sections.

[0005] Furthermore, if tables and images included in disclosure documents are provided to a language model without proper preprocessing, key financial figures or chart explanations are frequently not accurately reflected in the summary. Moreover, despite the fact that the summary points to be emphasized differ by item—such as business overview, financial statement explanation, and risk explanation—conventional single-prompt-based summarization methods have limitations in that they fail to adequately reflect these item-specific characteristics. Consequently, it has been pointed out that the generated summary results are difficult to utilize directly during the actual analysis and investment decision-making stages. Prior art literature

[0007] Korean Registered Patent No. 10-2576825, published September 11, 2022 The problem to be solved

[0008] To solve the aforementioned problems, the present invention aims to provide a structure capable of reliably generating summaries for long and complex electronic disclosure documents without being constrained by the input token limitations of a large-scale language model.

[0009] In addition, the present invention aims to ensure that necessary information for each item is not omitted and that a summary is generated in a form suitable for the purpose and nature of the corresponding section by applying a dedicated prompt tailored to the characteristics of each item within a public disclosure document.

[0010] In addition, the present invention aims to improve the quality and efficiency of summarization and reduce the cost of using the language model and resource consumption by generating a secondary overall document summary using the primary summary result for each item as input, instead of inputting the entire document directly into the language model.

[0011] The problems of the present invention are not limited to those mentioned above, and other unmentioned technical problems will be clearly understood by those skilled in the art from the description below. means of solving the problem

[0013] According to one embodiment for solving the above problem, a multi-stage summary generation system for item units of electronic disclosure documents includes a communication unit that transmits and receives data to and from a user terminal and an external electronic disclosure system, a database that stores original disclosure data, item-specific prompt data, and summary result data, a disclosure data collection unit that collects electronic disclosure documents, a data preprocessing unit that separates the collected electronic disclosure documents into item units and converts the format, a multi-stage summary output unit that generates a first summary for each item and integrates the first summaries to generate a second summary for the entire document, a result structuring unit that converts the generated summary results into a structured data format, and a control unit that controls the operation of each component and manages the data flow.

[0014] The above database may include original disclosure data that stores raw files in HTML, XML, and PDF formats and metadata including company name, disclosure type, and disclosure submission date; item-specific prompt data that stores prompt templates including summary guidelines and constraints optimized for the characteristics of each item; and summary result data that stores a primary summary for each item and a secondary summary for the entire document in JSON or XML format.

[0015] The above data preprocessing unit includes an item content separation module that divides the disclosure document into item units, a format conversion module that converts the content of the separated items into a text format optimized for input into a large-scale language model (LLM), and a table and image processing module that converts tables and images within the disclosure document into text form. The item content separation module identifies the start and end points of each item by analyzing the table of contents hyperlinks, HTML tags, or specific keyword patterns of the disclosure document, and the table and image processing module converts the HTML table structure into structured text in CSV, Markdown, or JSON format, and can generate text explaining the meaning of text or diagrams within the image by utilizing optical character recognition technology or an image captioning model.

[0016] The above-described multi-stage summary output unit includes an item-specific primary summary module that generates a primary summary for each item by loading a dedicated prompt mapped to the type of each item, and a document-wide secondary summary module that generates a secondary summary for the entire document by integrating the generated item-specific primary summaries. The above-described result structuring unit defines a schema that includes a top-level field containing overall document summary information and a bottom-level array field containing detailed summary information for each item, stores metadata including an item unique identifier, title, summary body text, key figures, and key entities in each item summary object, and can render the summary result in the form of a PDF report, a Word document, or HTML.

[0017] The above multi-stage summary output unit further includes a summary structure optimization module that inputs primary summaries for each item into a text embedding model to convert them into vectors and calculates semantic connection strength by calculating cosine similarity between items, and the summary structure optimization module defines a group of items whose semantic connection strength exceeds a threshold of 0.75 as a semantic cluster, and the document-wide secondary summary module generates intermediate summaries in units of the semantic clusters and then combines the intermediate summaries to generate a full summary, and the summary structure optimization module, when the total amount of tokens of the formed semantic clusters exceeds a preset threshold, sets the primary summaries of each item included in the semantic clusters as nodes and sets the cosine similarity values ​​between items as edge weights to construct a weighted graph, performs eigenvector centrality or PageRank algorithms to select the item with the highest influence score as a hub item, and can re-divide the semantic clusters into multiple sub-clusters through K-means clustering or Rubane community detection techniques with the hub item as the centroid. Effects of the invention

[0019] According to the method and system for generating item-unit multi-stage summaries of electronic disclosure documents according to one embodiment of the present invention, through a multi-stage structure in which disclosure documents are divided into item units to generate a first-stage summary and these are reconstructed to generate a second-stage overall summary, it is possible to improve the quality of the summary while mitigating the input token limitations of a large-scale language model.

[0020] Furthermore, the present invention can reduce information omission and context distortion by providing customized summaries that reflect item-specific characteristics and balanced summaries at the overall document level. In particular, by processing table and image information in combination with text, it is possible to ensure that key figures or explanations are reflected in the summary without being omitted.

[0021] In addition, according to the present invention, since the summary results are provided in the form of structured data, they can be immediately utilized in various subsequent systems such as search, financial analysis, and risk analysis, thereby increasing the efficiency and reliability of utilizing electronic disclosure documents.

[0022] The effects according to the embodiments are not limited to those exemplified above, and a wider variety of effects are included in this specification. Brief explanation of the drawing

[0024] FIG. 1 is a diagram schematically showing the interlocking structure of an electronic disclosure document multi-stage summary generation system according to one embodiment. FIG. 2 is a schematic diagram showing the configuration of a summary generation server of an electronic disclosure document multi-stage summary generation system according to one embodiment. FIG. 3 is a diagram schematically showing the configuration of a database of an electronic disclosure document multi-stage summary generation system according to one embodiment. FIG. 4 is a diagram schematically showing the configuration of the data preprocessing unit of an electronic disclosure document multi-stage summary generation system according to one embodiment. FIG. 5 is a diagram schematically showing the operation flow of a method for generating a multi-stage summary of an electronic disclosure document according to one embodiment. Specific details for implementing the invention

[0025] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims.

[0026] All flowcharts, state transition diagrams, and pseudocode should be understood as representing various processes performed by a computer or processor that can be substantially represented on a computer-readable medium, regardless of whether the computer or processor is explicitly depicted.

[0027] The functions of the various elements illustrated in the drawings, including functional blocks represented as processors or similar concepts, may be provided by the use of dedicated hardware as well as hardware capable of executing software in conjunction with appropriate software. When provided by a processor, said functions may be provided by a single dedicated processor, a single shared processor, or a plurality of individual processors, some of which may be shared. Throughout the specification, the same reference numerals refer to the same components.

[0028] The terms "...part" and "module" as used in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware or software, or a combination of hardware and software.

[0029] The contents within parentheses used in the specification may be examples. Specifically, where indicated as "Network (Bluetooth)," "Bluetooth" may be proposed as an example of "Network." In this case, the "Network" of the present disclosure is not limited to "Bluetooth," and "Bluetooth" is proposed as an example of "Network."

[0030] In this specification, the term "Electronic Disclosure Document" comprehensively refers to official documents prepared and submitted in electronic form by a legal entity, such as a company, for the purpose of transparently disclosing management activities, financial status, governance structure, and other important information to investors, regulatory authorities, and the general public in accordance with laws or regulations. Representative examples may include annual reports (Form 10-K), quarterly reports (Form 10-Q), and foreign company reports (Form 20-F) disclosed through the U.S. Securities and Exchange Commission (SEC)'s EDGAR (Electronic Data Gathering, Analysis, and Retrieval) system. Furthermore, since the scope of application of this invention is not limited to specific countries or forms, it should be interpreted as a broad concept that includes not only business reports, semi-annual reports, quarterly reports, and major event reports disclosed through the Financial Supervisory Service of Korea's Electronic Disclosure System (DART), but also all forms of electronic disclosure documents submitted to regulatory authorities in various countries.

[0031] As used in this specification, "Item" refers to a logical unit of information constituting an electronic disclosure document and typically refers to a section consisting of a heading having a unique number and title and a corresponding body. For example, in the U.S. Securities and Exchange Commission (SEC) 10-K report, information on a specific subject required by regulations forms an item, such as "Item 1. Business," "Item 1A. Risk Factors," and "Item 7. Management's Discussion and Analysis of Financial Condition and Results of Operations."

[0033] Specific embodiments will be described below with reference to the attached drawings.

[0034] FIG. 1 is a diagram schematically showing the interlocking structure of an electronic disclosure document multi-stage summary generation system according to one embodiment.

[0035] Referring to FIG. 1, an electronic disclosure document multi-stage summary generation system (10) according to one embodiment may be configured to include a summary generation server (100) and a user terminal (200).

[0036] The summary generation server (100) is a computing device responsible for core data processing of the system, and collects a vast amount of disclosure documents from external electronic disclosure systems (e.g., SEC EDGAR, Financial Supervisory Service DART, etc.), divides them into items or sections, and performs multi-stage summarization. The summary generation server (100) performs the function of generating a primary summary by applying optimized prompts for each item through integration with a large-scale language model (LLM) or by utilizing an embedded model, and generating a secondary summary for the entire document by integrating and reconstructing the generated primary summaries. Additionally, the summary generation server (100) converts the generated summary data into a structured data format such as JSON, stores it, or provides it to the user.

[0037] A user terminal (200) is a device that provides an interface for a user to access a summary generation server (100), request the generation of a summary for a specific company or a specific disclosure document, and check the results. The user terminal (200) communicates with the summary generation server (100) through a web browser or a dedicated application, and can visually display the generated item-by-item summary and overall summary, or download them in the form of data files for subsequent analysis. The user terminal (200) can be implemented as various electronic devices equipped with communication functions and data output functions, such as a desktop PC, a laptop, a smartphone, a tablet PC, or a wearable device.

[0038] The summary generation server (100) and the user terminal (200) are interconnected via a wired or wireless network to transmit and receive data. The network may include various communication standards such as a local area network (LAN), a wide area network (WAN), a mobile communication network (5G, LTE, etc.), and a satellite communication network. Additionally, although not specifically illustrated in the drawing, the summary generation server (100) may be linked via API or crawling with servers of regulatory agencies in various countries or servers of financial data providers to secure public disclosure documents.

[0039] Meanwhile, in FIG. 1, the summary generation server (100) and the user terminal (200) are shown as a separated client-server structure, but this is merely one embodiment of the present invention and is not limited thereto. For example, the functions performed by the summary generation server (100) may be integrated and implemented in the form of a standalone application or a local program installed and running within the user terminal (200).

[0041] FIG. 2 is a schematic diagram showing the configuration of a summary generation server of an electronic disclosure document multi-stage summary generation system according to one embodiment.

[0042] Referring to FIG. 2, the summary generation server (100) may be configured to include a communication unit (110), a database (120), a public data collection unit (130), a data preprocessing unit (140), a multi-stage summary calculation unit (150), a result structuring unit (160), and a control unit (170).

[0043] The communication unit (110) provides a communication interface for the summary generation server (100) to transmit and receive data with an external device or network. The communication unit (110) is connected to a user terminal (200) and an external electronic disclosure system server (e.g., SEC EDGAR, DART) or an external Large Language Model (LLM) API server via wired or wireless communication methods. Specifically, the communication unit (110) receives a request signal from the user terminal (200) for the generation of a summary of a disclosure document for a specific company or a specific period, and transmits the finally generated structured summary result data to the user terminal (200). Additionally, the communication unit (110) can perform the function of connecting to an external electronic disclosure system to receive original data (HTML, PDF, text, etc.), or transmitting data preprocessed for text summary generation to an external LLM server and receiving the inference result (summary text) therefrom. To this end, the communication unit (110) supports various wired and wireless communication protocols such as LAN, WAN, the Internet, 5G, and LTE, and may include standard interfaces for data exchange such as REST API and SOAP.

[0045] FIG. 3 is a diagram schematically showing the configuration of a database of an electronic disclosure document multi-stage summary generation system according to one embodiment.

[0046] The database (120) systematically stores and manages various programs, configuration information, collected original data, and processed result data required for the operation of the summary generation server (100). The database (120) may include at least one type of storage medium among flash memory type, hard disk type, multimedia card micro type, card type memory (e.g., SD or XD memory), RAM (Random Access Memory), S-RAM (Static Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, magnetic disk, and optical disk, and may be implemented using a relational or non-relational database management system (DBMS) such as Oracle, MySQL, and MongoDB. The database (120) may be configured to include original disclosure data (121), item-specific prompt data (122), and summary result data (123).

[0047] The original disclosure data (121) stores the original disclosure documents and related metadata of various companies collected from an external electronic disclosure system. The original disclosure data (121) includes not only the raw files collected in various formats such as HTML, XML, PDF, and text files, but also metadata for document management such as the unique identification number of the document, company name, ticker, disclosure type (10-K, 10-Q, business report, etc.), and disclosure submission date. The original disclosure data (121) can be stored while maintaining the integrity of the original in preparation for cases where original reference or reprocessing is required during the subsequent summary generation process, and can also be stored in an indexed form with text data, images, and table data separated after text parsing.

[0048] Item-specific prompt data (122) stores prompt templates containing summary guidelines and constraints optimized for the characteristics of each item of each disclosure document. Item-specific prompt data (122) includes prompt engineering data that defines the types of key information that must be extracted from the relevant section (e.g., major products, sales trends, legal proceedings, market risks, etc.), the tone of the summary, output length limits, and how to handle tables or numerical data, in accordance with the purpose and nature of each item (e.g., Item 1. Business, Item 7. MD&A, Item 1A. Risk Factors, etc.) defined according to disclosure regulations (e.g., SEC Regulation SK). This data can be managed by subdividing it according to the type of document (10-K, 10-Q, etc.) or industry group, and can be continuously updated based on summary quality feedback.

[0049] Summary result data (123) stores the primary summary for each item and the secondary summary for the entire document, generated by the multi-stage summary output unit (150), in a structured format. The summary result data (123) is not a simple list of text, but is stored in a data format with a hierarchical structure such as JSON or XML, and includes an item identifier indicating which item of the original document each summary corresponds to, the time of summary generation, and prompt version information used. Additionally, the summary result data (123) includes tagging information for key figures or keywords included in the summary, thereby supporting easy extraction or filtering of specific data in a future search or analysis system. For example, the summary result data (123) is composed of a top field containing the summary of the entire document and an array field containing the summary for each item, and is maintained in a form that can be provided immediately upon user request.

[0050] The disclosure data collection unit (130) performs the function of automatically searching for and collecting disclosure documents to be summarized from an external electronic disclosure system or financial data provision platform. The disclosure data collection unit (130) may perform collection periodically according to a preset scheduling cycle, or may perform collection immediately in response to a real-time request from a user terminal (200).

[0051] The disclosure data collection unit (130) can be implemented by calling an Open API provided by the relevant agency to access regulatory disclosure repositories in each country, such as the SEC EDGAR system or the Financial Supervisory Service DART system, or by monitoring RSS (Really Simple Syndication) feeds distributed by the disclosure agency, or by using a combination of at least one of web crawling and scraping techniques to directly explore web pages. In particular, when using the RSS feed method, the disclosure data collection unit (130) periodically accesses the RSS channel URL of the target disclosure system to receive feed data in the format of XML (Extensible Markup Language). Next, the public disclosure data collection unit (130) can detect whether a new disclosure has been registered in real time by parsing tags such as 'pubDate (publication date)', 'link (document path)', 'title (disclosure title)', and 'guid (unique identifier)' within the received XML data, and can ensure the timeliness of the collection by downloading the original file through the link (URL) of the detected new disclosure.

[0052] The disclosure data collection unit (130) selectively collects only the necessary documents by applying filtering conditions such as the identification code (CIK, unique number, etc.) of the target company, document type (10-K, 10-Q, 8-K, business report, etc.), and disclosure period during the collection process. In addition, the disclosure data collection unit (130) manages the collected documents by considering that they may exist in various formats such as HTML, XML, XBRL, and PDF, and prioritizes securing the format most suitable for subsequent processing (e.g., HTML or XML, which facilitates text extraction), or by collecting multiple formats simultaneously so that they can be utilized complementarily. The collected data is loaded into the disclosure original data (121) area of ​​the database (120) along with metadata. For example, the disclosure data collection unit (130) can maintain the timeliness of the document by detecting a notification of a new disclosure registration through RSS feed monitoring and immediately accessing the original URL of the disclosure to download the document.

[0054] FIG. 4 is a diagram schematically showing the configuration of the data preprocessing unit of an electronic disclosure document multi-stage summary generation system according to one embodiment.

[0055] Referring to FIG. 4, the data preprocessing unit (140) performs the function of converting and refining collected unstructured or semi-structured public documents into structured data that can be effectively processed by a large-scale language model (LLM). The data preprocessing unit (140) may be configured to include an item content separation module (141), a format conversion module (142), and a table and image processing module (143).

[0056] The item content separation module (141) divides the entire collected disclosure document into logical semantic units called items or sections. The item content separation module (141) identifies the start and end points of each item by analyzing structural features of the disclosure document, such as hyperlinks in the table of contents, HTML tags (e.g., header tags, bold tags), or specific keyword patterns (e.g., "Item 1.", "Item 7.", "III. Matters concerning management," etc.). To do this, a pattern matching algorithm based on regular expressions, a machine learning model specialized in document structure recognition, or a layout analysis technique based on computer vision may be utilized. In this case, the machine learning model may be a supervised learning model constructed from a large amount of past SEC and DART electronic disclosure files (HTML and PDF) as a training dataset. Specifically, the above model learns data in which bounding box coordinates and class labels are tagged by experts or automation tools for areas such as 'Item headers', 'Sub-section titles', 'Body text', 'Tables', and 'Comments' within a public disclosure document. This model may include a LayoutLM, Graph Convolutional Network (GCN), or Object Detection-based deep learning architecture that learns by extracting not only text sequence information but also visual and spatial features such as text font size, boldness, indentation, and spacing between paragraphs.

[0057] The item content separation module (141) does not simply cut the document into page units, but extracts text blocks in units of information items defined by the regulatory body, thereby ensuring that the context of each item is not mixed and can be processed independently in the subsequent summarization stage.

[0058] The format conversion module (142) converts the content of each separated item into a text format optimized for LLM input. The format conversion module (142) performs a cleansing operation to remove noise data unnecessary for summary generation, such as HTML tags, script codes, and special characters, unify the encoding of the text, and clean up unnecessary spaces or line breaks between sentences. Additionally, the format conversion module (142) can process the text into a modified form (e.g., "The following is text about [Item 7. MD&A]: ...") by adding the title or identification code of the corresponding item as metadata to each text block so that the LLM can clearly recognize which item the text is about.

[0059] The table and image processing module (143) converts non-text information, such as tables, images, and charts included in the disclosure document, into a form that can be understood by a text-based LLM. The table and image processing module (143) parses the HTML table structure and converts it into structured text in CSV (Comma-Separated Values), Markdown table, or JSON format using specific mapping rules that maintain the column / row structure of the table as is, thereby ensuring that the semantic relationship between the numeric data and rows / columns within the table is maintained. In particular, the table and image processing module (143) identifies areas where cell merging has occurred due to 'rowspan' or 'colspan' attributes within HTML tags, and performs flattening processing by logically duplicating and assigning the corresponding text value to all individual cell coordinates occupied by the merged cells, or redefining it into a hierarchical header structure. Through this, even in financial statements with complex merge structures, the conversion can be performed while maintaining structural consistency regarding which item (Row Header) and period (Column Header) each numeric data is mapped to. In the case of images, the table and image processing module (143) can extract text within the image using optical character recognition (OCR) technology or generate text (Caption) explaining the meaning of the chart or graph using an image captioning model, which can then be inserted between the text streams of the main text or linked to separate reference data. This prevents key information contained in the table or image, such as financial statements or statistical data, from being omitted during the summarization process.

[0061] The multi-stage summary output unit (150) generates a final summary result by performing a hierarchical summary process based on preprocessed item-specific data. The multi-stage summary output unit (150) may be configured to include an item-specific first-level summary module (not shown) and a document-wide second-level summary module (not shown).

[0062] The item-specific primary summary module performs a summary individually for each separated item (section) data. The item-specific primary summary module refers to the item-specific prompt data (122) in the database (120) to load a dedicated prompt mapped to the type of item to be processed (e.g., financial status, risk factors, business overview, etc.). This dedicated prompt includes information that needs to be summarized in the item (e.g., causes of sales increase / decrease, major legal risks, etc.), as well as instructions regarding output format, length limits, etc. The item-specific primary summary module combines the text and converted table / image data of the item with the dedicated prompt and inputs it into the LLM, thereby generating a primary summary that compresses the core content of the item. This process can be performed in parallel for each item to increase processing speed and ensure that local details of the document are preserved without being diluted during the overall summarization process.

[0063] The document-wide secondary summary module synthesizes the generated item-specific primary summaries to create a high-level summary that encompasses the entire document. It configures integrated prompts by arranging the primary summaries of each item in a logical order or selecting them based on importance, and then inputs these into the LLM to generate the overall summary. In this process, by using already refined and compressed primary summaries as input instead of the entire original text, the LLM's input token limit issue is resolved, and computational costs are reduced. When generating the overall summary, the module can provide prompts focused on deriving macroscopic information, such as the relationships between individual items, key messages running through the entire document, and comprehensive assessments of financial soundness. Consequently, this module provides high-quality secondary summaries that allow users to quickly grasp the overall context and key issues without having to read the vast disclosure documents page by page.

[0064] The result structuring unit (160) performs the function of converting and systematizing the item-specific primary summary and the entire document secondary summary results generated through the multi-stage summary output unit (150) into a standardized data format with high machine readability and usability. The result structuring unit (160) maps the generated summary texts in natural language form into data objects with a hierarchical structure such as JSON (JavaScript Object Notation) and XML (Extensible Markup Language).

[0065] Specifically, the result structuring unit (160) defines a schema containing a top field (e.g., 'document_summary') containing overall document summary information and a lower array field (e.g., 'item_summaries') containing detailed summary information for each item, and populates the data. Each item summary object may store metadata such as the unique identifier of the item (e.g., 'Item 1A'), title, and summary body text, as well as key figures and entities extracted during the summarization process, in separate fields.

[0066] In addition, the result structuring unit (160) can also perform the function of rendering and providing summary results in a human-readable document form, such as a PDF report, a Word document, or an HTML snippet for a web dashboard, according to the user's request. Through this structured data processing, the output of the system is not limited to a simple text file, but is provided in a form that can be immediately linked with various subsequent applications, such as search engine indexing, input data for investment analysis algorithms, and knowledge bases for RAG (Retrieval-Augmented Generation) systems, thereby increasing data availability.

[0067] The control unit (170) is a central processing unit that controls the overall operation of the summary generation server (100) and manages the data flow between components. The control unit (170) can be implemented as a processor, a CPU (Central Processing Unit), an AP (Application Processor), etc., and drives each function module by executing program instructions stored in memory.

[0068] When a summary request is received from a user, the control unit (170) controls the public disclosure data collection unit (130) to instruct it to collect the relevant public disclosure documents, and when collection is completed, it calls the data preprocessing unit (140) to control it to perform item separation and format conversion operations. Subsequently, the control unit (170) schedules the first and second summary generation processes sequentially or in parallel through the multi-stage summary calculation unit (150), monitors for delays or error situations that may occur during LLM API calls, and performs exception handling.

[0069] Additionally, the control unit (170) performs management functions for the stable operation of the system, such as performing logic to verify the quality of the generated summary results, controlling data access based on user authority, and managing input / output of the database (120). For example, if the generation of a summary of a specific item fails, the control unit (170) may perform a retry only for that item, or provide real-time feedback on the progress of the entire process to the user terminal (200).

[0070] FIG. 5 is a diagram schematically showing the operation flow of a method for generating a multi-stage summary of an electronic disclosure document according to one embodiment.

[0071] Referring to FIG. 5, first, the summary generation server (100) performs the step of collecting disclosure documents to be summarized from an external electronic disclosure system (SEC, DART, etc.) through the communication unit (110) and the disclosure data collection unit (130) (S501). At this time, the documents collected may be in various formats such as HTML, PDF, etc., and related metadata is collected together.

[0072] Next, the data preprocessing unit (140) analyzes the collected public documents and performs a step of separating the content into logical information units, such as items or sections (S502). In this step, the beginning and end of each item are identified and text blocks are extracted through a machine learning model specialized in document structure recognition or regular expression pattern matching.

[0073] Next, the data preprocessing unit (140) performs the step of converting unstructured data (tables, images) contained within each separated item into structured text that the LLM can understand (S503). HTML tables are converted into structured text (Markdown, JSON, etc.), and images are converted into descriptive text through captioning technology.

[0074] After that, the multi-stage summary output unit (150) performs the step of generating a primary summary for each item by applying a prompt optimized for the type of each item (S504). In this step, the text of each item and the converted table / image data are input into the LLM to extract the core content.

[0075] Next, the multi-stage summary output unit (150) performs the step of integrating the generated primary summaries to generate a secondary summary for the entire document (S505).

[0076] Finally, the result structuring unit (160) performs the step of converting the final generated item-specific primary summary, document-wide secondary summary, and verification result metadata into a structured format such as JSON and storing it or providing it to a user terminal (200) (S506).

[0078] Meanwhile, according to some embodiments, the multi-stage summary output unit (150) may further include a summary structure optimization module (not shown) to generate an in-depth summary based on the logical association of information, regardless of the physical table of contents order of the public document.

[0079] The above summary structure optimization module inputs the primary summaries of each item generated by the primary summary module for each item into a Text Embedding Model to convert them into high-dimensional vector values, and calculates the 'semantic connection strength' by calculating the cosine similarity between items.

[0080] At this time, the summary structure optimization module defines a group of items whose calculated semantic connection strength exceeds a preset threshold (e.g., 0.75) as a 'semantic cluster'. For example, if a specific item within 'Item 1. Business Overview' located at the beginning of the document and 'Item 8. Notes to Financial Statements' located at the end of the document shows high similarity, they are grouped into a single cluster. Subsequently, instead of integrating summaries in the order of the document's table of contents, the document-wide secondary summary module can first generate 'intermediate summaries' based on the semantic clusters formed by the summary structure optimization module, and finally combine the generated intermediate summaries to produce a full summary. Through this configuration, the present invention can provide insightful summary results by organically connecting disclosure information that is geographically separated but logically closely related.

[0081] However, if the above summary structure optimization module groups items based solely on similarity, a phenomenon may occur where an excessive number of items are grouped into one large cluster because specific issues (e.g., mergers, pandemics, etc.) are mentioned throughout the disclosure documents. In this case, there is a risk that the text volume of the cluster may exceed the input token limit of the language model, or that the focus of the summary may be blurred due to the mixing of multiple pieces of information.

[0082] To prevent this, in this embodiment, the summary structure optimization module may perform a token balancing-based recursive partitioning logic. Specifically, the summary structure optimization module calculates the total amount of tokens in the formed semantic cluster and determines whether this value exceeds a preset threshold (e.g., 25,000 to 35,000 tokens). If it exceeds the threshold, the module re-analyzes the connection strength between items within the cluster to identify the hub item with the highest centrality within the cluster, and re-divides the large cluster into multiple sub-clusters centered around this hub item. In particular, the re-division process may be performed recursively until the amount of tokens in each divided sub-cluster falls below the preset threshold.

[0083] Specifically, the summary structure optimization module sets the primary summary of each item included in the large cluster as a node and sets the cosine similarity value between items as the edge weight to temporarily construct a 'weighted graph'. Then, the 'Eigenvector Centrality' or 'PageRank' algorithm is performed on the constructed graph to calculate an influence score indicating how closely each node is connected to other nodes, and the item with the highest score is selected as the 'hub item'.

[0084] Furthermore, the aforementioned re-segmentation process can be performed using a 'K-Means Clustering' technique, which sets a selected first hub item and a second hub item—which has the second highest centrality within the cluster and whose similarity to the first hub item is below a preset threshold—as initial centroids, and forms clusters by assigning the remaining items to the hub item with a vector distance closer to them, or a 'Louvain Community Detection' technique based on density within the graph. Through this, rather than simply mechanically splitting, contextual disconnection can be minimized by inducing subgroups with high information cohesion to be grouped together.

[0085] In addition, the summary structure optimization module described above can process the input of items within the cluster by rearranging them according to a logical sequence based on the timestamps or table of contents of the original document, rather than in a random order, so that the summary is performed while preserving the causal relationships of the information.

[0087] Additionally, according to some other embodiments, the electronic disclosure document multi-stage summary generation system (10) may further include an integrity verification unit (not shown) to verify the facts of the generated summary and to support the user in easily comparing it with the original text.

[0088] The integrity verification unit performs backtracking indexing and numerical cross-verification functions. Specifically, for each sentence of the summary generated by the multi-stage summary output unit (150), the integrity verification unit analyzes which paragraph or sentence of the original public document the sentence is derived from, and generates metadata that maps the start offset and end offset information of the original text. Through this, when a user clicks on the summary on the terminal (200), an interface can be provided that immediately moves to (highlights) the corresponding location in the original text or displays it as a popup.

[0089] In addition, the integrity verification unit extracts numerical information included in the summary and compares it with JSON table data structured by the table and image processing module (143) of the data preprocessing unit (140). If the error between the numerical value listed in the summary and the actual numerical value in the table data exceeds a preset range (e.g., 3% to 7%), the integrity verification unit automatically assigns a 'review needed' tag to the summary sentence or generates a separate comment to warn the user of the possibility of numerical discrepancies. This configuration effectively controls the hallucination phenomenon, a chronic problem of generative AI, and provides the effect of improving the reliability of the summary report.

[0091] Although embodiments of the present invention have been described above with reference to the attached drawings, those skilled in the art will understand that the present invention may be implemented in other specific forms without changing the technical concept or essential features thereof. Therefore, the embodiments described above should be understood as illustrative in all respects and not restrictive. Explanation of the symbols

[0093] 10: Multi-stage Summary Generation System for Electronic Disclosure Documents 100: Summary generation server 110: Communications Department 120: Database 121: Original disclosure data 122: Item-specific prompt data 123: Summary result data 130: Public Disclosure Data Collection Department 140: Data Preprocessing Section 141: Item Content Separation Module 142: Format conversion module 143: Table and Image Processing Module 150: Multi-stage summary output section 160: Result Structure 170: Control unit 200: User terminal

Claims

Claim 1 The system comprises a communication unit that transmits and receives data to and from a user terminal and an external electronic disclosure system; a database that stores original disclosure data, item-specific prompt data, and summary result data; a disclosure data collection unit that collects electronic disclosure documents; a data preprocessing unit that separates the collected electronic disclosure documents into item units and converts their formats; a multi-stage summary output unit that generates a primary summary for each item and integrates the primary summaries to generate a secondary summary for the entire document; a result structuring unit that converts the generated summary results into a structured data format; and a control unit that controls the operation of each component and manages data flow. The database includes original disclosure data that stores raw files in HTML, XML, or PDF formats and metadata including company name, disclosure type, and disclosure submission date; item-specific prompt data that stores prompt templates containing summary guidelines and constraints optimized for the characteristics of each item; and summary result data that stores item-specific primary summaries and secondary summaries for the entire document in JSON or XML format. The summary result data includes a top-level field containing summary information for the entire document and a bottom-level field containing detailed summary information for each item. It includes an array field, and detailed summary information for each item includes an item unique identifier, title, summary body text, key figures, and key entities; the multi-stage summary output unit includes an item-specific primary summary module that generates a primary summary for each item by loading a dedicated prompt mapped to the type of each item, and a document-wide secondary summary module that generates a secondary summary for the entire document by integrating the generated item-specific primary summaries; the dedicated prompt includes the type of key information to be extracted from the item to be processed, the tone of the summary, output length limits, and the method of processing table or numerical data; the item-specific primary summary module combines the dedicated prompt with the text of the corresponding item and converted table / image data and inputs it into a large-scale language model; and the result structuring unitDefines a schema including a top-level field containing overall document summary information and a sub-array field containing detailed summary information for each item, stores metadata including an item unique identifier, title, summary body text, key figures, and key entities in each item summary object, renders the summary results in the form of a PDF report, Word document, or HTML, and the multi-stage summary output unit further includes a summary structure optimization module that inputs primary summaries for each item into a text embedding model to convert them into vectors and calculates semantic connection strength by calculating cosine similarity between items, the summary structure optimization module defines item groups whose semantic connection strength exceeds a preset threshold as semantic clusters, the overall document secondary summary module generates intermediate summaries at the semantic cluster level and then combines the intermediate summaries to generate an overall summary, and if the total amount of tokens in the formed semantic cluster exceeds a preset threshold, the summary structure optimization module constructs a weighted graph by setting the primary summaries of each item included in the semantic cluster as nodes and setting the cosine similarity values ​​between items as edge weights, and Eigenvector Centrality or PageRank The algorithm is executed to select the item with the highest influence score as the hub item, and the semantic cluster is re-divided into multiple sub-clusters using K-means clustering or Rubane community detection techniques with the hub item as the center, and the re-division is repeated until the token amount of each divided sub-cluster is below the preset threshold, wherein the re-division sets the hub item as the first hub item, and sets the item having the second highest centrality within the semantic cluster, while having a similarity to the first hub item below a preset standard, as the second hub item.A system for generating item-unit multi-stage summaries of electronic disclosure documents, performed by setting the first hub item and the second hub item as initial center points, respectively. Claim 2 delete Claim 3 In claim 1, the data preprocessing unit comprises an item content separation module that divides the disclosure document into item units, a format conversion module that converts the content of the separated items into a text format optimized for input into a large-scale language model (LLM), and a table and image processing module that converts tables and images within the disclosure document into text form; wherein the item content separation module identifies the start and end points of each item by analyzing the table of contents hyperlinks, HTML tags, or specific keyword patterns of the disclosure document, and the table and image processing module converts the HTML table structure into structured text in CSV, Markdown, or JSON format, and generates text explaining the meaning of text or tables within the image using optical character recognition technology or an image captioning model, thereby forming a multi-stage summary generation system for item units of electronic disclosure documents. Claim 4 delete Claim 5 delete

Images