Financial report abstract generation system based on artificial intelligence generated content

Abstract

This invention discloses a financial statement summary generation system based on artificial intelligence-generated content, belonging to the field of artificial intelligence technology. The system performs segmentation and extraction of cited fragments from the financial statement text, constructs a closed-loop structure by combining candidate fragments from footnotes, and classifies the closure status through correspondence comparison. It then performs backfilling expansion and constraint reconstruction on the closed structure to generate summary generation primitives containing the main body of the matter, the main body of the explanation, and the main body of the limitation. Finally, it verifies the structural consistency and semantic integrity of the summary generation primitives through checks on the clarity of the matter, the correspondence of the explanation, and the completeness of the limitation. This system, focusing on the construction and verification process of the closed-loop structure, structurally reconstructs the cross-segment relationships between the main body of the financial statement and the footnotes, ensuring that the summary content has a clear subject focus, consistent sources of explanation, and complete limiting conditions. It is suitable for scenarios involving automatic financial statement summary generation and assisted review.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to a financial statement summary generation system based on artificial intelligence-generated content. Background Technology

[0002] Artificial intelligence (AI) has evolved from rule-driven to data-driven approaches in text processing. Early methods primarily relied on keyword matching and template generation to process structured text. Subsequently, with the development of statistical learning methods and deep learning models, text understanding and generation gradually acquired semantic modeling capabilities. During this process, the demand for text processing related to financial reports emerged. Financial reports, as typical semi-structured long texts containing a large amount of technical terminology and hierarchical information, became one of the important application scenarios for AI text processing. With technological advancements, the automatic parsing, information extraction, and summary generation of financial report texts have gradually formed independent research directions, driving the continuous development of financial report summary generation technology based on AI-generated content.

[0003] Existing methods for generating financial statement summaries based on artificial intelligence often employ keyword matching or simple semantic retrieval to concatenate relevant information from the text with footnotes when dealing with the relationship between citations in the main text and footnotes. However, in practice, these methods often only extract a single citation or footnote sentence, lacking recognition of contextual relationships, leading to gaps between the item name, the object of description, and the limiting conditions. Furthermore, they lack uniform constraints on pronouns in the text, definitions in the footnotes, and reasons for changes, resulting in unclear references, missing descriptions, or incomplete limitations in the generated results. In addition, they lack a systematic closure mechanism for the concatenated expression, making it difficult to determine whether the expression structure forms a complete semantic unit. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a financial statement summary generation system based on artificial intelligence-generated content, which solves the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a financial statement summary generation system based on artificial intelligence-generated content, comprising a text parsing module, a citation closure generation module, a closure state determination module, a summary primitive reconstruction module, and a summary verification module; The text parsing module is used to divide the financial statement text into sections, extract the source fragments cited in the main text, and generate a set of candidate fragments for footnotes; The reference closure generation module is used to filter the corresponding note segments from the note candidate segments according to the item name, and expand them to form note supplement segments. The text reference source segment and the note supplement segment are sequentially spliced ​​together and supplemented to generate a reference closure structure. The closed state determination module is used to perform a correspondence comparison on the referenced closed structure, and mark the referenced closed structure as a complete closed structure, a partially closed structure, or a closed structure to be filled in according to the comparison result; The summary primitive reconstruction module is used to construct extended reconstruction fragments and perform constraint reconstruction processing with reference closure structures to generate summary generation primitives. The summary verification module is used to verify the clarity of the execution items, the correspondence of the descriptions, and the completeness of the constraints of the summary generation primitives.

[0006] Preferably, the text parsing module includes a segment parsing unit and a retrieval condition generation unit; The segment parsing unit is used to read the financial report text, divide the financial report text into the main text segment, the footnote segment and the title segment according to the title line, paragraph line and footnote line in the financial report text, and search for the statement containing the footnote pointing expression in the main text segment, extract the sentences and segments that are connected to the current statement, and form the main text citation source segment. The notes refer to the following expressions: see notes, refer to notes, see notes for details, see notes for specifics, and see notes to the financial statements for relevant content.

[0007] Preferably, the retrieval condition generation unit is used to extract the footnote number information, item name, current title content and continuation description in adjacent paragraphs from the text citation source fragment, and use them as footnote retrieval conditions to retrieve multiple corresponding footnote candidate contents in the footnote section, generate a footnote candidate fragment set, and then bind each text citation source fragment to its corresponding footnote candidate fragment set to form an original associated object.

[0008] Preferably, the reference closure generation module includes an item filtering unit and a closure splicing unit; The item filtering unit is used to extract the item name and main narrative content from the source fragments cited in the main text, and to filter the annotation segments directly corresponding to the current item from the annotation candidate fragment set based on the item name. It then extracts the scope and definition of the current item from the annotation segments and the composition and reasons for change from the current item, generating supplementary annotation fragments.

[0009] Preferably, the closing splicing unit is used to splice the main text section and the supplementary note section in the main text source fragment in a fixed order to form the expression fragment to be closed, and to perform supplementation and repair on the repeated expressions, pronoun references, missing subjects, missing objects and incomplete qualifiers in the expression fragment to be closed, so as to obtain a single closed expression; The completed and revised single closed expression, along with the item name, the source fragment in the main text, the supplementary fragment in the footnote, the source location in the main text, and the source location in the footnote, are encapsulated to generate a closed citation structure.

[0010] Preferably, the closure state determination module includes a correspondence comparison unit and a closure determination unit; The correspondence comparison unit is used to perform a correspondence comparison between the source fragment cited in the main text and the supplementary fragment in the footnote, as follows; Extract the event name from the source fragment cited in the main text and extract the explanatory notes from the supplementary notes fragment. Perform same name matching and synonym matching on the event name and the explanatory notes fragment. When the event name and the explanatory notes fragment are the same, or when the event name and the explanatory notes fragment refer to the same disclosure event in the financial statement, determine that the event name has a corresponding explanation in the supplementary notes fragment. Locate the citation trigger from the source fragment in the main text, and check whether there is a description of the scope, composition, reasons for change, and range corresponding to the name of the current matter after the location of the citation trigger. When the description of the name of the current matter has been formed after the citation trigger, it is determined that the citation trigger has been replaced by the description. Pronoun retrieval is performed on single closed expressions to extract the referential pronouns and to search backwards along the preceding statements in the single closed expression for the item name or annotation object corresponding to the referential pronoun. When there are referential pronouns that cannot be returned to the item name or annotation object, it is determined that there is an unreturned object in the single closed expression. The limiting component retrieval is performed on the limiting phrase completion fragment, and the time limitation, scope limitation, scope limitation and object limitation are extracted. It is then checked whether each limiting component forms a complete correspondence with the item name and the object of the annotation in the annotation completion fragment. When any limiting component is missing any of the limiting object, limiting scope, limiting time or limiting scope, it is determined that there is incomplete limiting content in the limiting phrase completion fragment. Extract the conclusion expression from a single closed expression and check whether there are corresponding explanations of scope, composition, reasons for change, and range before and after the current conclusion expression. If the conclusion expression does not match any explanation, it is determined that there is a break in the single closed expression.

[0011] Preferably, the closure determination unit is used to perform closure state determination on each referenced closed structure based on the correspondence comparison processing result, and summarize the closure state determination results to generate a closure state mark result. The closure state determination is as follows: When the item name has a corresponding description, the reference trigger has been replaced by the description content, there is no unreferential object in the single closed expression, the qualifier supplement fragment is complete and there is no broken expression in the single closed expression, the current reference closed structure is marked as a complete closed structure; When the item name has a partial corresponding description, but there are still broken expressions in a single closed expression, the currently referenced closed structure is marked as a partially closed structure; When any of the following conditions are met—that the item name does not have a corresponding description, the reference trigger is not replaced by the description content, or there is still a broken expression in a single closed expression—the current reference closed structure is marked as a closed structure to be filled back.

[0012] Preferably, the summary primitive reconstruction module includes a back-extension unit and a primitive reconstruction unit; The back-complementary extension unit is used to read the complete closed structure and the partially closed structure in the closed state marking result; Extract the item name, text source fragment, footnote supplement fragment, text source position, and footnote source position of each referenced closed structure from the complete closed structure set. At the same time, extract the preceding sentence and following sentence of the text corresponding to the text source position of each referenced closed structure in the partial closed structure. Simultaneously, read the footnote adjacent sub-item title and footnote subsequent explanation sentence corresponding to the footnote source position in each referenced closed structure. Add the preceding sentence, following sentence, adjacent sub-item title, and subsequent explanation sentence of the footnote into the current referenced closed structure to generate an extended reconstructed fragment.

[0013] Preferably, the primitive reconstruction unit is used to perform constraint reconstruction on each referenced closed structure and extended reconstruction fragment in the complete closed structure to obtain the main body of the matter, the main body of the description and the main body of the limitation, and combine them in a fixed order to generate the summary generation primitive; The constraint reconstruction process includes the following: The item name is fixed as a single main object as the item's main body; Extract the scope, composition, reasons for change, and range of the items from the supplementary and expanded fragments of the notes, and organize them into the main body of the explanation according to the principle of unique correspondence between the objects of the notes; Extract the limiting object, limiting scope, limiting time, or limiting definition that directly corresponds to the event name from the limiting phrase supplement or extended reconstruction phrase, and organize them into the limiting main body.

[0014] Preferably, the summary verification module includes a matter clarity verification unit, a description correspondence verification unit, and a limitation completeness verification unit; The item clarification verification unit is used to extract the financial statement item names in the item trunk and detect whether there are any pronouns in the item trunk and whether the item trunk contains two or more parallel financial statement item names at the same time. When there are no pronouns in the main body of a matter and it corresponds to only a single financial statement matter name, the main body of the matter is clearly defined. When there are pronouns in the main body of a matter, or when there are two or more parallel financial statement matter names, the main body of the matter is determined to be unclear, and the current summary generation primitive is marked as a primitive to be reviewed; The description correspondence verification unit is used to extract the annotation objects and item names in the description trunk, and to perform consistency matching between the annotation objects and the item names in the item trunk. When all the notes to the financial statements belong to the financial items corresponding to the main subject matter, it is determined that the notes to the financial statements are consistent with the main subject matter. When any note in the main body of a matter does not belong to the financial items of the main body of the matter, the current note is written into the information to be removed and the removal process is performed. After the removal process, if the main body of the matter retains at least one note of financial item, the main body of the matter is deemed valid; if the main body of the matter does not retain any note of financial item, the main body of the current summary generation element is deemed invalid, and the current summary generation element is marked as a element to be reviewed. The defined integrity verification unit is used to extract the time limit, scope limit, caliber limit and object limit in the defined backbone; Perform corresponding verification on each of the specified contents and the item names in the main body of the item and the supplementary notes in the main body of the description; The main body of a limiting structure is deemed qualified when at least one item name or a description of an item is included in the limiting structure. If the constraint trunk does not contain the constraint content of the item name or description object, the constraint trunk is deemed unqualified. In this case, the current summary generation primitive is returned to the summary primitive reconstruction module to re-execute the constraint reconstruction process.

[0015] This invention provides a financial statement summary generation system based on artificial intelligence-generated content. It has the following beneficial effects: (1) This method divides the financial report text into title lines, paragraph lines, and footnote lines using a text parsing module. It then extracts textual source fragments containing footnote references from the main text sections, parsing the footnote number, item name, current title content, and connecting descriptions from adjacent paragraphs to form a set of candidate footnote fragments and establish original related objects. Based on this, a citation closure generation module further filters and expands the set of candidate footnote fragments, generating supplementary footnote fragments. The main text source fragments and supplementary footnote fragments are then sequentially concatenated and, after supplementation and refinement, encapsulated into a citation closure structure. This process completes the hierarchical construction of extracting main text source fragments, supplementary footnote fragments, and citation closure structures from the original financial report text, establishing a correspondence between item names, main text source locations, and footnote source locations, thus constructing a structured expression unit with complete contextual connections.

[0016] (2) This method performs multi-dimensional correspondence comparison on the closed structure of the reference through the closed state determination module, including the same name matching and synonym matching between the item name and the object of the annotation, the substitution relationship detection between the reference trigger and the explanatory content, the back-reference path verification of the pronoun, the completeness retrieval of time limitation, scope limitation, caliber limitation and object limitation in the qualifier supplement fragment, and the correspondence detection between the conclusion expression and the caliber explanation, composition explanation, change reason explanation and scope explanation, and accordingly divides the closed structure of the reference into a complete closed structure, a partially closed structure and a closed structure to be supplemented. On this basis, the abstract primitive reconstruction module reads the complete closed structure and the partially closed structure, extracts the preceding sentence of the main text corresponding to the source position of the main text, the following sentence of the main text, and the adjacent sub-item title and subsequent explanation sentence of the annotation corresponding to the source position of the annotation, and generates an extended reconstruction fragment; then, the main body of the reference and the extended reconstruction fragment are separated and organized into the main body of the item, the explanatory body and the qualifier body, forming an abstract generation primitive with a fixed structural order. This process completes the structural reconstruction task from the reference closed structure to the summary generation primitive, so that the broken relationships and implicit explanatory relationships in the original expression are uniformly organized after reconstruction.

[0017] (3) This method performs item-by-item verification of the summary generation primitives through the summary verification module: the clarity verification unit checks the uniqueness and pronouns of the names in the main body of the matter; the correspondence verification unit performs consistency matching between the objects of the notes in the main body of the description and the names of the matters, and removes the description content that does not belong to the corresponding financial items in the main body of the matter; the completeness verification unit checks the correspondence between the time limit, scope limit, caliber limit and object limit in the limit main body and the names of the matters and the objects of the notes, and performs a process of reverting to the summary primitive reconstruction module for the summary generation primitives that do not meet the conditions. Through the above verification and revert mechanism, the final output summary generation primitives form a closed and consistent expression result in terms of the object name pointing, the correspondence of the objects of the notes and the limit structure. Compared with the processing method that only relies on direct extraction or simple splicing, it forms a more rigorous structural expression in terms of the integrity of the reference relationship, the consistency of semantic inheritance and the completeness of the limit information. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the financial statement summary generation system based on artificial intelligence content according to the present invention; Figure 2 This is a block diagram illustrating the operational logic of the financial statement summary generation system based on artificial intelligence, as described in this invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1 Please see Figure 1 This invention provides a financial statement summary generation system based on artificial intelligence-generated content. To achieve the above objectives, this invention is implemented through the following technical solutions: including a text parsing module, a reference closure generation module, a closure state determination module, a summary primitive reconstruction module, and a summary verification module; The text parsing module is used to divide the financial statement text into sections, extract the source fragments cited in the main text, and generate a set of candidate fragments for footnotes; The reference closure generation module is used to filter the corresponding note segments from the note candidate segments according to the item name, and expand them to form note supplement segments. The text reference source segment and the note supplement segment are sequentially spliced ​​together and supplemented to generate a reference closure structure. The closed state determination module is used to perform a correspondence comparison on the referenced closed structure, and mark the referenced closed structure as a complete closed structure, a partially closed structure, or a closed structure to be filled in according to the comparison result; The summary primitive reconstruction module is used to construct extended reconstruction fragments and perform constraint reconstruction processing with reference closure structures to generate summary generation primitives. The summary verification module is used to verify the clarity of the execution items, the correspondence of the descriptions, and the completeness of the constraints of the summary generation primitives.

[0021] In this embodiment, the text parsing module divides the financial statement text into sections and extracts the source fragments of the main text citations, while simultaneously generating a set of candidate footnote fragments. This establishes a coherent data foundation for information originally scattered across the main text and footnote sections. Based on this, the citation closure generation module selects corresponding footnote segments from the candidate footnote fragments according to the item name, and expands them forward and backward to form supplementary footnote fragments. Finally, the source fragments of the main text citations and the supplementary footnote fragments are sequentially concatenated and trimmed to generate a citation closure structure. This process completes the structural transformation from the original text to source fragments of the main text citations, supplementary footnote fragments, and a citation closure structure, establishing a clear correspondence between item names and footnote descriptions, and completing the organization and integration of cross-section citation content. The closed-state determination module performs a correspondence comparison on the referenced closed structures and marks them as complete closed structures, partially closed structures, or closed structures to be supplemented based on the comparison results, thus distinguishing structures with different levels of completeness. On this basis, the summary primitive reconstruction module constructs extended reconstruction fragments for the above structures and performs constraint reconstruction processing with the referenced closed structures to generate summary generation primitives. This process completes the hierarchical processing and structural reorganization of referenced closed structures, ensuring that the expression of matters, explanatory content, and limiting information form a unified structural organization in the summary generation primitives, changing the processing method of simply splicing together fragments from the original text. The summary verification module performs checks on the clarification of matters, the correspondence of explanations, and the completeness of limitations on the summary generation primitives. Summary generation primitives that do not meet the requirements are rolled back or modified, ensuring that the final result has a consistent structure in terms of matter name references, correspondence of annotations, and limiting expressions. Compared to methods that rely solely on text extraction or model generation, the collaborative processing of these five modules introduces two-level structural units—a citation closure structure and a summary generation primitive—in the generation process. Combined with closure state determination and multi-dimensional verification mechanisms, this results in a clearer structured summary in terms of coherence, correspondence, and completeness of limiting information.

[0022] Example 2 Please refer to Figure 2 Specifically: the text parsing module includes a segment parsing unit and a retrieval condition generation unit; The segment parsing unit is used to read the financial report text, divide the financial report text into the main text segment, the footnote segment and the title segment according to the title line, paragraph line and footnote line in the financial report text, and search for the statement containing the footnote pointing expression in the main text segment, extract the sentences and segments that are connected to the current statement, and form the main text citation source segment. The notes refer to the following expressions: see notes, refer to notes, see notes for details, see notes for specifics, and see notes to the financial statements for relevant content.

[0023] The retrieval condition generation unit is used to extract the footnote number information, item name, current title content and continuation description in adjacent paragraphs from the text citation source fragment, and use them as footnote retrieval conditions to retrieve multiple corresponding footnote candidate contents in the footnote section, generate a footnote candidate fragment set, and then bind each text citation source fragment to its corresponding footnote candidate fragment set to form an original associated object.

[0024] In this embodiment, the segment parsing unit divides the title lines, paragraph lines, and footnote lines in the financial statement text. Within the main text segment, it locates statements containing footnotes such as "See footnotes for details," "Refer to footnotes," "Specific details are in footnotes," and "Related content is in the footnotes to the financial statements." Simultaneously, it extracts the preceding and following sentences to form a main text citation source fragment containing the event name, conclusion description, and reference point. Here, not only a single citation sentence is retained, but the preceding and following sentences are also extracted to preserve the complete context between the event name, conclusion description, and reference point. Subsequently, the retrieval condition generation unit extracts from the... The method parses the footnote number information, item name, current title content, and continuation description in adjacent paragraphs from the source fragment of the main text citation. These elements are used as composite search conditions to perform targeted searches within the footnote section, obtaining multiple corresponding candidate footnote contents and constructing a set of candidate footnote fragments. Simultaneously, each source fragment of the main text citation is bound to its corresponding set of candidate footnote fragments, forming an original association object. The footnote number information is used to limit the search scope, the item name to narrow the semantic scope of candidate footnotes, and the current title content and continuation description in adjacent paragraphs to distinguish different sub-items under the same number. This implementation method creates a structured mapping path between the main text citation relationship and the footnote description content, transforming cross-segment information that originally relied on manual location into a search process with number constraints, semantic constraints, and context constraints. This provides clear criteria for the footnote search scope, the semantic orientation of the item, and the distinction between sub-items with the same number, providing a stable input foundation for subsequent footnote filtering, supplementation, and closure construction.

[0025] Example 3 Please refer to Figure 2 Specifically: the reference closure generation module includes an item filtering unit and a closure splicing unit; The item filtering unit is used to extract the item name and main narrative content from the source fragments cited in the main text, and to filter the annotation segments directly corresponding to the current item from the annotation candidate fragment set based on the item name. It then extracts the scope and definition of the current item from the annotation segments and the composition and reasons for change from the current item, generating supplementary annotation fragments.

[0026] The closing splicing unit is used to splice the main text section and the supplementary note section in the main text source fragment in a fixed order to form the expression fragment to be closed, and to perform supplementation and repair on the repeated expressions, pronoun references, missing subjects, missing objects and incomplete qualifiers in the expression fragment to be closed, so as to obtain a single closed expression; The supplementary adjustments include the following: Remove identical expressions that appear repeatedly in the main text and footnotes; Replace the pronouns with the specific names of the items; Add the subject of the item to the descriptive sentence that is missing a subject; To fill in the missing object in the concluding sentence; Complete any incomplete qualifiers regarding consolidated figures, ending balances, and the deduction of relevant impacts. The completed and revised single closed expression, along with the item name, the source fragment in the main text, the supplementary fragment in the footnote, the source location in the main text, and the source location in the footnote, are encapsulated to generate a closed citation structure. The pronouns refer to this item, the above content, the relevant circumstances, and this part.

[0027] In this embodiment, the item filtering unit first extracts the item name and main text content from the source text citation fragment. Using the item name as the retrieval core, it locates the annotation sentence segment directly corresponding to the current item in the annotation candidate fragment set. Simultaneously, it performs forward and backward adjacency expansion processing on this annotation sentence segment, extracting the scope and definition forward and the composition and reasons for change backward, forming an annotation supplement fragment. This serves to avoid extracting only an isolated annotation sentence, resulting in incomplete supplementary content. Through adjacency expansion, it incorporates the explanatory content in the annotation that truly supports the expression of the current item. Subsequently, the closing and splicing unit splices the main text segment in the source text citation fragment with the annotation supplement fragment in a fixed order of "item main sentence first, supplementary explanation second," forming a pending closing. The process involves combining explanatory fragments and supplementing them with repetitive expressions, pronouns such as "this item" or "the above content," incomplete subjects, incomplete objects, and incomplete qualifiers to obtain a single closed expression that is semantically continuous and structurally complete. This closed expression is then encapsulated along with the item name, the source fragment in the main text, the supplementary fragment in the footnote, the source location in the main text, and the source location in the footnote to generate a closed citation structure. Through this process, the explanatory information that was originally scattered in the main text and footnotes is organized into a consistent, clearly directed, and fully defined structure within the same explanatory unit. This completes the reorganization of cross-segment citation content and the construction of closed expressions. At the same time, the process unifies and organizes referential relationships, incomplete sentences, and qualifying information, ensuring that the generated closed citation structure has a clear item-directing relationship and a complete explanatory chain.

[0028] Example 4 Please refer to Figure 2 Specifically: the closed state determination module includes a correspondence comparison unit and a closed determination unit; The correspondence comparison unit is used to perform a correspondence comparison between the source fragment cited in the main text and the supplementary fragment in the footnote, as follows; Extract the event name from the source fragment cited in the main text and extract the explanatory notes from the supplementary notes fragment. Perform same name matching and synonym matching on the event name and the explanatory notes fragment. When the event name and the explanatory notes fragment are the same, or when the event name and the explanatory notes fragment refer to the same disclosure event in the financial statement, determine that the event name has a corresponding explanation in the supplementary notes fragment. Locate the citation trigger from the source fragment in the main text, and check whether there is a description of the scope, composition, reasons for change, and range corresponding to the name of the current matter after the location of the citation trigger. When the description of the name of the current matter has been formed after the citation trigger, it is determined that the citation trigger has been replaced by the description. Pronoun retrieval is performed on single closed expressions to extract the referential pronouns and to search backwards along the preceding statements in the single closed expression for the item name or annotation object corresponding to the referential pronoun. When there are referential pronouns that cannot be returned to the item name or annotation object, it is determined that there is an unreturned object in the single closed expression. The limiting component retrieval is performed on the limiting phrase completion fragment, and the time limitation, scope limitation, scope limitation and object limitation are extracted. It is then checked whether each limiting component forms a complete correspondence with the item name and the object of the annotation in the annotation completion fragment. When any limiting component is missing any of the limiting object, limiting scope, limiting time or limiting scope, it is determined that there is incomplete limiting content in the limiting phrase completion fragment. Extract the conclusion expression from a single closed expression and check whether there are corresponding explanations of scope, composition, reasons for change, and range before and after the current conclusion expression. If the conclusion expression does not match any explanation, it is determined that there is a break in the single closed expression.

[0029] The closure determination unit is used to perform closure state determination on each referenced closed structure based on the correspondence comparison processing result, and summarize the closure state determination results to generate closure state marking results. The closure state determination is as follows: When the item name has a corresponding description, the reference trigger has been replaced by the description content, there is no unreferential object in the single closed expression, the qualifier supplement fragment is complete and there is no broken expression in the single closed expression, the current reference closed structure is marked as a complete closed structure; When the item name has a partial corresponding description, but there are still broken expressions in a single closed expression, the currently referenced closed structure is marked as a partially closed structure; When any of the following conditions are met—that the item name does not have a corresponding description, the reference trigger is not replaced by the description content, or there is still a broken expression in a single closed expression—the current reference closed structure is marked as a closed structure to be filled back.

[0030] In this embodiment, the correspondence comparison unit performs name matching and synonym matching between the item name and the object described in the footnote, detects substitution relationships of citation triggers, retrieves the reference path of pronouns, searches for the completeness of time limits, scope limits, caliber limits, and object limits in the qualifier supplement fragment, and detects the correspondence between the conclusion expression and the caliber description, composition description, explanation of reasons for change, and scope description, forming a multi-dimensional correspondence comparison result. Based on this, the closure determination unit determines the result based on whether the item name has a corresponding description, whether the citation trigger is replaced by the description content, whether there is an unreferenced object in a single closed expression, and whether the qualifier supplement fragment is complete. The system determines whether there are broken expressions in a single closed expression, performs a closure status judgment on each cited closed structure, and generates a closure status marker result. The cited closed structure is divided into a complete closed structure, a partially closed structure, or a closed structure to be supplemented. This implementation method unifies the semantic correspondence, referential relationship, and limiting relationship between the source fragment of the text citation and the supplementary fragment of the footnote into the structured judgment process. It makes the expression completeness of the cited closed structure have a distinguishable judgment basis, and provides hierarchical input conditions for the construction of extended reconstruction fragments and constraint reconstruction processing in the subsequent abstract primitive reconstruction module. This transforms the abstract generation process from directly splicing the original fragments to a hierarchical processing flow based on the closure status marker result.

[0031] Example 5 Please refer to Figure 2 Specifically: the summary primitive reconstruction module includes a back-complementation extension unit and a primitive reconstruction unit; The back-complementary extension unit is used to read the complete closed structure and the partially closed structure in the closed state marking result; Extract the item name, text source fragment, footnote supplement fragment, text source position, and footnote source position of each referenced closed structure from the complete closed structure set. At the same time, extract the preceding sentence and following sentence of the text corresponding to the text source position of each referenced closed structure in the partial closed structure. Simultaneously, read the footnote adjacent sub-item title and footnote subsequent explanation sentence corresponding to the footnote source position in each referenced closed structure. Add the preceding sentence, following sentence, adjacent sub-item title, and subsequent explanation sentence of the footnote into the current referenced closed structure to generate an extended reconstructed fragment.

[0032] The primitive reconstruction unit is used to perform constraint reconstruction on each referenced closed structure and extended reconstruction fragment in the complete closed structure to obtain the main body of the matter, the main body of the description and the main body of the limitation, and combine them in a fixed order to generate the summary generation primitive; The constraint reconstruction process includes the following: The item name is fixed as a single main object as the item's main body; Extract the scope, composition, reasons for change, and range of the items from the supplementary and expanded fragments of the notes, and organize them into the main body of the explanation according to the principle of unique correspondence between the objects of the notes; Extract the limiting object, limiting scope, limiting time, or limiting definition that directly corresponds to the event name from the limiting phrase supplement or extended reconstruction phrase, and organize them into the limiting main body.

[0033] In this embodiment, the summary primitive reconstruction module first uses the backfilling extension unit to read the complete closed structure and the partially closed structure from the closed state marking results. For the complete closed structure, it directly extracts the item name, the text citation source fragment, the footnote supplement fragment, the text source position, and the footnote source position. Simultaneously, for the partially closed structure, it obtains the corresponding preceding and following sentences based on the text source position, and obtains the adjacent sub-item titles and subsequent explanatory sentences based on the footnote source position. This content is then added to the original citation closed structure to form an extended reconstruction fragment. Subsequently, the primitive reconstruction unit performs constraint reconstruction processing on the complete closed structure and the extended reconstruction fragment, fixing the item name as the item trunk and extracting the footnote supplement fragment... The process extracts the scope, composition, reasons for change, and scope from the segment and extended reconstruction fragments and organizes them into the explanatory backbone. At the same time, it extracts the limiting object, limiting scope, limiting time, and limiting scope to form the limiting backbone. These are then combined in a fixed order to generate the summary generation primitive. Through this process, the expression breaks and scattered descriptions existing in the original closed structure of the citation are reorganized into a structure of matter backbone, explanatory backbone, and limiting backbone. This allows the summary content to form a clear, traceable, and complete expression form around the matter name. The process forms a continuous execution link from distinguishing the closed state, expanding and supplementing, to structural rearrangement, giving the summary generation primitive a unified structural framework and clear semantic boundaries.

[0034] Example 6 Please refer to Figure 2 Specifically: the summary verification module includes a matter clarity verification unit, a description correspondence verification unit, and a limitation completeness verification unit; The item clarification verification unit is used to extract the financial statement item names in the item trunk and detect whether there are any pronouns in the item trunk and whether the item trunk contains two or more parallel financial statement item names at the same time. When there are no pronouns in the main body of a matter and it corresponds to only a single financial statement matter name, the main body of the matter is clearly defined. When there are pronouns in the main body of a matter, or when there are two or more parallel financial statement matter names, the main body of the matter is determined to be unclear, and the current summary generation primitive is marked as a primitive to be reviewed; The description correspondence verification unit is used to extract the annotation objects and item names in the description trunk, and to perform consistency matching between the annotation objects and the item names in the item trunk. When all the notes to the financial statements belong to the financial items corresponding to the main subject matter, it is determined that the notes to the financial statements are consistent with the main subject matter. When any note in the main body of a matter does not belong to the financial items of the main body of the matter, the current note is written into the information to be removed and the removal process is performed. After the removal process, if the main body of the matter retains at least one note of financial item, the main body of the matter is deemed valid; if the main body of the matter does not retain any note of financial item, the main body of the current summary generation element is deemed invalid, and the current summary generation element is marked as a element to be reviewed. The defined integrity verification unit is used to extract the time limit, scope limit, caliber limit and object limit in the defined backbone; Perform corresponding verification on each of the specified contents and the item names in the main body of the item and the supplementary notes in the main body of the description; The main body of a limiting structure is deemed qualified when at least one item name or a description of an item is included in the limiting structure. If the constraint trunk does not contain the constraint content of the item name or description object, the constraint trunk is deemed unqualified. In this case, the current summary generation primitive is returned to the summary primitive reconstruction module to re-execute the constraint reconstruction process.

[0035] In this embodiment, the summary verification module extracts the names of financial reporting items from the main body of the items through the item clarity verification unit, and detects pronouns and parallel financial reporting item names to distinguish whether the main body of the items has a single pointing relationship. Subsequently, the description correspondence verification unit extracts the notes and descriptions in the description main body and performs consistency matching with the item names in the main body of the items. Descriptions that do not belong to the corresponding financial items are written with information to be removed and removed. The validity of the description main body is determined based on the retention results. On this basis, the limitation completeness verification unit further extracts the time limitation, scope limitation, caliber limitation and object limitation in the limitation main body, and performs correspondence verification with the item names in the main body of the items and the notes and descriptions in the description main body. Summary generation primitives with missing correspondence are rolled back to the summary primitive reconstruction module for reprocessing. Through the above processing flow, a clear correspondence is formed between the main body of the summary generation, the main body of the description, and the main body of the limitation. The structural consistency check of the summary generation results and the rollback of unqualified content are completed. Compared with the processing method that does not distinguish the relationship between the main body of the summary and the object of the footnote description, this process forms structural constraints in terms of the relationship between the subject, the attribution of the description content, and the matching of the limiting information, making the summary expression more complete and verifiable in terms of content correspondence and limiting structure.

[0036] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended technical solutions and their equivalents.

Claims

1. A financial statement summary generation system based on artificial intelligence-generated content, characterized in that: It includes a text parsing module, a reference closure generation module, a closure state determination module, a summary primitive reconstruction module, and a summary verification module; The text parsing module is used to divide the financial statement text into sections, extract the source fragments cited in the main text, and generate a set of candidate fragments for footnotes; The reference closure generation module is used to filter the corresponding note segments from the note candidate segments according to the item name, and expand them to form note supplement segments. The text reference source segment and the note supplement segment are sequentially spliced ​​together and supplemented to generate a reference closure structure. The closed state determination module is used to perform a correspondence comparison on the referenced closed structure, and mark the referenced closed structure as a complete closed structure, a partially closed structure, or a closed structure to be filled in according to the comparison result; The summary primitive reconstruction module is used to construct extended reconstruction fragments and perform constraint reconstruction processing with reference closure structures to generate summary generation primitives. The summary verification module is used to verify the clarity of the execution items, the correspondence of the descriptions, and the completeness of the constraints of the summary generation primitives.

2. The financial statement summary generation system based on artificial intelligence-generated content according to claim 1, characterized in that: The text parsing module includes a segment parsing unit and a search condition generation unit; The segment parsing unit is used to read the financial report text, divide the financial report text into the main text segment, the footnote segment and the title segment according to the title line, paragraph line and footnote line in the financial report text, and search for the statement containing the footnote pointing expression in the main text segment, extract the sentences and segments that are connected to the current statement, and form the main text citation source segment. The notes refer to the following expressions: see notes, refer to notes, see notes for details, see notes for specifics, and see notes to the financial statements for relevant content.

3. The financial statement summary generation system based on artificial intelligence-generated content according to claim 2, characterized in that: The retrieval condition generation unit is used to extract the footnote number information, item name, current title content and continuation description in adjacent paragraphs from the text citation source fragment, and use them as footnote retrieval conditions to retrieve multiple corresponding footnote candidate contents in the footnote section, generate a footnote candidate fragment set, and then bind each text citation source fragment to its corresponding footnote candidate fragment set to form an original associated object.

4. The financial statement summary generation system based on artificial intelligence-generated content according to claim 3, characterized in that: The reference closure generation module includes an item filtering unit and a closure splicing unit; The item filtering unit is used to extract the item name and main narrative content from the source fragments cited in the main text, and to filter the annotation segments directly corresponding to the current item from the annotation candidate fragment set based on the item name. It then extracts the scope and definition of the current item from the annotation segments and the composition and reasons for change from the current item, generating supplementary annotation fragments.

5. The financial statement summary generation system based on artificial intelligence-generated content according to claim 4, characterized in that: The closing splicing unit is used to splice the main text section and the supplementary note section in the main text source fragment in a fixed order to form the expression fragment to be closed, and to perform supplementation and repair on the repeated expressions, pronoun references, missing subjects, missing objects and incomplete qualifiers in the expression fragment to be closed, so as to obtain a single closed expression; The completed and revised single closed expression, along with the item name, the source fragment in the main text, the supplementary fragment in the footnote, the source location in the main text, and the source location in the footnote, are encapsulated to generate a closed citation structure.

6. The financial statement summary generation system based on artificial intelligence-generated content according to claim 5, characterized in that: The closure state determination module includes a correspondence comparison unit and a closure determination unit; The correspondence comparison unit is used to perform a correspondence comparison between the source fragment cited in the main text and the supplementary fragment in the footnote, as follows; Extract the event name from the source fragment cited in the main text and extract the explanatory notes from the supplementary notes fragment. Perform same name matching and synonym matching on the event name and the explanatory notes fragment. When the event name and the explanatory notes fragment are the same, or when the event name and the explanatory notes fragment refer to the same disclosure event in the financial statement, determine that the event name has a corresponding explanation in the supplementary notes fragment. Locate the citation trigger from the source fragment in the main text, and check whether there is a description of the scope, composition, reasons for change, and range corresponding to the name of the current matter after the location of the citation trigger. When the description of the name of the current matter has been formed after the citation trigger, it is determined that the citation trigger has been replaced by the description. Pronoun retrieval is performed on single closed expressions to extract the referential pronouns and to search backwards along the preceding statements in the single closed expression for the item name or annotation object corresponding to the referential pronoun. When there are referential pronouns that cannot be returned to the item name or annotation object, it is determined that there is an unreturned object in the single closed expression. The limiting component retrieval is performed on the limiting phrase completion fragment, and the time limitation, scope limitation, scope limitation and object limitation are extracted. It is then checked whether each limiting component forms a complete correspondence with the item name and the object of the annotation in the annotation completion fragment. When any limiting component is missing any of the limiting object, limiting scope, limiting time or limiting scope, it is determined that there is incomplete limiting content in the limiting phrase completion fragment. Extract the conclusion expression from a single closed expression and check whether there are corresponding explanations of scope, composition, reasons for change, and range before and after the current conclusion expression. If the conclusion expression does not match any explanation, it is determined that there is a break in the single closed expression.

7. The financial statement summary generation system based on artificial intelligence-generated content according to claim 6, characterized in that: The closure determination unit is used to perform closure state determination on each referenced closed structure based on the correspondence comparison processing result, and summarize the closure state determination results to generate closure state marking results. The closure state determination is as follows: When the item name has a corresponding description, the reference trigger has been replaced by the description content, there is no unreferential object in the single closed expression, the qualifier supplement fragment is complete and there is no broken expression in the single closed expression, the current reference closed structure is marked as a complete closed structure; When the item name has a partial corresponding description, but there are still broken expressions in a single closed expression, the currently referenced closed structure is marked as a partially closed structure; When any of the following conditions are met—that the item name does not have a corresponding description, the reference trigger is not replaced by the description content, or there is still a broken expression in a single closed expression—the current reference closed structure is marked as a closed structure to be filled back.

8. The financial statement summary generation system based on artificial intelligence-generated content according to claim 7, characterized in that: The summary primitive reconstruction module includes a back-extension unit and a primitive reconstruction unit; The back-complementary extension unit is used to read the complete closed structure and the partially closed structure in the closed state marking result; Extract the item name, text source fragment, footnote supplement fragment, text source position, and footnote source position of each referenced closed structure from the complete closed structure set. At the same time, extract the preceding sentence and following sentence of the text corresponding to the text source position of each referenced closed structure in the partial closed structure. Simultaneously, read the footnote adjacent sub-item title and footnote subsequent explanation sentence corresponding to the footnote source position in each referenced closed structure. Add the preceding sentence, following sentence, adjacent sub-item title, and subsequent explanation sentence of the footnote into the current referenced closed structure to generate an extended reconstructed fragment.

9. The financial statement summary generation system based on artificial intelligence-generated content according to claim 8, characterized in that: The primitive reconstruction unit is used to perform constraint reconstruction on each referenced closed structure and extended reconstruction fragment in the complete closed structure to obtain the main body of the matter, the main body of the description and the main body of the limitation, and combine them in a fixed order to generate the summary generation primitive; The constraint reconstruction process includes the following: The item name is fixed as a single main object as the item's main body; Extract the scope, composition, reasons for change, and range of the items from the supplementary and expanded fragments of the notes, and organize them into the main body of the explanation according to the principle of unique correspondence between the objects of the notes; Extract the limiting object, limiting scope, limiting time, or limiting definition that directly corresponds to the event name from the limiting phrase supplement or extended reconstruction phrase, and organize them into the limiting main body.

10. The financial statement summary generation system based on artificial intelligence-generated content according to claim 9, characterized in that: The summary verification module includes a matter clarity verification unit, a description correspondence verification unit, and a limitation completeness verification unit; The item clarification verification unit is used to extract the financial statement item names in the item trunk and detect whether there are any pronouns in the item trunk and whether the item trunk contains two or more parallel financial statement item names at the same time. When there are no pronouns in the main body of a matter and it corresponds to only a single financial statement matter name, the main body of the matter is clearly defined. When there are pronouns in the main body of a matter, or when there are two or more parallel financial statement matter names, the main body of the matter is determined to be unclear, and the current summary generation primitive is marked as a primitive to be reviewed; The description correspondence verification unit is used to extract the annotation objects and item names in the description trunk, and to perform consistency matching between the annotation objects and the item names in the item trunk. When all the notes to the financial statements belong to the financial items corresponding to the main subject matter, it is determined that the notes to the financial statements are consistent with the main subject matter. When any note in the main body of a matter does not belong to the financial items of the main body of the matter, the current note is written into the information to be removed and the removal process is performed. After the removal process, if the main body of the matter retains at least one note of financial item, the main body of the matter is deemed valid; if the main body of the matter does not retain any note of financial item, the main body of the current summary generation element is deemed invalid, and the current summary generation element is marked as a element to be reviewed. The defined integrity verification unit is used to extract the time limit, scope limit, caliber limit and object limit in the defined backbone; Perform corresponding verification on each of the specified contents and the item names in the main body of the item and the supplementary notes in the main body of the description; The main body of a limiting structure is deemed qualified when at least one item name or a description of an item is included in the limiting structure. If the constraint trunk does not contain the constraint content of the item name or description object, the constraint trunk is deemed unqualified. In this case, the current summary generation primitive is returned to the summary primitive reconstruction module to re-execute the constraint reconstruction process.

Images