Construction Methods and Systems for Report Document Generation Model

Abstract

This application provides a method and system for constructing a report document generation model. Based on report document-related knowledge extracted from banking professional knowledge documents, a report document knowledge base is constructed. The banking professional knowledge documents include at least one of the following: historical report documents, banking regulations and rules, and banking business standards. By identifying and analyzing the report document knowledge base, N report modules are obtained, each used to generate different parts of the report document. The N report modules are then integrated to construct a report document generation agent for the report document generation model. Based on a pre-configured report drafting workflow and the report document generation agent, the report document generation model is generated. This improves the quality and efficiency of the generated report documents.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method and system for constructing a report document generation model. Background Technology

[0002] In the daily operations of commercial banks, the writing of various reports is extremely frequent, such as write-off investigation reports, industry analysis reports, and market research reports. Traditional report drafting methods heavily rely on manual writing and proofreading. This method has significant drawbacks, including inefficiency and poor quality.

[0003] With the rise of large language model technology, its powerful text generation capabilities have led to its widespread application in many fields, including report writing. While general-purpose large language models can address basic issues such as typos, omissions, and missing subjects to some extent, generative AI relies on complex algorithms and models, resulting in a degree of randomness in the generated content. The financial industry has strict reporting requirements, and this randomness makes it difficult for the generated content to directly meet those requirements. Repeated manual proofreading and revisions are still necessary, leading to issues such as poor quality and low efficiency.

[0004] Therefore, improving the quality and efficiency of model-generated report documents is an urgent problem to be solved. Summary of the Invention

[0005] In view of this, the purpose of this application is to provide a method and system for constructing a report document generation model, so as to improve the quality and efficiency of the report document generated by the model.

[0006] In conjunction with the first aspect of this application, a method for constructing a report document generation model is provided, the method comprising:

[0007] A report document knowledge base is constructed based on the report document-related knowledge extracted from banking professional knowledge documents. The banking professional knowledge documents include at least one of the following: historical report documents, banking system rules and regulations, and banking business standards.

[0008] By identifying and analyzing the report document knowledge base, N report modules corresponding to the report document knowledge base are obtained. Different report modules are used to generate different parts of the content in the report document.

[0009] By integrating the N reporting modules, a report document generation intelligent agent is constructed to form a report document generation model;

[0010] Based on the pre-configured report drafting workflow and the report document generation agent, the report document generation model is generated. The report drafting workflow is used to determine the execution order and interaction method between the sub-agents corresponding to each report module in the report document generation agent.

[0011] Optionally, the construction of a report document knowledge base based on banking professional knowledge documents includes:

[0012] The initial bank professional knowledge document is parsed and the parsing result of the initial bank professional knowledge document is processed by a document format processing tool to generate the bank professional knowledge document. The initial bank professional knowledge document comes from at least one channel, such as the bank's internal system or a regulatory database. The initial bank professional knowledge document includes at least one of a text document and an image document. The text document is parsed based on natural language processing, and the image document is parsed based on optical character recognition (OCR).

[0013] The content entities in the banking professional knowledge document are located by means of predefined grammar rules and grammar patterns, and the metadata of the banking professional knowledge document is determined from the entities based on the frequency, positional relationship and context information of word segmentation in the banking professional knowledge document;

[0014] A pre-trained word vector model is used to obtain the lexical units and sentence units included in the banking professional knowledge document. Based on the semantic understanding model, the semantic connotation of the lexical units and sentence units is deeply mined according to the context information of the lexical units, so as to map the semantic information in the banking professional knowledge document into the corresponding semantic feature vector.

[0015] By using data association technology, the original text of the bank professional knowledge document, the determined metadata, and the generated semantic feature vector are associated and integrated through the identifiers of each document in the bank professional knowledge document to generate the report document knowledge base.

[0016] Optionally, the step of identifying and analyzing the report document knowledge base to divide it into N report modules corresponding to the report document knowledge base includes:

[0017] Based on the semantic feature vectors and metadata in the report document knowledge base, the semantic similarity between documents in the report document knowledge base is calculated, and the documents in the report document knowledge base are divided into different candidate sets of report modules based on the semantic similarity.

[0018] The structural and syntactic features of the documents included in the candidate sets of each report module are analyzed to determine the hierarchical structure and logical relationship of different parts of each document. Key entity information is extracted by identifying structural elements such as titles, paragraphs, and lists in each document. The key entity information is used to clarify the core descriptive information of different parts of each document.

[0019] Based on the hierarchical structure, logical connections, and core descriptive information, the documents in each candidate set of report modules are summarized and a module description for each candidate set of report modules is generated. The module description is used to reflect the role and function of the modules in the candidate set of report modules in the report document.

[0020] Based on the module descriptions in each candidate set of report modules, the content of the documents in each candidate set of report modules, and the preset scoring rules, target document fragments in each candidate set of report modules are determined as module examples for each candidate set of report modules.

[0021] Based on the module descriptions and module examples in each candidate set of report modules, the module verification criteria for each candidate set of report modules are determined from the preset verification rules. The module verification criteria are used to clarify the generation accuracy and completeness of each module in the candidate set of report modules.

[0022] Based on the consistency of module descriptions, module examples, and module verification standards for each module in the candidate sets of each report module, N report modules corresponding to the report document knowledge base are determined in each candidate set of each report module.

[0023] Optionally, the report document generation agent includes a report document synthesis agent, a report document content generation sub-agent corresponding to each of the report modules, and a report document content verification sub-agent; the report document generation agent that constructs the report document generation model based on the N report modules includes:

[0024] Based on the module description of each report module, the functional features, core features, and expected content features of each report module are obtained, and the functional features, core features, and expected content features are input into the large language model to guide the large language model to generate prompt word templates corresponding to each report module.

[0025] Based on the prompt word template and the module example of the report module, the initial language model is fine-tuned and trained to generate the report document content generation sub-agent corresponding to each of the report modules;

[0026] The module verification standards corresponding to each of the report modules are converted into executable code logic to construct the verification algorithm corresponding to each of the module verification standards, and the report document content verification sub-intelligent agent of the corresponding report module is generated based on each of the verification algorithms.

[0027] By deeply mining the overall framework of the report document and the relationships between each report module, and based on the overall document framework and the relationships, the workflow of each report document content generation sub-agent and each report document content verification sub-agent is determined, and a report document synthesis agent related to the workflow is generated;

[0028] The report document generation agent of the report document generation model is constructed based on the report document synthesis agent, the report document content generation sub-agent corresponding to each report module, and the report document content verification sub-agent corresponding to each report module.

[0029] Optionally, the step of generating the report document generation model based on the report compilation workflow and the report document generation agent includes:

[0030] By formally describing the report compilation workflow, the nodes, edges, and transformation rules between the nodes in the report compilation workflow are obtained. The nodes represent work nodes, the edges represent workflow directions, and the transformation rules represent the triggering conditions for work node transformations.

[0031] Based on the nodes, edges, and transformation rules, determine the execution order and interaction method of each report document content generation sub-agent, each report document content verification sub-agent, and the report document synthesis agent;

[0032] Based on the execution order and the interaction method, the agent scheduling parameters in the initial report document generation model are adjusted to generate the report document generation model.

[0033] Optionally, the optical character recognition (OCR)-based parsing of the image document includes:

[0034] The original image data of the image document is obtained, and the original image data is preprocessed by at least one of grayscale processing, binarization processing, and noise reduction processing to generate standardized image data.

[0035] The text region in the standardized image data is located based on the edge detection algorithm, and the tilted or distorted text region is geometrically corrected using the affine transformation algorithm to generate the corrected text region image.

[0036] The text region image is subjected to text recognition using a pre-trained OCR recognition model. The text sequence in the text region image is extracted, and the text sequence is corrected by character-level confidence verification and contextual semantic verification to generate preliminary parsed text.

[0037] Based on the layout structure features of the image document, the paragraphs, headings, and tables in the preliminary parsed text are logically segmented, and structured parsed text is generated by aligning and matching them with the paragraph marks, heading levels, and table identifiers of the text document.

[0038] The structured parsed text and the parsing results of the text document are subjected to cross-modal semantic consistency verification. Based on the verification results, the semantically ambiguous segments in the structured parsed text are manually annotated and corrected to generate the target parsed text of the image document.

[0039] Optionally, calculating the semantic similarity between documents in the report document knowledge base based on the semantic feature vector and the metadata includes:

[0040] Sentence vectors corresponding to sentence units of each document are extracted from the semantic feature vectors, and the sentence vectors of each document are aggregated into document-level semantic vectors by a weighted average algorithm.

[0041] The first similarity value between the document-level semantic vectors is calculated based on the cosine similarity algorithm, and the keyword overlap rate between the documents is calculated based on the keyword set in the metadata as the second similarity value.

[0042] The first similarity value and the second similarity value are weighted and fused according to a preset weight allocation strategy to generate a comprehensive semantic similarity score between the documents.

[0043] A document similarity matrix is ​​constructed based on the comprehensive semantic similarity score, and unsupervised clustering analysis is performed on the document similarity matrix based on the spectral clustering algorithm to generate document clustering results;

[0044] Based on the density distribution characteristics and inter-cluster distance characteristics of each cluster in the document clustering results, the clustering parameters are dynamically adjusted and the document clusters are re-divided until the variance of the comprehensive semantic similarity score of documents within each cluster is lower than a preset threshold, and the candidate set of the report module is generated.

[0045] Optionally, the step of inputting the functional features, the core point features, and the expected content features into the large language model to guide the large language model to generate prompt word templates corresponding to each of the report modules includes:

[0046] The functional features are converted into module role description statements, the core key features are converted into a keyword constraint list, and the expected content features are converted into content generation format specifications.

[0047] Based on preset template generation rules, the module role description statement, the keyword constraint list, and the content generation format specification are sequentially concatenated into an initial prompt word sequence;

[0048] The initial prompt word sequence is syntactically optimized and semantically enhanced using the large language model to generate a set of candidate prompt word templates.

[0049] Using the module example as a verification sample, simulated content is generated based on the candidate prompt word template through the large language model, and the content overlap and logical coherence index between the simulated content and the module example are calculated.

[0050] Candidate prompt word templates with content overlap higher than a first preset threshold and logical coherence index higher than a second preset threshold are selected as the prompt word templates corresponding to the reporting module.

[0051] Optionally, when locating content entities in the banking expertise document using predefined syntax rules and patterns, the method further includes:

[0052] Identify table regions with fixed syntax structures in the banking professional knowledge documents, and extract the row and column headers in the table regions as entity attribute tags;

[0053] Analyze the numerical data type and unit symbol of the cell data within the table area, and associate the numerical data type and unit symbol with the entity attribute label to form a structured field;

[0054] Detect logical connectors between consecutive paragraphs in the banking professional knowledge document, and divide the dependency relationship chain of the content entities based on the logical connectors;

[0055] The structured fields and dependency chains are mapped to a preset entity relationship graph to generate a multi-level semantic association network of the content entities;

[0056] Based on the multi-level semantic association network, entity alignment is performed on the cross-referenced chapter content in the banking professional knowledge document to eliminate entity nodes with duplicate descriptions.

[0057] In conjunction with the second aspect of this application, a system for constructing a report document generation model is provided. The system includes a machine-readable storage medium and a processor. The machine-readable storage medium stores machine-executable instructions. When the processor executes the machine-executable instructions, the system implements the aforementioned method for constructing a report document generation model.

[0058] In conjunction with a third aspect of this application, a computer-readable storage medium is provided, wherein computer-executable instructions are stored therein, and when the computer-executable instructions are executed, the aforementioned method for constructing a report document generation model is implemented.

[0059] In conjunction with the fourth aspect of this application, a computer program product is provided, wherein when the computer program is executed by a processor, it implements the aforementioned method for constructing the report document generation model.

[0060] Combining any of the above aspects, a report document knowledge base is constructed based on report document-related knowledge extracted from banking professional knowledge documents. These banking professional knowledge documents include at least one of the following: historical report documents, banking regulations and rules, and banking business standards. By identifying and analyzing the report document knowledge base, N report modules are obtained, each used to generate different parts of the report document. These N report modules are then integrated to construct a report document generation agent for a report document generation model. Based on a pre-configured report drafting workflow and the report document generation agent, the report document generation model is generated. Using this model to generate report documents improves both the quality and efficiency of report document generation. Attached Figure Description

[0061] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained in conjunction with these drawings without creative effort.

[0062] Figure 1 This application provides a flowchart illustrating the method for constructing a report document generation model. Detailed Implementation

[0063] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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.

[0064] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.

[0065] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0066] Figure 1 This diagram illustrates a flowchart of a method for constructing a report document generation model according to an embodiment of this application. It should be understood that in other embodiments, the order of some steps in the method for constructing the report document generation model of this embodiment can be shared according to actual needs, or some steps can be omitted or maintained. The detailed method for constructing the report document generation model includes:

[0067] S110. Construct a report document knowledge base based on the report document-related knowledge extracted from banking professional knowledge documents.

[0068] Among them, banking professional knowledge documents include at least one of the following: historical report documents, banking system rules and regulations, and banking business standards.

[0069] In this step, the initial banking expertise document can be parsed first. This initial document may come from sources such as the bank's internal systems or regulatory databases, and may include text documents and image documents. Text documents can be parsed using natural language processing, while image documents can be parsed using optical character recognition (OCR).

[0070] Next, the parsed results can be formatted using document formatting tools to generate banking professional knowledge documents. Then, content entities within the banking professional knowledge documents are located using predefined grammatical rules and patterns. Based on the frequency, positional relationships, and contextual information of word segments in the document, the metadata of the banking professional knowledge documents is determined from these entities. A pre-trained word vector model is then used to obtain lexical units and sentence units included in the document. Based on a semantic understanding model, the semantic connotations of lexical units and sentence units are deeply mined according to their contextual information, mapping the semantic information in the document into corresponding semantic feature vectors. Finally, using data association technology, the original text of the banking professional knowledge documents, the determined metadata, and the generated semantic feature vectors are linked and integrated through the identifiers of each document in the document, generating a report document knowledge base.

[0071] For example, suppose a bank wants to build a knowledge base of credit business reporting documents. It can collect historical credit business reports from the past few years, along with relevant credit business rules and regulations, as initial professional banking documents. After parsing and processing these documents, it can locate content entities related to loan amounts, loan terms, etc., determine metadata, mine semantic feature vectors, and finally integrate them to generate the knowledge base.

[0072] S120. By identifying and analyzing the report document knowledge base, N report modules corresponding to the report document knowledge base are obtained.

[0073] Different reporting modules are used to generate different parts of the report document.

[0074] In this step, we can first calculate the semantic similarity between documents based on the semantic feature vectors and metadata in the report document knowledge base, and then divide the documents into different candidate sets of report modules based on the similarity.

[0075] Then, we carefully analyze the structural features (such as paragraph layout and chapter settings) and grammatical relationship features (such as subject-verb-object structure and use of conjunctions) of the documents in each candidate set to clarify the hierarchical structure and logical relationship of different parts of the document. By identifying structural elements such as titles, paragraphs, and lists in the document, we extract key entity information, which can clearly reflect the core descriptive content of different parts of the document.

[0076] Subsequently, based on the hierarchical structure, logical connections, and core descriptive information, the documents in each candidate set are summarized to generate module descriptions that reflect the role and function of the set in the report document.

[0077] Then, based on the module description, document content, and preset scoring rules of each candidate set, the target document fragment is selected from the candidate set as a module example.

[0078] Finally, based on the module description and module examples, the module verification criteria for each candidate set are determined from the preset verification rules. Taking into account the consistency between the module description, module examples and module verification criteria, N report modules corresponding to the report document knowledge base are determined from each candidate set.

[0079] S130, Integrating N report modules to construct a report document generation intelligent agent for a report document generation model.

[0080] In this step, for each report module, functional features, core elements, and expected content features can be extracted based on its module description. These features are then input into the large language model to guide it in generating prompt word templates for each report module.

[0081] Based on the generated prompt word templates and module examples of the report module, the initial language model is fine-tuned and trained to obtain the report document content generation sub-agents corresponding to each report module.

[0082] The module verification standards corresponding to each report module are transformed into code logic that can be executed by a computer. Verification algorithms for each module verification standard are constructed, and a report document content verification sub-intelligent agent for each report module is generated based on these algorithms.

[0083] By deeply analyzing the overall framework of the report document and the inherent relationships between each report module, the workflow of each report document content generation sub-agent and each report document content verification sub-agent is determined, and then a report document synthesis agent adapted to this workflow is generated.

[0084] Finally, the report document synthesis agent, the report document content generation sub-agent corresponding to each report module, and the report document content verification sub-agent corresponding to each report module are integrated to construct the report document generation agent of the report document generation model.

[0085] For example, suppose we want to build an intelligent agent for generating an annual comprehensive business report document for a bank. The report module is divided into three modules: "Savings Business," "Credit Business," and "Intermediary Business." Taking the "Savings Business" report module as an example:

[0086] First, based on the module description, its functional characteristics can be determined as showcasing the overall situation of savings business. Key features include savings types, interest rate levels, and customer groups. The expected content features are detailed data reports and trend analysis. These characteristics are then input into a large language model to generate prompt templates such as "As a presenter of bank savings business, generate content based on keywords such as savings types, interest rate levels, and customer groups, following a format of data reports plus trend analysis."

[0087] Then, the initial language model is fine-tuned and trained using the prompt word template and the "Savings Business" module example (such as the content about savings business in past savings business reports) to obtain the sub-agent for generating the content of the savings business report document.

[0088] Using the same method, corresponding content generation sub-intelligent agents can be generated for the "credit business" and "intermediary business" modules respectively.

[0089] For the report document content verification sub-agent, taking the "Savings Business" module as an example, the module verification standards (such as data accuracy, trend analysis rationality, etc.) are translated into code logic to build a verification algorithm. For example, it checks whether the data is consistent with the bank's system records and whether the trend analysis conforms to business logic, and generates a "Savings Business Report Document Content Verification Sub-agent" based on this. Verification sub-agents for other modules are then generated accordingly.

[0090] Next, the overall framework of the report document is analyzed, determining that the content generation sub-agents of each module will first generate the corresponding content, then the verification sub-agent will perform content verification, and finally the report document synthesis agent will integrate and format the content that has passed the verification of each module. Following this workflow, the report document synthesis agent is generated.

[0091] Finally, the report document synthesis agent, the report document content generation sub-agents corresponding to the three report modules, and the report document content verification sub-agents are integrated to construct the report document generation agent of the report document generation model.

[0092] S140. Based on the pre-configured report compilation workflow and report document generation agent, generate the report document generation model.

[0093] The report drafting workflow is used to determine the execution order and interaction method between the sub-agents corresponding to each report module in the report document generation agent.

[0094] In this step, the report compilation workflow can be formally described first to obtain the nodes (representing various work tasks), edges (representing the direction of the workflow), and the transformation rules between nodes (i.e. the triggering conditions for work node transformation).

[0095] Based on the acquired nodes, edges, and transformation rules, the execution order and interaction method of each report document content generation sub-agent, each report document content verification sub-agent, and the report document synthesis sub-agent are determined.

[0096] Finally, the agent scheduling parameters in the initial report document generation model are adjusted according to the determined execution order and interaction method, thereby generating the report document generation model.

[0097] Based on the above steps, and using the report document-related knowledge extracted from banking professional knowledge documents, a report document knowledge base is constructed. These banking professional knowledge documents include at least one of the following: historical report documents, banking regulations and rules, and banking business standards. Through identification and analysis of the report document knowledge base, N report modules are identified, each used to generate different parts of the report document. These N report modules are then integrated to construct a report document generation agent for a report document generation model. Based on a pre-configured report drafting workflow and the report document generation agent, the report document generation model is generated. This improves the quality and efficiency of the model-generated report documents.

[0098] In one possible implementation, the step of building a report document knowledge base based on banking expertise documents includes:

[0099] The initial bank professional knowledge document is parsed and the parsing result of the initial bank professional knowledge document is processed by a document format processing tool to generate the bank professional knowledge document. The initial bank professional knowledge document comes from at least one channel, such as the bank's internal system or a regulatory database. The initial bank professional knowledge document includes at least one of a text document and an image document. The text document is parsed based on natural language processing, and the image document is parsed based on optical character recognition (OCR).

[0100] The content entities in the banking professional knowledge document are located by means of predefined grammar rules and grammar patterns, and the metadata of the banking professional knowledge document is determined from the entities based on the frequency, positional relationship and context information of word segmentation in the banking professional knowledge document;

[0101] A pre-trained word vector model is used to obtain the lexical units and sentence units included in the banking professional knowledge document. Based on the semantic understanding model, the semantic connotation of the lexical units and sentence units is deeply mined according to the context information of the lexical units, so as to map the semantic information in the banking professional knowledge document into the corresponding semantic feature vector.

[0102] By using data association technology, the original text of the bank professional knowledge document, the determined metadata, and the generated semantic feature vector are associated and integrated through the identifiers of each document in the bank professional knowledge document to generate the report document knowledge base.

[0103] In this embodiment, initial banking expertise documents can be obtained from at least one channel, such as the bank's internal system or regulatory databases. These documents include at least one type: text documents and image documents. For text documents, natural language processing (NLP) techniques are used to parse them, converting the text into a structured representation that computers can understand, such as lexical analysis and syntactic analysis. For image documents, OCR technology is used for parsing. The image is first preprocessed, such as by grayscale conversion, binarization, and noise reduction, to improve image quality. Then, the text region is located and geometrically corrected. Finally, an OCR recognition model is used to extract the text information. After parsing, document formatting tools are used to format the parsing results, unifying the data format and generating a compliant banking expertise document.

[0104] Then, based on predefined grammatical rules and patterns, content entities with specific structures or characteristics can be searched and located in banking professional knowledge documents. For example, table areas with fixed grammatical structures can be identified, and the row and column headers can be extracted as entity attribute labels; the numerical types and unit symbols of the table cell data can be analyzed and associated with attribute labels to form structured fields. Simultaneously, logical connectors between consecutive paragraphs in the document can be detected to segment dependency chains of content entities. Afterwards, combining the frequency of word segmentation, their positional relationships within the document, and contextual information, metadata representing key document information can be filtered and determined from these located entities. This metadata can include document themes, key indicators, important time nodes, etc.

[0105] Next, pre-trained word vector models can be used to decompose the text in banking professional knowledge documents into lexical units and sentence units, generating corresponding vector representations for each lexical unit and sentence unit. These pre-trained models have learned rich linguistic knowledge on large-scale text data, enabling them to provide effective semantic representations for words and sentences. Then, based on semantic understanding models and combined with the contextual information of lexical units, the semantic connotations of each lexical unit and sentence unit are further mined and understood. Semantic understanding models can capture information such as semantic relationships between words and semantic structure of sentences. Finally, this deeply understood semantic information is mapped into corresponding semantic feature vectors. These vectors serve as a digital representation of the document's semantic information, facilitating subsequent analysis and processing.

[0106] Then, using data association technology, relationships are established between the original content of each document, the previously determined metadata, and the generated semantic feature vectors through the unique identifiers of each document in the banking professional knowledge documents. For example, each document can be assigned an identifier, and this identifier can be used to link the tables storing the original document text, the metadata, and the semantic feature vectors in the database. Through this association and integration, different types of interrelated information are combined to form a complete report document knowledge base, which can comprehensively store and represent various knowledge related to report documents.

[0107] For example, suppose a bank wants to build a knowledge base for risk management report documents. First, it can retrieve historical risk management reports (text documents) from its internal systems over the past few years and relevant risk management regulations (partially in image format) from a regulatory database as initial bank professional knowledge documents. Natural language processing is used to parse the text documents, and OCR is used to parse the image documents. Then, a document formatting tool is used to standardize the format and generate bank professional knowledge documents. Next, based on predefined grammar rules, content entities such as risk indicator tables and risk event description paragraphs are located in the documents. Metadata such as risk type and risk occurrence time is determined by combining word segmentation information. Then, a pre-trained word vector model and semantic understanding model are used to convert the words and sentences in the documents into semantic feature vectors. Finally, the original document text, metadata, and semantic feature vectors are linked and integrated through document identification to form a risk management report document knowledge base, providing comprehensive data support for the construction of subsequent report generation models.

[0108] In one possible implementation, the step of identifying and analyzing the report document knowledge base to divide it into N report modules corresponding to the report document knowledge base includes:

[0109] Based on the semantic feature vectors and metadata in the report document knowledge base, the semantic similarity between documents in the report document knowledge base is calculated, and the documents in the report document knowledge base are divided into different candidate sets of report modules based on the semantic similarity.

[0110] The structural and syntactic features of the documents included in the candidate sets of each report module are analyzed to determine the hierarchical structure and logical relationship of different parts of each document. Key entity information is extracted by identifying structural elements such as titles, paragraphs, and lists in each document. The key entity information is used to clarify the core descriptive information of different parts of each document.

[0111] Based on the hierarchical structure, logical connections, and core descriptive information, the documents in each candidate set of report modules are summarized and a module description for each candidate set of report modules is generated. The module description is used to reflect the role and function of the modules in the candidate set of report modules in the report document.

[0112] Based on the module descriptions in each candidate set of report modules, the content of the documents in each candidate set of report modules, and the preset scoring rules, target document fragments in each candidate set of report modules are determined as module examples for each candidate set of report modules.

[0113] Based on the module descriptions and module examples in each candidate set of report modules, the module verification criteria for each candidate set of report modules are determined from the preset verification rules. The module verification criteria are used to clarify the generation accuracy and completeness of each module in the candidate set of report modules.

[0114] Based on the consistency of module descriptions, module examples, and module verification standards for each module in the candidate sets of each report module, N report modules corresponding to the report document knowledge base are determined in each candidate set of each report module.

[0115] In this embodiment, sentence vectors corresponding to sentence units of each document are first extracted from the semantic feature vectors contained in the report document knowledge base. These sentence vectors reflect the semantic information of the sentences. For example, the vectors of the sentences "Bank deposit interest rate adjustments affect customer savings behavior" and "Banks launch new wealth management products to attract customers" will show semantic differences. Next, these sentence vectors are aggregated using a weighted average algorithm to obtain document-level semantic vectors. In this process, weights can be set according to factors such as the importance of sentences in the document. For example, sentences in paragraphs that explain core viewpoints are given higher weights. At the same time, the keyword set is found from the metadata, and the keyword overlap rate is obtained by calculating the keyword overlap between documents. For example, if the keywords of document A are "bank loans, interest rates, enterprises" and the keywords of document B are "loan business, interest rate adjustments, SMEs", the degree of keyword overlap between them can be calculated. Then, according to a pre-set weight allocation strategy, the similarity value (such as cosine similarity) calculated based on the document-level semantic vectors and the keyword overlap rate are weighted and fused to obtain a comprehensive semantic similarity score between documents. This score can more comprehensively reflect the semantic similarity between documents. A document similarity matrix is ​​constructed based on the comprehensive semantic similarity score, where each element represents the similarity between documents. Finally, an unsupervised clustering analysis is performed on this matrix using a spectral clustering algorithm, grouping documents with high semantic similarity together to form preliminary document clustering results. Then, based on features such as the density distribution and inter-cluster distance of each cluster, the clustering parameters are dynamically adjusted, and the document clusters are re-divided until the variance of the comprehensive semantic similarity score of documents within each cluster is lower than a pre-set threshold. This yields different candidate sets for report modules. For example, in a knowledge base containing various banking business reports, documents related to credit card business, due to their high semantic similarity, will be grouped into the same candidate set for report modules.

[0116] Then, the documents in the candidate set of each report module can be structurally analyzed. For example, syntactic analysis techniques from natural language processing can be used to parse the grammatical structure of the documents and determine the grammatical relationships between sentences, such as subject-verb-object. The hierarchical structure can be determined by identifying the heading levels in the document, such as which second-level headings and paragraphs are contained under a first-level heading. For list structures, the logical relationships between list items can be analyzed, such as whether they are parallel or progressive. Key entity information can be extracted from these structural elements. For example, in a document about bank loan business, key entity information such as "large enterprise customers" and "small and medium-sized enterprise customers" can be extracted from the title "Loan Customer Classification" and the paragraph descriptions below. This information clarifies the core descriptive content of this part of the document.

[0117] Next, the hierarchical structure, logical connections, and core descriptive information of the documents in a candidate set of report modules can be integrated. For example, the documents in a candidate set of report modules mainly revolve around the bank's deposit business, describing information such as deposit product types, interest rate policies, and customer groups from different perspectives. By sorting and summarizing this information, the generated module description could be: "This module mainly introduces information related to various types of bank deposit businesses, including the characteristics of different deposit products, applicable customer groups, and corresponding interest rate policies, providing basic information for a comprehensive understanding of the bank's deposit business, and serving as the core descriptive part of the deposit business section in the report document," thus reflecting the module's role and function in the report document.

[0118] Next, the most relevant parts of the document content can be identified based on the module description. For example, for the deposit business module mentioned above, the module description emphasizes key information such as deposit product types and interest rate policies. Therefore, the document content can be searched for segments that contain this key information clearly and completely. Pre-defined scoring rules can be formulated based on factors such as the frequency of keyword occurrences and the logical coherence between the segment and the module description. For instance, if a document segment details multiple deposit products and their corresponding interest rate policies, closely matching the module description, and receives a high score according to the scoring rules, it can be identified as a module example in the candidate module set for this report.

[0119] The module description, module example, and module validation criteria are compared within each candidate set of report modules. If these three elements in a candidate set are consistent in terms of theme, key information, etc. (e.g., all revolving around bank loan business, the module description accurately summarizes the loan business-related content, the module example effectively reflects this content, and the module validation criteria are also consistent), then this candidate set of report modules can be identified as a report module. By performing this consistency check on all candidate sets of report modules, N report modules corresponding to the report document knowledge base are ultimately determined.

[0120] In one possible implementation, the report document generation agent includes a report document synthesis agent, a report document content generation sub-agent corresponding to each of the report modules, and a report document content verification sub-agent; the report document generation agent that constructs the report document generation model based on the N report modules includes:

[0121] Based on the module description of each report module, the functional features, core features, and expected content features of each report module are obtained, and the functional features, core features, and expected content features are input into the large language model to guide the large language model to generate prompt word templates corresponding to each report module.

[0122] Based on the prompt word template and the module example of the report module, the initial language model is fine-tuned and trained to generate the report document content generation sub-agent corresponding to each of the report modules;

[0123] The module verification standards corresponding to each of the report modules are converted into executable code logic to construct the verification algorithm corresponding to each of the module verification standards, and the report document content verification sub-intelligent agent of the corresponding report module is generated based on each of the verification algorithms.

[0124] By deeply mining the overall framework of the report document and the relationships between each report module, and based on the overall document framework and the relationships, the workflow of each report document content generation sub-agent and each report document content verification sub-agent is determined, and a report document synthesis agent related to the workflow is generated;

[0125] The report document generation agent of the report document generation model is constructed based on the report document synthesis agent, the report document content generation sub-agent corresponding to each report module, and the report document content verification sub-agent corresponding to each report module.

[0126] In this embodiment, the functional features, core features, and expected content features of each reporting module can be extracted based on the module description. For example, for the "Bank Credit Business Analysis" module, features such as "analyzing and interpreting credit business data" can be extracted from the module description. These features are then input into a large language model to generate corresponding prompt word templates, such as templates like "analyzing and interpreting the core points of credit business and presenting the results in a specific form".

[0127] Next, the initial language model is fine-tuned and trained using the generated prompt word templates and module examples. Taking past reports from the "Bank Credit Business Analysis" module as examples, the initial language model learns the content style, structure, and professional expression of the examples, enabling it to generate content that conforms to the characteristics of the module, and thus generate the report document content generation sub-agent corresponding to each report module.

[0128] Next, the module validation standards for each reporting module are converted into executable code logic. For example, the validation standards for the "Bank Credit Business Analysis" module, such as covering data from the past year and referencing authoritative indicators, are used to write corresponding code logic to build a validation algorithm. Based on this, a report document content validation sub-agent is generated to automatically validate the generated report content.

[0129] Further analysis is conducted to explore the overall framework of the report document and the relationships between each report module. For example, the arrangement order and causal relationships of each business module in a bank's comprehensive report are used to determine the workflow of the content generation and verification sub-agents for each report document. First, the content is generated by the sub-agents of each module, then verified by the verification sub-agents before being passed to the report document synthesis agent, which integrates the content according to the framework and relationships.

[0130] Finally, the report document synthesis agent, the report document content generation sub-agents corresponding to each report module, and the report document content verification sub-agents are integrated. By establishing a unified interface, the sub-agents combine according to a predetermined collaboration method to form a complete agent system, thus constructing the report document generation agent of the report document generation model to generate report documents efficiently and accurately.

[0131] In one possible implementation, generating the report document generation model based on the report drafting workflow and the report document generation agent includes:

[0132] By formally describing the report compilation workflow, the nodes, edges, and transformation rules between the nodes in the report compilation workflow are obtained. The nodes represent work nodes, the edges represent workflow directions, and the transformation rules represent the triggering conditions for work node transformations.

[0133] Based on the nodes, edges, and transformation rules, determine the execution order and interaction method of each report document content generation sub-agent, each report document content verification sub-agent, and the report document synthesis agent;

[0134] Based on the execution order and the interaction method, the agent scheduling parameters in the initial report document generation model are adjusted to generate the report document generation model.

[0135] In this embodiment, it is assumed that a bank annual comprehensive report needs to be generated, involving multiple reporting modules such as credit business, deposit business, and financial status. A report document generation intelligent agent has been built, and now a report document generation model is generated in conjunction with the report drafting workflow.

[0136] First, the report compilation workflow is formally described. Taking the compilation process of a bank's annual comprehensive report as an example, the workflow starts at "Data Collection," collecting raw data from various business systems. Nodes such as "Credit Business Analysis," "Deposit Business Analysis," and "Financial Status Analysis" are carried out in parallel, processing their respective business operations. The workflow then converges at the "Report Integration" node, integrating the analysis results from each part. Finally, it reaches the "Report Review" node. Edges represent the workflow direction, pointing from the "Data Collection" node to each business analysis node, from each business analysis node to the "Report Integration" node, and from the "Report Integration" node to the "Report Review" node. Regarding the conversion rules, after the "Data Collection" node completes data collection, it triggers the start of work for each business analysis node; after each business analysis node completes its analysis, it triggers the "Report Integration" node; and after "Report Integration" is completed, it triggers the "Report Review" node.

[0137] Next, based on the aforementioned nodes, edges, and transformation rules, the execution order and interaction method of the agents are determined. The report document generation sub-agents for "Credit Business Analysis," "Deposit Business Analysis," and "Financial Status Analysis" are launched in parallel after "Data Collection" is completed, generating the report content for their respective business modules. Subsequently, the corresponding report document content verification sub-agents verify the generated content. After successful verification, the results are passed to the report document synthesis agent for integration. Once integrated, the report document synthesis agent submits the integrated report to the process responsible for the review stage for approval.

[0138] Finally, based on the determined execution order and interaction method, the agent scheduling parameters in the initial report document generation model are adjusted. For example, the number of parallel processing operations for each business content generation sub-agent is set, and the waiting time for data transfer between the verification sub-agent and the synthesis agent is set. After these adjustments, the final report document generation model is generated, which can coordinate the work of each sub-agent in the report document generation agent according to the workflow of compiling the bank's annual comprehensive report, and generate reports efficiently and accurately.

[0139] In one possible implementation, the optical character recognition (OCR)-based parsing of the image document includes:

[0140] The original image data of the image document is obtained, and the original image data is preprocessed by at least one of grayscale processing, binarization processing, and noise reduction processing to generate standardized image data.

[0141] The text region in the standardized image data is located based on the edge detection algorithm, and the tilted or distorted text region is geometrically corrected using the affine transformation algorithm to generate the corrected text region image.

[0142] The text region image is subjected to text recognition using a pre-trained OCR recognition model. The text sequence in the text region image is extracted, and the text sequence is corrected by character-level confidence verification and contextual semantic verification to generate preliminary parsed text.

[0143] Based on the layout structure features of the image document, the paragraphs, headings, and tables in the preliminary parsed text are logically segmented, and structured parsed text is generated by aligning and matching them with the paragraph marks, heading levels, and table identifiers of the text document.

[0144] The structured parsed text and the parsing results of the text document are subjected to cross-modal semantic consistency verification. Based on the verification results, the semantically ambiguous segments in the structured parsed text are manually annotated and corrected to generate the target parsed text of the image document.

[0145] In this embodiment, it is assumed that we want to parse an image document of a bank loan contract.

[0146] First, the original image data of the document can be obtained. This image may have issues such as color interference, uneven brightness, and noise. To generate standardized image data, it can be converted to grayscale, simplifying the image information. For example, the RGB value of each pixel in the image can be converted into a single grayscale value using a specific algorithm. Next, binarization is performed, setting an appropriate threshold to divide the pixels in the image into foreground (text portion) and background, making the text portion black and the background white, further highlighting the text information. Simultaneously, noise reduction processing is performed, using algorithms such as Gaussian filtering to remove random noise from the image, making the image clearer.

[0147] Based on edge detection algorithms, such as the Canny algorithm, text regions in standardized image data are located to find the text's outline boundaries. If the text region is found to be tilted or distorted, an affine transformation algorithm is used for geometric correction. For example, if the text in a document image is tilted due to scanning angle issues, an affine transformation algorithm is used to calculate the transformation matrix to correct the text region to a horizontal and regular state, generating a corrected text region image.

[0148] Using a pre-trained OCR recognition model, text recognition is performed on the corrected text region image to extract the text sequence. During the recognition process, a character-level confidence score is assigned to each recognized character, and characters with low confidence scores are marked. Then, combined with contextual semantic verification, error correction is performed on the marked characters based on the semantic logic of the surrounding text, generating preliminary parsed text.

[0149] Based on the layout structure of the loan contract image document, such as paragraph spacing and title font size, the paragraphs, titles, and tables in the initial parsed text are logically segmented. The segmented content is then aligned and matched with the paragraph marks, title levels, and table identifiers of the same loan contract text document to make the parsing results more structured, generating structured parsed text.

[0150] Finally, cross-modal semantic consistency checks are performed on the parsed results of the structured parsed text and the text document. For example, the key information such as loan amount and interest rate parsed from the image is checked to see if they are consistent with those in the text document. If there are semantically ambiguous segments, such as a clause in the image parsing that is unclear or differs from the text document, manual annotation and correction are performed, and finally, the target parsed text of the image document is generated.

[0151] In one possible implementation, calculating the semantic similarity between documents in the report document knowledge base based on the semantic feature vector and the metadata includes:

[0152] Sentence vectors corresponding to sentence units of each document are extracted from the semantic feature vectors, and the sentence vectors of each document are aggregated into document-level semantic vectors by a weighted average algorithm.

[0153] The first similarity value between the document-level semantic vectors is calculated based on the cosine similarity algorithm, and the keyword overlap rate between the documents is calculated based on the keyword set in the metadata as the second similarity value.

[0154] The first similarity value and the second similarity value are weighted and fused according to a preset weight allocation strategy to generate a comprehensive semantic similarity score between the documents.

[0155] A document similarity matrix is ​​constructed based on the comprehensive semantic similarity score, and unsupervised clustering analysis is performed on the document similarity matrix based on the spectral clustering algorithm to generate document clustering results;

[0156] Based on the density distribution characteristics and inter-cluster distance characteristics of each cluster in the document clustering results, the clustering parameters are dynamically adjusted and the document clusters are re-divided until the variance of the comprehensive semantic similarity score of documents within each cluster is lower than a preset threshold, and the candidate set of the report module is generated.

[0157] In this embodiment, it is assumed that the report document knowledge base contains multiple documents about different banking businesses, such as reports related to credit business, deposit business, and wealth management business.

[0158] First, sentence vectors corresponding to each sentence unit in the document can be extracted from the semantic feature vectors. For example, in a credit business document, there are sentences such as "Recently, banks have increased their credit support for small and medium-sized enterprises" and "The increase in the non-performing loan rate has affected the revenue of credit business." Their corresponding sentence vectors can be extracted separately. Next, these sentence vectors are aggregated into a document-level semantic vector using a weighted average algorithm. For example, sentences in important paragraphs are given higher weights; sentences explaining core credit policies have a weight of 0.6, while other supplementary sentences have a weight of 0.4. This is used to calculate the document-level semantic vector of the credit business document.

[0159] Then, the first similarity value between the semantic vectors of each document is calculated based on the cosine similarity algorithm. For example, the document-level semantic vectors of credit business documents and deposit business documents are calculated to obtain a numerical value representing their similarity at the semantic vector level. Simultaneously, the keyword overlap rate is calculated based on the keyword set in the metadata as a second similarity value. For instance, if the keywords for the credit business document are "credit limit, SMEs, non-performing loans," and the keywords for the deposit business document are "deposit interest rate, customer savings, credit business," their keyword overlap rate is calculated.

[0160] Based on the preset weighting strategy, assuming the first similarity value has a weight of 0.6 and the second similarity value has a weight of 0.4, the two are weighted and fused to generate a comprehensive semantic similarity score between documents.

[0161] A document similarity matrix is ​​constructed based on the comprehensive semantic similarity score, where each element represents the similarity between documents. For example, the element in the i-th row and j-th column represents the similarity between the i-th and j-th documents. Unsupervised clustering analysis is then performed on the document similarity matrix using a spectral clustering algorithm, grouping documents with high semantic similarity together to form preliminary document clustering results. This might result in clusters such as credit-related documents forming one group, deposit-related documents forming another, and so on.

[0162] Finally, examine the density distribution and inter-cluster distance characteristics of each cluster in the document clustering results. If documents in a cluster are scattered (low density) and close to other clusters, the clustering effect is poor. Dynamically adjust clustering parameters, such as changing the cluster radius or minimum sample size, to re-divide the document clusters. Repeat this process until the variance of the comprehensive semantic similarity score of documents within each cluster is lower than a preset threshold, ultimately generating a candidate set of report modules, such as different candidate sets for credit business, deposit business, and wealth management business.

[0163] In one possible implementation, the step of inputting the functional features, the core feature, and the expected content feature into a large language model to guide the large language model in generating prompt word templates corresponding to each of the report modules includes:

[0164] The functional features are converted into module role description statements, the core key features are converted into a keyword constraint list, and the expected content features are converted into content generation format specifications.

[0165] Based on preset template generation rules, the module role description statement, the keyword constraint list, and the content generation format specification are sequentially concatenated into an initial prompt word sequence;

[0166] The initial prompt word sequence is syntactically optimized and semantically enhanced using the large language model to generate a set of candidate prompt word templates.

[0167] Using the module example as a verification sample, simulated content is generated based on the candidate prompt word template through the large language model, and the content overlap and logical coherence index between the simulated content and the module example are calculated.

[0168] Candidate prompt word templates with content overlap higher than a first preset threshold and logical coherence index higher than a second preset threshold are selected as the prompt word templates corresponding to the reporting module.

[0169] In this embodiment, it is assumed that there is a "Bank Customer Credit Assessment Report" module. First, its functional characteristic can be determined as "assessing the credit status of bank customers", which is converted into a module role description statement: "You are a professional bank customer credit assessor"; the core key characteristics include "credit record, income level, and debt status", which is converted into a keyword constraint list: "credit record, income level, debt status"; the expected content characteristics are "giving a credit score out of 100 points and classifying it into levels, and explaining the assessment basis in paragraphs", which is converted into a content generation format specification: "giving a customer credit score out of 100 points, classifying the credit level (excellent, good, average, poor), and explaining the assessment basis in clear paragraphs".

[0170] Next, based on the preset template generation rules, the module role description statement, keyword constraint list, and content generation format specifications are sequentially assembled into the initial prompt word sequence: "You are a professional bank customer credit assessor. Based on key points such as credit history, income level, and debt situation, give the customer a credit score out of a total of 100 points, classify the credit level (excellent, good, average, poor), and explain the basis of each assessment in clear paragraphs."

[0171] Then, this initial sequence of prompt words is input into a large language model, which performs syntactic optimization and semantic enhancement to generate a set of candidate prompt word templates. For example, the model may optimize the sentence expression to make the semantics clearer and more accurate, generating multiple candidate prompt word templates with different expressions but similar meanings.

[0172] The module example of this report module was used as a validation sample; this module example is a past bank customer credit assessment report. Simulated content was generated based on each candidate prompt word template using a large language model. After generating the simulated content, the content overlap and logical coherence indicators between the simulated content and the module example were calculated. For example, it was checked whether the simulated content mentioned the key credit assessment factors in the module example, and whether the logic in explaining the assessment basis was clear and organized.

[0173] Finally, assuming a first preset threshold of 70% and a second preset threshold of 80%, a template with a content overlap of over 70% and a logical coherence index of over 80% can be selected from the candidate prompt templates as the prompt templates for the "Bank Customer Credit Assessment Report" module. This prompt template will be used for subsequent fine-tuning and training of the initial language model to generate report content that meets the requirements of this module.

[0174] In one possible implementation, when locating content entities in the banking expertise document using predefined grammar rules and patterns, the method may further include:

[0175] Identify table regions with fixed syntax structures in the banking professional knowledge documents, and extract the row and column headers in the table regions as entity attribute tags;

[0176] Analyze the numerical data type and unit symbol of the cell data within the table area, and associate the numerical data type and unit symbol with the entity attribute label to form a structured field;

[0177] Detect logical connectors between consecutive paragraphs in the banking professional knowledge document, and divide the dependency relationship chain of the content entities based on the logical connectors;

[0178] The structured fields and dependency chains are mapped to a preset entity relationship graph to generate a multi-level semantic association network of the content entities;

[0179] Based on the multi-level semantic association network, entity alignment is performed on the cross-referenced chapter content in the banking professional knowledge document to eliminate entity nodes with duplicate descriptions.

[0180] In this embodiment, it is assumed that there exists a professional knowledge document from a bank regarding its loan business. First, table areas with a fixed grammatical structure can be identified within the document. For example, the document might contain a table displaying information on different types of loan products, with row headings such as "Loan Product Name," "Interest Rate," and "Loan Term," and column headings corresponding to different loan product types. These row and column headings are extracted and used as entity attribute tags, representing different dimensions of loan product-related information.

[0181] Next, we analyze the numeric data types and unit symbols of the cells within the table area. In the "Interest Rate" column, the numeric data type of the cells is decimal, and the unit symbol is "%". We associate this with the entity attribute label "Interest Rate" to form a structured field such as "Interest Rate (Number Type: Decimal, Unit: %)". The numeric data type of the "Loan Term" column is integer, and the unit is "month", forming a structured field "Loan Term (Number Type: Integer, Unit: Month)".

[0182] Next, examine the logical connectors between consecutive paragraphs in the document. For example, in a paragraph describing the loan approval process, logical connectors such as "firstly," "secondly," and "then" appear. Based on these connectors, the dependency chain of content entities can be identified: first, customer information is collected; second, credit assessment is conducted; and third, approval decisions are made, clearly defining the sequence and dependencies between each step.

[0183] Next, the generated structured fields and dependency chains are mapped to a pre-defined entity relationship graph. This graph may already contain the relationship framework of various entities in the bank's loan business, integrating structured fields such as "loan product," "interest rate," and "loan term," as well as dependency chains of the loan approval process, to generate a multi-level semantic relationship network of content entities, making loan business-related information interconnected and hierarchically distinct within a single network.

[0184] Finally, based on this multi-level semantic association network, entity alignment is performed on cross-referenced content in the document. For example, the entity "high-quality customer" is mentioned in both the chapters introducing different loan products and the loan approval process. By aligning entities through the semantic association network, duplicate entity nodes are eliminated, information redundancy is avoided, and the knowledge structure of the entire document becomes clearer and more unified, providing a more efficient foundation for subsequent knowledge processing and application.

[0185] In the above embodiments, the system for building the report document generation model for performing the above method embodiments has at least one processor, a control module (chipset) coupled to at least one of the processors, a memory coupled to the control module, a non-volatile memory (NVM) / storage device coupled to the control module, at least one input / output device coupled to the control module, and a network interface coupled to the control module.

[0186] The processor may include at least one single-core or multi-core processor, and may include any combination of general-purpose processors or special-purpose processors (e.g., graphics processors, application processors, baseband processors, etc.). For some alternative implementations, the report document generation model construction system can serve as the gateway or other report document generation model construction system device described in the embodiments of this application.

[0187] In some alternative implementations, the system for building the report document generation model may include at least one computer-readable medium (e.g., memory or NVM / storage device) having instructions and at least one processor fused with the at least one computer-readable medium and configured to execute the instructions to implement the module thereby performing the actions described in this disclosure.

[0188] In one embodiment, the control module may include any suitable interface controller to provide any suitable interface to at least one of the processors and / or any suitable device or component communicating with the control module.

[0189] The control module may include a memory controller module to provide an interface to the memory. The memory controller module may be a hardware module, a software module, and / or a firmware module.

[0190] Memory can be used, for example, to load and store data and / or instructions in a system for building models of report documents. In one embodiment, memory may include any suitable volatile memory, such as suitable DRAM.

[0191] In one embodiment, the control module may include at least one input / output controller to provide an interface to the NVM / storage device and (at least one) input / output device.

[0192] For example, an NVM / storage device can be used to store data and / or instructions. An NVM / storage device may include any suitable non-volatile memory (e.g., flash memory) and / or may include any suitable (at least one) non-volatile storage device (e.g., at least one hard disk drive (HDD), at least one optical disc (CD) drive, and / or at least one digital universal optical disc (DVD) drive).

[0193] An NVM / storage device may include storage resources that are physically part of a device on which a building system for a report document generation model is mounted, or that can be accessed by the device without being part of the device. For example, an NVM / storage device may be accessed via a network through at least one input / output device.

[0194] At least one input / output device may provide an interface for the report document generation model building system to communicate with any other suitable device. The input / output device may include communication components, spelling components, sensor components, etc. A network interface may provide an interface for the report document generation model building system to communicate via at least one network. The report document generation model building system may wirelessly communicate with at least one component of a wireless network according to any standard and / or protocol of at least one wireless network standard and / or protocol, for example, by accessing a wireless network based on a communication standard.

[0195] In one embodiment, at least one of the processors may be integrated with the logic of at least one controller of the control module (e.g., a memory controller module). In one embodiment, at least one of the processors may be integrated with the logic of at least one controller of the control module to form a system-on-a-chip (SoC). In one embodiment, at least one of the processors may be integrated with the logic of at least one controller of the control module on the same die. In one embodiment, at least one of the processors may be integrated with the logic of at least one controller of the control module on the same die to form a system-on-a-chip (SoC).

[0196] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

[0197] This invention discloses a computer read storage medium that stores a computer program for electronic data interchange, wherein the computer program causes a computer to execute the steps in the method for constructing a report document generation model as described in the foregoing embodiments.

[0198] This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps in the method for constructing a report document generation model as described in the foregoing embodiments.

[0199] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0200] Through the detailed description of the above embodiments, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, including read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electronically erasable rewritable read-only memory (EEPROM), compact optical disc (CD-ROM) or other optical disc storage, disk storage, magnetic tape storage, or any other computer-readable medium that can be used to have or store data.

[0201] Finally, it should be noted that the above-disclosed embodiments are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a report document generation model, characterized in that, include: A report document knowledge base is constructed based on the report document-related knowledge extracted from banking professional knowledge documents. The banking professional knowledge documents include at least one of the following: historical report documents, banking system rules and regulations, and banking business standards. By identifying and analyzing the report document knowledge base, N report modules corresponding to the report document knowledge base are obtained. Different report modules are used to generate different parts of the content in the report document. By integrating the N reporting modules, a report document generation intelligent agent is constructed to form a report document generation model; Based on the pre-configured report drafting workflow and the report document generation agent, the report document generation model is generated. The report drafting workflow is used to determine the execution order and interaction method between the sub-agents corresponding to each report module in the report document generation agent. The step of generating the report document generation model based on the report compilation workflow and the report document generation agent includes: By formally describing the report compilation workflow, the nodes, edges, and transformation rules between the nodes in the report compilation workflow are obtained. The nodes represent work nodes, the edges represent workflow directions, and the transformation rules represent the triggering conditions for work node transformations. Based on the nodes, edges, and transformation rules, determine the execution order and interaction method of each report document content generation sub-agent, each report document content verification sub-agent, and the report document synthesis agent; Based on the execution order and the interaction method, the agent scheduling parameters in the initial report document generation model are adjusted to generate the report document generation model.

2. The method for constructing the report document generation model according to claim 1, characterized in that, The report document knowledge base, constructed based on report document-related knowledge extracted from banking professional knowledge documents, includes: The initial bank professional knowledge document is parsed and the parsing result of the initial bank professional knowledge document is processed by a document format processing tool to generate the bank professional knowledge document. The initial bank professional knowledge document comes from at least one channel, such as the bank's internal system or a regulatory database. The initial bank professional knowledge document includes at least one of a text document and an image document. The text document is parsed based on natural language processing, and the image document is parsed based on optical character recognition (OCR). The content entities in the banking professional knowledge document are located by means of predefined grammar rules and grammar patterns, and the metadata of the banking professional knowledge document is determined from the entities based on the frequency, positional relationship and context information of word segmentation in the banking professional knowledge document; A pre-trained word vector model is used to obtain the lexical units and sentence units included in the banking professional knowledge document. Based on the semantic understanding model, the semantic connotation of the lexical units and sentence units is deeply mined according to the context information of the lexical units, so as to map the semantic information in the banking professional knowledge document into the corresponding semantic feature vector. By using data association technology, the original text of the bank professional knowledge document, the determined metadata, and the generated semantic feature vector are associated and integrated through the identifiers of each document in the bank professional knowledge document to generate the report document knowledge base.

3. The method for constructing the report document generation model according to claim 2, characterized in that, The process involves identifying and analyzing the report document knowledge base to divide it into N report modules, including: Based on the semantic feature vectors and metadata in the report document knowledge base, the semantic similarity between documents in the report document knowledge base is calculated, and the documents in the report document knowledge base are divided into different candidate sets of report modules based on the semantic similarity. The structural and syntactic features of the documents included in the candidate sets of each report module are analyzed to determine the hierarchical structure and logical relationship of different parts of each document. Key entity information is extracted by identifying the title, paragraph, and list structure elements in each document. The key entity information is used to clarify the core descriptive information of different parts of each document. Based on the hierarchical structure, logical connections, and core descriptive information, the documents in each candidate set of report modules are summarized and a module description for each candidate set of report modules is generated. The module description is used to reflect the role and function of the modules in the candidate set of report modules in the report document. Based on the module descriptions in each candidate set of report modules, the content of the documents in each candidate set of report modules, and the preset scoring rules, target document fragments in each candidate set of report modules are determined as module examples for each candidate set of report modules. Based on the module descriptions and module examples in each candidate set of report modules, the module verification criteria for each candidate set of report modules are determined from the preset verification rules. The module verification criteria are used to clarify the generation accuracy and completeness of each module in the candidate set of report modules. Based on the consistency of module descriptions, module examples, and module verification standards for each module in the candidate sets of each report module, N report modules corresponding to the report document knowledge base are determined in each candidate set of each report module.

4. The method for constructing the report document generation model according to claim 3, characterized in that, The report document generation agent includes a report document synthesis agent, a report document content generation sub-agent corresponding to each of the report modules, and a report document content verification sub-agent; the report document generation agent that constructs the report document generation model based on the N report modules includes: Based on the module description of each report module, the functional features, core features, and expected content features of each report module are obtained, and the functional features, core features, and expected content features are input into the large language model to guide the large language model to generate prompt word templates corresponding to each report module. Based on the prompt word template and the module example of the report module, the initial language model is fine-tuned and trained to generate the report document content generation sub-agent corresponding to each of the report modules; The module verification standards corresponding to each of the report modules are converted into executable code logic to construct the verification algorithm corresponding to each of the module verification standards, and the report document content verification sub-intelligent agent of the corresponding report module is generated based on each of the verification algorithms. By deeply mining the overall framework of the report document and the relationships between each report module, and based on the overall document framework and the relationships, the workflow of each report document content generation sub-agent and each report document content verification sub-agent is determined, and a report document synthesis agent related to the workflow is generated; The report document generation agent of the report document generation model is constructed based on the report document synthesis agent, the report document content generation sub-agent corresponding to each report module, and the report document content verification sub-agent corresponding to each report module.

5. The method for constructing the report document generation model according to claim 2, characterized in that, The optical character recognition (OCR)-based parsing of the image document includes: The original image data of the image document is obtained, and the original image data is preprocessed by at least one of grayscale processing, binarization processing, and noise reduction processing to generate standardized image data. The text region in the standardized image data is located based on the edge detection algorithm, and the tilted or distorted text region is geometrically corrected using the affine transformation algorithm to generate the corrected text region image. The text region image is subjected to text recognition using a pre-trained OCR recognition model. The text sequence in the text region image is extracted, and the text sequence is corrected by character-level confidence verification and contextual semantic verification to generate preliminary parsed text. Based on the layout structure features of the image document, the paragraphs, headings, and tables in the preliminary parsed text are logically segmented, and structured parsed text is generated by aligning and matching them with the paragraph marks, heading levels, and table identifiers of the text document. The structured parsed text and the parsing results of the text document are subjected to cross-modal semantic consistency verification. Based on the verification results, the semantically ambiguous segments in the structured parsed text are manually annotated and corrected to generate the target parsed text of the image document.

6. The method for constructing the report document generation model according to claim 3, characterized in that, The step of calculating the semantic similarity between documents in the report document knowledge base based on the semantic feature vector and the metadata includes: Sentence vectors corresponding to sentence units of each document are extracted from the semantic feature vectors, and the sentence vectors of each document are aggregated into document-level semantic vectors by a weighted average algorithm. The first similarity value between the document-level semantic vectors is calculated based on the cosine similarity algorithm, and the keyword overlap rate between the documents is calculated based on the keyword set in the metadata as the second similarity value. The first similarity value and the second similarity value are weighted and fused according to a preset weight allocation strategy to generate a comprehensive semantic similarity score between the documents. A document similarity matrix is ​​constructed based on the comprehensive semantic similarity score, and unsupervised clustering analysis is performed on the document similarity matrix based on the spectral clustering algorithm to generate document clustering results; Based on the density distribution characteristics and inter-cluster distance characteristics of each cluster in the document clustering results, the clustering parameters are dynamically adjusted and the document clusters are re-divided until the variance of the comprehensive semantic similarity score of documents within each cluster is lower than a preset threshold, and the candidate set of the report module is generated.

7. The method for constructing a report document generation model according to claim 4, characterized in that, The step of inputting the functional features, the core feature features, and the expected content features into the large language model to guide the large language model in generating prompt word templates corresponding to each of the report modules includes: The functional features are converted into module role description statements, the core key features are converted into a keyword constraint list, and the expected content features are converted into content generation format specifications. Based on preset template generation rules, the module role description statement, the keyword constraint list, and the content generation format specification are sequentially concatenated into an initial prompt word sequence; The initial prompt word sequence is syntactically optimized and semantically enhanced using the large language model to generate a set of candidate prompt word templates. Using the module example as a verification sample, simulated content is generated based on the candidate prompt word template through the large language model, and the content overlap and logical coherence index between the simulated content and the module example are calculated. Candidate prompt word templates with content overlap higher than a first preset threshold and logical coherence index higher than a second preset threshold are selected as the prompt word templates corresponding to the reporting module.

8. The method for constructing a report document generation model according to claim 2, characterized in that, The process of locating content entities in the banking expertise document using predefined syntax rules and patterns also includes: Identify table regions with fixed syntax structures in the banking professional knowledge documents, and extract the row and column headers in the table regions as entity attribute tags; Analyze the numerical data type and unit symbol of the cell data within the table area, and associate the numerical data type and unit symbol with the entity attribute label to form a structured field; Detect logical connectors between consecutive paragraphs in the banking professional knowledge document, and divide the dependency relationship chain of the content entities based on the logical connectors; The structured fields and dependency chains are mapped to a preset entity relationship graph to generate a multi-level semantic association network of the content entities; Based on the multi-level semantic association network, entity alignment is performed on the cross-referenced chapter content in the banking professional knowledge document to eliminate entity nodes with duplicate descriptions.

9. A system for constructing a report document generation model, characterized in that, The report document generation model construction system includes a processor and a memory, the memory and the processor are connected, the memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the report document generation model construction method according to any one of claims 1-8.

Images