A business investigation report automatic generation method based on multi-modal semantic fusion

Abstract

The present application relates to the technical field of business investigation report generation, and particularly relates to a business investigation report automatic generation method based on multi-modal semantic fusion, which comprises the following steps: obtaining a customer demand document and demand clarification language communication content, analyzing the customer demand document to obtain document demand anchor points, analyzing the demand clarification language communication content to obtain clarification question and answer anchor points, associating the document demand anchor points with the clarification question and answer anchor points to generate an investigation anchor point chain; generating an original business workflow based on the investigation anchor point chain, and using the clarification question and answer anchor points to confirm, supplement and rewrite the business nodes in the original business workflow to generate a clarified and revised business workflow. The present application is a business investigation report automatic generation method based on multi-modal semantic fusion, which integrates customer demand documents and demand clarification language communication content, automatically generates and dynamically revises a business workflow, automatically performs a financial attribute caliber integrity check and generates a phased business caliber list.

Description

Technical Field

[0001] This invention relates to the field of business research report generation technology, and in particular to an automatic method for generating business research reports based on multimodal semantic fusion. Background Technology

[0002] As enterprises expand their digital transformation, higher demands are placed on the completeness, accuracy, and standardization of business research reports during the construction of business systems, data governance, and financial information systems. In actual business research, researchers typically need to simultaneously analyze customer requirement documents, meeting minutes, interview transcripts, and requirement clarification records. Based on multiple rounds of communication, they must organize business processes, rules, and data definitions to ultimately generate a business research report. Current methods for generating business research reports largely rely on manual reading of requirement documents, manual organization of meeting minutes, and manual drawing of business process diagrams. Researchers must manually extract business objectives, rules, data items, and process relationships from large amounts of unstructured text, voice communication content, and multiple rounds of requirement revision records, and then manually summarize these to generate business processes and research conclusions. Because customer requirements are often expressed colloquially, incompletely, use inconsistent business terminology, and undergo repeated revisions through multiple rounds of communication, traditional manual methods are prone to omissions of business nodes, misunderstandings of business logic, and inconsistencies in requirement versions.

[0003] Furthermore, in existing business research processes, customer requirement documents and requirement clarification communication content are usually stored independently, lacking a unified semantic association mechanism. Some business rules only appear in meeting Q&A or verbal communication but are not promptly synchronized to the formal requirement document, easily leading to inconsistencies between the subsequently generated business processes and the actual business scenarios. Meanwhile, traditional business process construction methods typically use fixed templates or manual drag-and-drop to generate business process diagrams, making it difficult to dynamically revise business nodes, input / output data items, and business rules based on requirement clarification content. This results in delayed business process updates, poor process consistency, and difficulties in requirement traceability. On the other hand, in financial-related business research scenarios, data definitions often differ between different departments, systems, and research stages. For example, the same field may have different statistical ranges, different measurement rules, or different coding methods. Existing business research reports typically lack automatic integrity verification mechanisms for financial attribute fields, relying mainly on manual verification of data items and business definitions. This easily leads to problems such as missing fields, inconsistent field formats, and inconsistent business definitions, thus affecting the accuracy of subsequent system construction, data governance, and financial analysis. Summary of the Invention

[0004] This invention provides an automatic business research report generation method based on multimodal semantic fusion, which integrates customer requirement documents and requirement clarification language communication content, automatically generates and dynamically revises business workflows, automatically performs financial attribute integrity verification, and generates a phased business scope list.

[0005] A method for automatically generating business research reports based on multimodal semantic fusion, characterized by the following steps: S1: Obtain customer requirement documents and requirement clarification language communication content, parse the customer requirement documents to obtain document requirement anchors, parse the requirement clarification language communication content to obtain clarification Q&A anchors, associate the document requirement anchors with the clarification Q&A anchors to generate a survey anchor chain; S2: Generate the original business workflow based on the survey anchor chain, and use the clarification question and answer anchors to confirm, supplement and rewrite the business nodes in the original business workflow to generate the clarification and revision business workflow; S3: Connect the data items in the survey anchor chain to the business nodes corresponding to the clarification and revision business workflow, perform a caliber integrity check based on the financial attributes of the data items, generate a phased business caliber list with missing markers, and generate a business survey report based on the clarification and revision business workflow and the phased business caliber list.

[0006] Optionally, S1 specifically includes: S11: Read customer requirement documents from the customer requirement management system through the preset data interface, and obtain requirement clarification language communication content, including questions and answers, from the requirement communication record library; S12: Use natural language processing technology to perform semantic parsing on the customer requirement document, extract the concept nodes that represent business objectives, functional scope, data entities and constraints, and mark each concept node as a document requirement anchor point; S13: Perform speech-to-text processing and dialogue structure analysis on the language communication content of the request clarification, identify each question and its corresponding answer, extract core semantic triples from each question-answer pair, and mark each semantic triple as a clarification question-answer anchor point; S14: Calculate the semantic similarity between each document requirement anchor point and each clarification question and answer anchor point, establish an association relationship for anchor point pairs with similarity exceeding a preset threshold, and arrange and combine the anchor point pairs in the association relationship according to the time order of requirement clarification or the business logic order to form a survey anchor point chain with a sequential dependency relationship.

[0007] Optionally, S12 specifically includes: The customer requirement document is preprocessed with text, including invalid symbol removal, terminology standardization, paragraph splitting, and sentence segmentation, to generate a standardized set of sentences. The standardized set of sentences is then segmented and part-of-speech tagging is performed to identify action words, business object words, attribute words, and constraint words in the sentences. Named entity recognition is then performed on the tagged sentences to extract business target entities, functional action entities, data object entities, and rule constraint entities. Dependency parsing is performed on the extracted entities to identify subject-predicate, action, modification, and conditional constraint relationships between entities, generating business semantic units. Semantic classification processing is performed on the business semantic units to uniformly map business semantic units with consistent semantics or the same business meaning. Concept nodes are constructed based on the classified business semantic units, and a unique identifier is assigned to each concept node. Each concept node is marked as a corresponding document requirement anchor point.

[0008] Optionally, S13 specifically includes: The requirements for speech format are clarified, and noise reduction, silence segment clipping, sampling rate unification, and speech segment segmentation are performed to obtain effective speech segments. The effective speech segments are input into the speech recognition model. Through acoustic feature extraction, acoustic feature encoding, contextual semantic modeling, character decoding, and text post-processing, the effective speech segments are converted into text transcription results with timestamps. Speaker separation and dialogue turn segmentation are performed on the text transcription results to identify the speech content of the customer side, the speech content of the survey side, and the corresponding speech time order. Based on the interrogative sentence structure, inquiry keywords, confirmation keywords, answer connectors, and the contextual semantic relationship between adjacent dialogue rounds, question-and-answer matching is performed on the text transcription results to obtain question-and-answer pairs consisting of the question content and the corresponding answer content. Semantic relation extraction is performed on each question-answer pair to identify the subject semantic item, predicate semantic item, and object semantic item in the question-answer pair and generate core semantic triples. Each core semantic triple is marked as a clarifying question-answer anchor point, and each clarifying question-answer anchor point is bound to the timestamp, speaker identifier, and business semantic relation information of the corresponding question-answer pair.

[0009] Optionally, S2 specifically includes: S21: According to the order of dependencies between anchors in the survey anchor chain, map each anchor to a candidate business node in the preset business component library; the business component library stores standard business node templates and connection rules between templates; S22: Based on the mapping results, taking the business node corresponding to the starting anchor point as the starting point and the business node corresponding to the ending anchor point as the ending point, the original business workflow is automatically arranged and generated according to the association direction between anchor points and the connection rules in the business component library. The original business workflow includes several business nodes and the flow path between nodes. S23: Traverse each business node in the original business workflow and retrieve the clarification question and answer anchor points associated with the corresponding business node; if there are associated clarification question and answer anchor points, perform confirmation, supplementation and rewriting operations based on the question and answer content carried by the clarification question and answer anchor points. S24: Reconnect the confirmed, supplemented or rewritten business nodes according to the original flow path to generate a clarification and revision business workflow.

[0010] Optionally, the execution confirmation, supplementation, and rewriting operations specifically include: Confirmation: When the question and answer content is consistent with the business semantics of the current business node and there is no new information, the corresponding node will be marked as confirmed. Additional information: When the Q&A content mentions new data items or business rules that are not covered by the current node but fall within the scope of the node's responsibilities, add the corresponding data fields or rule constraints to the node; Rewrite: When the question and answer content conflicts with the business semantics of the current node or explicitly requires modification of the node definition, update the node's business name, input and output data items, or processing logic based on the question and answer content.

[0011] Optionally, S21 specifically includes: Read the survey anchor chain and obtain the sequential dependencies between each anchor according to the order in the survey anchor chain; Read multiple standard business node templates from the business component library. The standard business node template includes a node name, node function description, input data items, output data items, node rule constraints, and node connection rules. Extract the business semantic features corresponding to each anchor point and extract the business semantic features corresponding to each standard business node template. Semantic matching is performed between the business semantic features corresponding to the anchor point and the business semantic features corresponding to each standard business node template to determine the semantic consistency between business actions, business objects, input and output data items, and rule constraints. From multiple standard business node templates, the standard business node template that is closest to the business semantics of the current anchor point is selected as the candidate business node. The mapping relationship between the anchor point and the candidate business nodes is established, and the anchor point business node mapping set is generated according to the order of dependencies of the anchor point in the survey anchor point chain.

[0012] Optionally, S3 specifically includes: S31: Extract the data item definition carried by each anchor point from the survey anchor point chain. The data item definition includes the data item name, data type, value range and business stage information. Match each data item to the corresponding business node in the clarification and revision business workflow according to the business stage. If multiple data items are attached to the same business node, a list of data items in the node is established. S32: Obtain the predefined financial attribute caliber rule base, traverse each business node with attached data items, retrieve the corresponding list of financial attribute fields to be filled from the rule base according to the business type of the current node, and check whether the data items attached to the corresponding business node have included each financial attribute field in the list. S33: For missing necessary financial attribute fields, generate a missing marker in the data item list corresponding to the business node. The missing marker includes the field name, the reason for the missing field, and the recommended source of supplementary data. For fields that already exist but whose value format or business definition does not meet the requirements of the rule base, generate a definition exception marker. S34: According to the execution order of the nodes in the clarification and revision business workflow, summarize each business node and data item list, missing marker and abnormal caliber marker into a phased business caliber list; the list is displayed in segments by node, and each node is marked with its integrity status.

[0013] Optionally, S32 specifically includes: Read the predefined financial attribute caliber rule library, which stores financial attribute field rules corresponding to different business types; parse the business node of each attached data item and extract the business type identifier of the current business node; Match the corresponding financial attribute field rules in the financial attribute caliber rule base according to the business type identifier; retrieve the list of financial attribute fields to be filled for the current business node from the matched financial attribute field rules; Read the list of data items attached to the current business node, and extract the data item name, field type, field format and business definition information from the list of data items; compare the list of data items with the list of financial attribute fields to be filled item by item, and determine whether there is a data item in the list of data items that corresponds to each financial attribute field in the list of financial attribute fields to be filled; When a corresponding data item exists, the consistency of the field type, field format, and business definition information of the corresponding data item is checked; when a corresponding data item does not exist, the corresponding financial attribute field is identified as a missing necessary financial attribute field.

[0014] Optionally, S3 also includes using the clarification and revision business workflow as the main framework of the report, embedding the corresponding phased business scope list content under each business node in the framework, and attaching overall integrity statistics, and finally generating a visual or exportable business research report.

[0015] The beneficial effects of this invention are: This invention achieves multimodal semantic fusion analysis of requirements by simultaneously introducing customer requirement documents and requirement clarification language communication content, constructing document requirement anchors, clarification Q&A anchors, and survey anchor chains. Compared to the traditional method of relying solely on manual compilation of business processes from requirement documents, this invention can synchronously incorporate spoken clarification information, supplementary explanations, and business revision information generated during the requirement survey process into the business analysis process. Through a semantic association mechanism, it automatically establishes the dependency relationship between requirement documents and Q&A clarification content, thereby reducing business omissions caused by incomplete requirement expression, misunderstandings of business, and differences in cross-departmental communication, and improving the completeness and accuracy of business requirement identification. Utilizing natural language processing technology, speech-to-text technology, and semantic triple extraction mechanisms, unstructured survey content is structurally transformed, enabling business requirements to form a survey anchor chain with sequential dependencies, enhancing the traceability of the requirement survey process and the continuity of business semantics.

[0016] This invention constructs a business component library and automatically generates original business workflows based on survey anchor chain. It then combines clarification question-and-answer anchors to confirm, supplement, and rewrite business nodes, achieving dynamic revision and automatic evolution of business workflows. Compared to traditional methods of manually drawing business process diagrams or manually maintaining requirement process documents, this invention can automatically adjust business node definitions, input / output data items, and node processing logic based on new business rules, business logic conflicts, and data item changes during the requirement clarification process. This makes the generated clarification and revision business workflows more closely aligned with real business scenarios. Through a three-tiered node revision mechanism of confirmation, supplementation, and rewriting, the business workflow not only retains the original requirement structure but also dynamically absorbs revisions from subsequent requirement communications. This solves the problems of scattered requirement versions, delayed process updates, and poor business logic consistency in traditional business surveys, improving business process construction efficiency and consistency.

[0017] This invention establishes a rule base for financial attribute definitions and links data item definitions to corresponding business nodes in the clarification and revision workflow. This enables automatic verification of business definition completeness and automatic generation of a phased business definition list. Compared to traditional business research reports that rely primarily on manual verification of field completeness, this invention can automatically retrieve the corresponding list of financial attribute fields to be filled based on the business type of different business nodes. It performs field completeness verification, field format verification, and business definition consistency verification on the data items in the business nodes, and automatically generates missing data markers and definition anomaly markers, proactively identifying data missing issues and definition conflicts during the business research process. Furthermore, using the clarification and revision workflow as the main framework of the business research report, it embeds the phased business definition list into the corresponding business nodes and generates overall completeness statistics. This ensures that the final output business research report not only has the ability to display business processes but also possesses data governance, financial definition verification, and business risk identification capabilities, improving the standardization, data reliability, and implementation efficiency of subsequent system construction. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only for this invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This is a schematic diagram illustrating the process revision of an embodiment of the present invention. Detailed Implementation

[0020] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. For some well-known technologies, those skilled in the art may also use other alternative methods to implement the invention. Moreover, the accompanying drawings are only for more specific description of the embodiments and are not intended to specifically limit the present invention.

[0021] like Figures 1-2 As shown, a method for automatically generating business research reports based on multimodal semantic fusion includes the following steps: S1: Obtain customer requirement documents and requirement clarification language communication content, parse the customer requirement documents to obtain document requirement anchors, parse the requirement clarification language communication content to obtain clarification Q&A anchors, associate the document requirement anchors with the clarification Q&A anchors to generate a survey anchor chain; S1 specifically includes: S11: Through a pre-defined data interface, retrieve customer requirement documents from the customer requirement management system and obtain requirement clarification language communication content, including Q&A formats, from the requirement communication record library. The customer requirement documents include business requirement specifications, functional requirement lists, prototype description documents, and project scope description documents; the requirement clarification language communication content includes meeting recordings, online communication records, email Q&A records, and instant messaging Q&A records. The data interface is implemented using an application programming interface (API), database connection interface, or file exchange interface. After obtaining the requirement clarification language communication content, perform uniform formatting processing on the content, converting text content from different sources into unified structured text data, and generating a requirement clarification corpus sequence in chronological order, providing input data for the subsequent extraction of document requirement anchors and clarification Q&A anchors.

[0022] S12: Utilize natural language processing technology to perform semantic analysis on customer requirement documents, extract several concept nodes that represent business objectives, functional scope, data entities, and constraints, and mark each concept node as a document requirement anchor point.

[0023] Specifically, the system performs word segmentation, part-of-speech tagging, named entity recognition, and dependency parsing on customer requirement documents to identify business target entities, functional action entities, data object entities, and rule constraint entities from the text, and constructs a set of concept nodes. ; In section S12, after obtaining the customer requirements document, the document content is first preprocessed, including removing invalid symbols, standardizing terminology, splitting paragraphs, and segmenting sentences to form a standardized set of statements. Each statement is then segmented into semantic units. Part-of-speech tagging (POT) is used to identify the part-of-speech category of each unit, distinguishing between action words, business object words, attribute words, and restrictive condition words. After POT, named entity recognition (MID) is performed on the statements to identify business target entities, functional action entities, data object entities, and rule constraint entities. For example, in "generating financial statements," "financial statements" can be identified as a data object entity, and "generating" can be identified as a functional action entity; in "approval is required before export," "approval" can be identified as a business process constraint entity, and "export" can be identified as a functional action entity. Dependency parsing is then performed on the identified entities to analyze subject-predicate relationships, action relationships, modification relationships, and conditional constraint relationships between different entities to determine the association structure between business semantics. For example, in "system-generated daily reports," the system can be identified as the executing subject, "generation" as a business action, and the daily report as a data object, forming a complete business semantic unit. Furthermore, semantic classification is performed on each business semantic unit, unifying statements with the same business meaning but different expressions. For example, exporting reports, outputting report files, and generating report download files can be uniformly classified as the report export function. After semantic classification, each independent business semantic unit is constructed into a concept node. Concept nodes primarily represent the core business semantic content in the customer requirement document, and each concept node includes at least information such as business type, business action, data object, constraint rules, and the business stage it belongs to. A unique identifier is assigned to each concept node, and each concept node is marked as a document requirement anchor point, enabling document requirement anchor points to correspond to explicit business semantic units in customer requirements. This provides a unified business semantic foundation for subsequent semantic association analysis between document requirement anchor points and clarification question-and-answer anchor points.

[0024] Natural Language Processing (NLP) technology refers to a class of intelligent text analysis technologies used to automatically understand, semantically decompose, and extract business information from the natural language content of customer requirement documents. NLP technology can include word segmentation, part-of-speech tagging, named entity recognition, keyword extraction, dependency parsing, semantic role analysis, and contextual semantic vector encoding. Its role is to transform the business meaning in the original requirement text into structured business semantic nodes, thereby providing a foundation for subsequently generating document requirement anchors. This represents the set of document requirement anchor points. Indicates the first Anchor points for document requirements This indicates the total number of document requirement anchor points.

[0025] Furthermore, a corresponding semantic feature vector is generated for each document's requirement anchor point: ; in, Indicates the first The semantic feature vector of each document's requirement anchor point Represents the first semantic feature vector. Dimensional semantic feature values.

[0026] Semantic feature vectors are used to characterize the business semantic content of the corresponding concept nodes, so as to facilitate subsequent semantic association analysis between document requirement anchors and clarification question and answer anchors.

[0027] S13: Perform speech-to-text processing and dialogue structure analysis on the language communication content for clarifying needs, identify each question and its corresponding answer, extract the core semantic triples from each question-answer pair, namely subject-predicate-object, and mark each semantic triple as a clarification question-answer anchor point.

[0028] S131, to clarify the content of language communication regarding the need for speech format, firstly, a speech recognition model is used to convert speech data into text data; then, speaker separation, question-and-answer matching, and contextual analysis are performed on the converted text data to obtain a set of question-and-answer pairs: ; in, Represents a set of question-answer pairs. Indicates the first The question and answer were correct.

[0029] The speech recognition model in S131 adopts an end-to-end speech-to-text model structure consisting of an acoustic feature encoding layer, a contextual semantic modeling layer, a character decoding layer, and a text post-processing layer. The model input is speech data from the content of the clarification of needs, and the model output is a text transcription result with a timestamp. The text transcription result is used for subsequent speaker separation, question-and-answer matching, and contextual association analysis.

[0030] The model structure specifically includes: First, a speech preprocessing module, used for noise reduction, silence clipping, volume normalization, and frame segmentation of the raw speech data, ensuring a unified input format for speech data collected from different conference devices, call systems, or recording environments; Second, an acoustic feature extraction module, used to extract Mel-spectrum features or filter bank features from each frame of speech, converting continuous speech signals into acoustic feature sequences that the model can recognize; Third, an acoustic feature encoding module, used to encode the acoustic feature sequences using convolutional networks or Transformer encoders, extracting phonemes, syllables, pauses, stresses, and continuous pronunciation features from the speech; Fourth, a contextual semantic modeling module, used to determine the true meaning of homophones, business terms, system names, process names, and data item names in the current context by combining the contextual relationships between preceding and following speech segments; Fifth, a character or word decoding module, used to progressively decode the encoded speech features into a text sequence; Sixth, a text post-processing module, used to perform punctuation restoration, number normalization, proper noun correction, business term replacement, and timestamp alignment on the initial text sequence, ultimately generating structured text data.

[0031] During model construction, training samples were first collected, including recordings of requirements gathering meetings, customer communication recordings, project interview recordings, and their manually transcribed texts. These training samples were then cleaned to remove severe noise, invalid blank segments, and duplicate fragments. Next, each audio segment was time-aligned with its corresponding manually transcribed text, forming audio-text training sample pairs. A business terminology lexicon was established, including customer system names, business process names, financial account names, data table names, field names, functional module names, and commonly used requirement expressions, to enhance the model's ability to recognize business-specific vocabulary. Then, the audio segments were input into the model, which predicted the corresponding text content. The predicted text was compared with the manually transcribed text, and the model parameters of the acoustic feature encoding layer, contextual semantic modeling layer, and character decoding layer were adjusted based on recognition errors until the model could stably output recognition results consistent with or close to the manually transcribed text. Finally, the trained speech recognition model, along with the business terminology lexicon, punctuation recovery rules, and text correction rules, were encapsulated to form a speech-to-text model for transcribing requirements clarification communication. During model execution, the system first receives raw speech data from the clarification of requirements. This raw speech data undergoes sampling rate unification, channel merging, background noise suppression, and silence detection to obtain continuous, valid speech segments. Then, each valid speech segment is framed and windowed, breaking it down into multiple short-time speech frames. Acoustic features are extracted from each short-time speech frame to form an acoustic feature sequence. Next, the acoustic feature sequence is input into the acoustic feature encoding module, which identifies the articulatory units and their continuous changes in the speech, generating a high-level speech representation. The contextual semantic modeling module combines the semantic content before and after the current speech segment, the business terminology vocabulary, and the requirements communication scenario information to semantically enhance the high-level speech representation, reducing ambiguity caused by homophones, abbreviations, technical terms, and colloquial expressions. After semantic enhancement, the character or word decoding module generates candidate text sequences step-by-step according to the speech pronunciation order. Multiple candidate text sequences are ranked by confidence, and the text sequence with the highest confidence and conforming to the business terminology vocabulary and contextual semantic relationships is selected as the initial transcribed text. Subsequently, the text post-processing module performs punctuation completion, sentence segmentation, number and unit standardization, business terminology correction, and removal of redundant colloquialisms on the initial transcribed text. For example, the requirement for the system to be able to query and export customer balances is standardized to ensure the system supports both customer balance querying and exporting. Finally, the processed text content is bound to the start and end times of the corresponding audio segments, outputting a timestamped text transcription result. This text transcription result serves as input for subsequent dialogue structure analysis, used to further identify the correspondence between question content, answer content, and question-answer pairs.

[0032] S132, perform semantic relation extraction on each question-answer pair to obtain the corresponding semantic triples: ; in, Indicates the first A semantic triple, Indicates the main semantic item, Indicates the semantic term of the predicate; Represents the semantic term of the object.

[0033] Each semantic triple is labeled as a clarifying question-and-answer anchor, and a set of clarifying question-and-answer anchors is constructed: ; in, This represents the set of anchor points for clarifying questions and answers. Indicates the first A clarification question and answer anchor point.

[0034] Each clarification question and answer anchor includes business semantic relationship information for the corresponding question and answer pair, which is used to represent the business confirmation content, rule supplementary content, and constraint revision content generated during the requirement clarification process.

[0035] S14: Calculate the semantic similarity between each document requirement anchor and each clarification question and answer anchor, establish a relationship between document pairs with similarity exceeding a preset threshold, and arrange and combine the anchors in the relationship according to the time sequence of requirement clarification or the business logic sequence to form a research anchor chain with a sequential dependency relationship.

[0036] Semantic feature vectors for document requirement anchors and clarification question-and-answer anchors are obtained separately, and cosine similarity is used to calculate the semantic similarity between them: ; in, Indicates the first The document's requirement anchor point and the first Semantic similarity between clarification question and answer anchors Indicates the first The semantic feature vector of each document's requirement anchor point Indicates the first The semantic feature vector of a clarifying question-and-answer anchor point This represents the vector dot product operation. The magnitude of the semantic feature vector representing the document requirement anchor point. This represents the magnitude of the semantic feature vector of the clarification question-and-answer anchor point.

[0037] When the following conditions are met: At the same time, establish the relationship between document requirement anchors and clarification question and answer anchors; in, The semantic similarity threshold is used to determine whether document requirement anchors and clarification question-and-answer anchors belong to the same business semantic content. The semantic similarity threshold can be set according to the complexity of the requirement document, the standardization of business terminology, and the colloquialism of the requirement clarification statements. Typically, a value range of 0.70–0.90 is acceptable. When the business terminology in the customer requirement document is relatively uniform, the expression is standardized, and the requirement boundaries are clear, the semantic similarity threshold can be appropriately increased, for example, set to 0.82–0.90, to reduce the possibility of mis-associating different business semantics and improve the accuracy of the research anchor chain. When the requirement clarification language contains many colloquial expressions, abbreviations, omissions, or significant differences in the expression habits of business personnel, the semantic similarity threshold can be appropriately decreased, for example, set to 0.70–0.80, to improve the recall capability of semantic associations and avoid the inability to establish a connection between anchors that should be associated due to differences in expression. Furthermore, in actual business research scenarios, the semantic similarity threshold is set to 0.78–0.85. When the similarity is below 0.78, anchor points with only partial keyword overlap but different business meanings are easily misassociated. When the similarity is above 0.85, the ability to identify and associate colloquial clarifications, supplementary explanations, and implicit business expressions will significantly decrease. Setting the semantic similarity threshold between 0.78 and 0.85 can achieve a good balance between semantic association accuracy and business semantic recall.

[0038] An anchor point association set is generated based on the relationships. Then, taking into account the time of requirement clarification, the order of business processing, and business dependencies, the associated anchor points in the anchor point association set are sorted and chained together to generate a survey anchor point chain. ; in, Indicates the anchor chain of the survey. This node represents a combination of document requirement anchors and clarification question / answer anchors that establish a relationship.

[0039] The survey anchor chain is used to characterize the business semantic evolution relationship in the process of customer demand formation, clarification and revision, and provides basic data support for the subsequent generation of clarification and revision business workflows.

[0040] S2: Generate the original business workflow based on the survey anchor chain, and use the clarification question and answer anchors to confirm, supplement and rewrite the business nodes in the original business workflow to generate the clarification and revision business workflow; S2 specifically includes: S21: Map each anchor point to a candidate business node in the preset business group library according to the order of dependencies between anchor points in the survey anchor point chain.

[0041] S211: First, read the survey anchor chain generated by S14, and obtain the dependencies between each anchor according to the order of the anchors in the survey anchor chain. The business group library stores multiple standard business node templates. Each standard business node template corresponds to a standard business processing function, and each standard business node template includes a node name, node function description, input data items, output data items, node rule constraints, and node connection rules.

[0042] S212: Extract the corresponding business semantic features for each anchor point in the survey anchor point chain, and perform semantic matching with the standard business node templates in the business group library. Calculate the node matching degree between the anchor point and each standard business node template. ; S212 first reads the current anchor point in the survey anchor point chain and parses the corresponding business semantic content, such as business actions, business objects, input data items, output data items, and business rules. This business semantic content is then converted into unified business semantic features. For example, customer balance inquiry can be broken down into three core business semantic features: customer, balance, and inquiry. The system then reads standard business node templates from the business component library and extracts the node function description, input / output data items, business processing actions, and rule constraints from each template, also converting these into corresponding business semantic features. After extracting the semantic features from both sides, the system compares the business semantic features of the current anchor point with those of each standard business node template item by item to determine the similarity in business actions, business objects, data item names, and rule semantics. If most of the business semantic features of the current anchor point can be found in a corresponding standard business node template, it indicates a high degree of node matching between the current anchor point and that standard business node template. For example, if the current anchor point includes business semantics such as customer information query, query conditions, and result list output, and the data query node template in the business component library also includes business semantics such as information query, condition filtering, and result output, then the system will determine that the node matching degree between the two is high, and will use this data query node template as a candidate business node corresponding to the current anchor point. Finally, the system selects the standard business node template with the highest node matching degree and that meets the preset matching conditions from all standard business node templates, as the final candidate business node corresponding to the current anchor point, and establishes a mapping relationship between the anchor point and the candidate business nodes, providing a business node foundation for the subsequent automatic generation of the original business workflow. Indicates the first The anchor point and the first Node matching degree between standard business node templates Indicates the first Anchor points, Indicates the first A standard business node template, Indicates the first The business semantic feature vector corresponding to each anchor point Indicates the first The business semantic feature vector corresponding to each standard business node template; S213: Select the standard business node template with the highest node matching degree that exceeds the preset node matching threshold as the candidate business node corresponding to the current anchor point, and establish a mapping relationship between the anchor point and the candidate business nodes to generate an anchor point business node mapping set. The preset node matching threshold is used to determine whether there is sufficient business semantic consistency between the current anchor point and the standard business node template. The preset node matching threshold is set to 0.75–0.90. When the standard business node templates in the business component library are finely divided and have high business semantic differentiation, the preset node matching threshold can be appropriately increased, for example, set to 0.85–0.90, to reduce mismatches between different business nodes. When the customer requirement document contains many colloquial expressions, abbreviations, or incomplete business descriptions, the preset node matching threshold can be appropriately decreased, set to 0.80–0.85. This is because when the node matching degree is below 0.80, nodes with similar business actions but different business responsibilities are easily mismapped; when the node matching degree is above 0.85, some business nodes with different expressions but consistent business logic may not be successfully matched. Therefore, setting the preset node matching threshold to 0.80–0.85 can balance the accuracy and coverage of business node matching.

[0043] S22: Based on the mapping results, starting from the business node corresponding to the starting anchor point and ending from the business node corresponding to the ending anchor point, the original business workflow is automatically generated according to the association direction between anchor points and the connection rules in the business component library. The original business workflow includes several business nodes and the flow path between nodes.

[0044] Identify the starting and ending anchors in the survey anchor chain from the anchor business node mapping set, and obtain the corresponding business nodes as the starting and ending nodes of the business workflow. Based on the anchor association direction in the survey anchor chain, determine the data flow order and business execution order between business nodes. Simultaneously, read the node connection rules from the business component library to determine whether the output data item of the current business node meets the input data item requirements of the next business node. When the connection conditions are met, establish node connection relationships between the corresponding business nodes to form a business node flow path. Construct the original business workflow from all business nodes and node connection relationships. ; in, This represents the original business workflow. Represents a set of business nodes. It represents the set of flow paths between nodes; the flow path is used to characterize the data flow relationship, business execution order relationship and business dependency relationship between business nodes, forming a complete original business processing flow structure.

[0045] S23: Traverse each business node in the original business workflow and retrieve the clarification question and answer anchor points associated with that node; if an associated clarification question and answer anchor point exists, perform confirmation, supplementation, or rewriting operations based on the question and answer content carried by that anchor point.

[0046] For each business node in the original business workflow, extract the node semantic identifier and retrieve the corresponding clarification Q&A anchor point in the survey anchor point chain based on the node semantic identifier. Perform semantic consistency analysis and difference analysis on the Q&A content in the clarification Q&A anchor point. When the Q&A content is consistent with the business semantics of the current business node and no new data items, new rule constraints, or new processing logic errors are detected, the current business node is marked as confirmed, indicating that the current business node has passed requirement clarification verification. When the Q&A content contains new data items, new rule constraints, or new business conditions that are not included in the current business node but fall within the scope of the current business node's responsibilities, perform supplementary processing on the current business node, adding corresponding data fields, rule constraints, or condition judgment logic to the current business node to expand the business processing scope of the current business node. When the Q&A content conflicts with the business semantics of the current business node, or when the Q&A content explicitly proposes modifying the node definition, adjusting the processing logic, changing input and output data items, or changing the business function, perform rewriting processing on the current business node, updating the business name, input data items, output data items, processing logic, and node rule constraints of the current business node according to the Q&A content. Furthermore, the semantic differences before and after the revision of business nodes are recorded, generating a set of node revision records: ; in, Represents the set of node revision records. Indicates the first Node revision record.

[0047] Each node revision record is used to record the exact results, supplementary content, or rewritten content corresponding to the business node.

[0048] S24: Reconnect the confirmed, supplemented or rewritten business nodes according to the original flow path to generate a clarification and revision business workflow.

[0049] The revised business nodes are re-verified to check the matching relationships of input and output data items and the constraints of node rules. When the data flow relationship between adjacent business nodes meets the connection conditions, the original node flow path is retained. When the input and output data items change due to the supplementation or rewriting of business nodes, the flow path between the corresponding nodes is re-adapted, and the node connection relationship is updated. All revised business nodes and updated node flow paths are recombined to generate a clarified revision business workflow. ; The clarification and revision business workflow is used to characterize the complete business process after requirements clarification, business confirmation, and business revision, providing a business process foundation for the subsequent generation of phased business scope lists and business research reports.

[0050] in, This indicates a clarification and revision of the business workflow. This represents the revised set of business nodes. This represents the revised set of node flow paths.

[0051] S3: Connect the data items in the survey anchor chain to the business nodes corresponding to the clarification and revision business workflow, perform a caliber integrity check based on the financial attributes of the data items, generate a phased business caliber list with missing markers, and generate a business survey report based on the clarification and revision business workflow and the phased business caliber list.

[0052] S3 specifically includes: S31: Extract the data item definitions carried by each anchor point from the survey anchor point chain, and attach the data items to the corresponding business nodes in the clarification and revision business workflow.

[0053] Read each anchor point in the survey anchor point chain and parse the data item definition information corresponding to each anchor point. The data item definition should include at least the data item name, data type, value range, and the business stage information to which it belongs. Match the data item's business stage information with the business node stage identifier in the clarification and revision business workflow to determine the target business node corresponding to the data item. When multiple data items match the same business node, create a list of data items within the node according to the data item name and data type. Construct the association relationship between business nodes and data items. ; in, This represents the set of connections between business nodes and data items. Indicates the first Each business node Indicates the first One data item.

[0054] Furthermore, a corresponding list of data items is generated for each business node: ; The data item list is used to represent the data content and business scope information involved in the current business node.

[0055] in, Indicates the first A list of data items corresponding to each business node. Represents the first item in the list of data items. One data item; S32: Obtain the predefined financial attribute caliber rule base and perform financial attribute integrity verification according to the business type of the business node.

[0056] The system reads the financial attribute definition rule base, which pre-stores financial attribute field rules corresponding to multiple business types. Each financial attribute field rule includes a field name, field type, field format requirements, field value rules, and field integrity verification conditions. It then iterates through each business node with attached data items, extracts the business type identifier corresponding to the current business node, and retrieves the corresponding list of required financial attribute fields from the financial attribute definition rule base based on the business type identifier. ; Furthermore, the data item list of the current business node is compared item by item with the list of financial attribute fields to be filled in. The system checks whether the corresponding financial attribute fields exist in the data item list and whether the corresponding fields meet the field format requirements and business definition requirements, thus completing the financial attribute integrity verification of the business node.

[0057] in, Indicates the first List of financial attribute fields to be filled in for each business node Indicates the first The financial attribute rule base is a collection of rules used to uniformly manage the standards, business scope definitions, and integrity verification rules for various financial data fields. It is primarily used to perform financial attribute integrity verification, field standardization verification, and business scope consistency verification on data items linked to business nodes, ensuring that financial data across different business stages meets unified data governance requirements. The rule base can pre-store financial attribute field rules corresponding to multiple business types. For example, revenue-related business nodes can correspond to financial attribute fields such as revenue recognition date, revenue amount, tax rate, and customer code; cost-related business nodes can correspond to financial attribute fields such as cost account, expense department, and budget number; and settlement-related business nodes can correspond to financial attribute fields such as payment method, settlement cycle, and invoice type. Each financial attribute field rule can include field name, field type, field format requirements, field value range, field coding rules, mandatory field conditions, and the relationships and constraints between fields. Furthermore, the rule base can also store business scope definitions for different business scenarios. These business definition criteria can avoid the problem of different departments, systems, or researchers using different statistical standards for the same financial field. In actual implementation, the system can automatically retrieve the corresponding financial attribute field rules from the financial attribute caliber rule base according to the business type of the current business node, and compare the data items attached to the business node with the standard fields in the rule base item by item to identify problems such as missing fields, abnormal field formats, and inconsistent business criteria.

[0058] S33: For missing necessary financial attribute fields, generate a missing flag in the data item list of the corresponding business node; for fields that exist but do not meet the rule requirements, generate an abnormal scope flag.

[0059] When a financial attribute field from the list of required financial attribute fields is missing from the data item list, a corresponding missing flag is generated in the current business node. The missing flag includes at least the field name, the reason for the missing field, and the recommended data source for supplementation. The reason for the missing field indicates the business reason for its absence; the recommended data source indicates the data system, business form, or business department that can be used to supplement the field. When a corresponding financial attribute field already exists in the data item list, but the field value format, field encoding method, field value range, or business definition does not meet the requirements of the financial attribute definition rule base, a definition exception flag is generated in the current business node.

[0060] The caliber anomaly marker is used to characterize inconsistencies between field content and standard financial caliber. Subsequently, a set of node markers is constructed: ; in, Indicates the first The set of node tags corresponding to each business node Indicates the first Missing markers or caliber aberration markers.

[0061] The node tag set is used to record the missing business definitions and abnormal business definitions in the current business node.

[0062] S34: Following the execution order of the clarification and revision workflow nodes, summarize the business nodes and their corresponding data item lists, missing item markers, and caliber exception markers into a phased business caliber list. Specifically, according to the node flow order in the clarification and revision workflow, read each business node sequentially and obtain the data item list and node marker set corresponding to the current business node.

[0063] The data item list, missing data markers, and data caliber exception markers are segmented and organized using business nodes as units to generate a phased business caliber list: ; in, This indicates a phased list of business procedures. Indicates the first The phased business definitions corresponding to each business node.

[0064] Calculate the integrity status for each business node. If the current business node has no missing or abnormal data markers, it is marked as complete; if it has missing or abnormal data markers, it is marked as needing supplementation or abnormal. A phased business data specification list is used to display the completeness and any abnormalities of business data specifications according to the business execution stage.

[0065] S35: Use the clarification and revision business workflow as the main framework of the business research report. Embed the corresponding phased business scope list under each business node in the main framework, and attach overall completeness statistics to finally generate the business research report.

[0066] Read the node structure and node flow relationship in the clarification and revision business workflow, and construct the chapter structure of the business survey report according to the execution order of business nodes; embed the corresponding phased business scope list content in the report chapter corresponding to each business node, including data item list, missing mark, scope exception mark and node integrity status.

[0067] Perform statistical analysis on the integrity status of all business nodes to generate overall integrity statistics: ; in, Indicates the proportion of overall integrity. This indicates the number of business nodes in complete state. This indicates the total number of business nodes.

[0068] Then, the overall integrity statistics are added to the summary area of ​​the business research report, and a visual report file or an export report file is generated according to the output requirements of the business research report.

[0069] The business research report is used to show the customer's business process structure, business data scope, missing data items, and abnormal business rules, providing a research basis for subsequent system construction, data governance, and financial business analysis.

[0070] This invention encompasses any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of this invention. To provide the public with a thorough understanding of this invention, specific details are described in detail in the following preferred embodiments; however, those skilled in the art will fully understand the invention even without these details. Furthermore, to avoid unnecessary misunderstanding of the essence of this invention, well-known methods, processes, procedures, components, and circuits are not described in detail.

[0071] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for automatically generating business research reports based on multimodal semantic fusion, characterized in that, Includes the following steps: S1: Obtain customer requirement documents and requirement clarification language communication content, parse the customer requirement documents to obtain document requirement anchors, parse the requirement clarification language communication content to obtain clarification Q&A anchors, associate the document requirement anchors with the clarification Q&A anchors to generate a survey anchor chain; S2: Generate the original business workflow based on the survey anchor chain, and use the clarification question and answer anchors to confirm, supplement and rewrite the business nodes in the original business workflow to generate the clarification and revision business workflow; S3: Connect the data items in the survey anchor chain to the business nodes corresponding to the clarification and revision business workflow, perform a caliber integrity check based on the financial attributes of the data items, generate a phased business caliber list with missing markers, and generate a business survey report based on the clarification and revision business workflow and the phased business caliber list. S2 specifically includes: S21: According to the order of dependencies between anchors in the survey anchor chain, map each anchor to a candidate business node in the preset business component library; the business component library stores standard business node templates and connection rules between templates; S22: Based on the mapping results, taking the business node corresponding to the starting anchor point as the starting point and the business node corresponding to the ending anchor point as the ending point, the original business workflow is automatically arranged and generated according to the association direction between anchor points and the connection rules in the business component library. The original business workflow includes several business nodes and the flow path between nodes. S23: Traverse each business node in the original business workflow and retrieve the clarification question and answer anchor points associated with the corresponding business node; if there are associated clarification question and answer anchor points, perform confirmation, supplementation and rewriting operations based on the question and answer content carried by the clarification question and answer anchor points. S24: Reconnect the confirmed, supplemented or rewritten business nodes according to the original flow path to generate a clarification and revision business workflow.

2. The method for automatically generating business research reports based on multimodal semantic fusion according to claim 1, characterized in that, S1 specifically includes: S11: Read customer requirement documents from the customer requirement management system through the preset data interface, and obtain requirement clarification language communication content in the form of questions and answers from the requirement communication record library; S12: Use natural language processing technology to perform semantic parsing on the customer requirement document, extract the concept nodes that represent business objectives, functional scope, data entities and constraints, and mark each concept node as a document requirement anchor point; S13: Perform speech-to-text processing and dialogue structure analysis on the language communication content of the request clarification, identify each question and its corresponding answer, extract core semantic triples from each question-answer pair, and mark each semantic triple as a clarification question-answer anchor point; S14: Calculate the semantic similarity between each document requirement anchor point and each clarification question and answer anchor point, establish an association relationship for anchor point pairs with similarity exceeding a preset threshold, and arrange and combine the anchor point pairs in the association relationship according to the time order of requirement clarification or the business logic order to form a survey anchor point chain with a sequential dependency relationship.

3. The method for automatically generating business research reports based on multimodal semantic fusion according to claim 2, characterized in that, S12 specifically includes: The customer requirement document is preprocessed with text, including invalid symbol removal, terminology standardization, paragraph splitting, and sentence segmentation, to generate a standardized set of sentences. The standardized set of sentences is then segmented and part-of-speech tagging is performed to identify action words, business object words, attribute words, and constraint words in the sentences. Named entity recognition is then performed on the tagged sentences to extract business target entities, functional action entities, data object entities, and rule constraint entities. Dependency parsing is performed on the extracted entities to identify subject-predicate, action, modification, and conditional constraint relationships between entities, generating business semantic units. Semantic classification processing is performed on the business semantic units to uniformly map business semantic units with consistent semantics or the same business meaning. Concept nodes are constructed based on the classified business semantic units, and a unique identifier is assigned to each concept node. Each concept node is marked as a corresponding document requirement anchor point.

4. The method for automatically generating business research reports based on multimodal semantic fusion according to claim 2, characterized in that, S13 specifically includes: The requirements for speech format are clarified, and noise reduction, silence segment clipping, sampling rate unification, and speech segment segmentation are performed to obtain effective speech segments. The effective speech segments are input into the speech recognition model. Through acoustic feature extraction, acoustic feature encoding, contextual semantic modeling, character decoding, and text post-processing, the effective speech segments are converted into text transcription results with timestamps. Speaker separation and dialogue turn segmentation are performed on the text transcription results to identify the speech content of the customer side, the speech content of the survey side, and the corresponding speech time order. Based on the interrogative sentence structure, inquiry keywords, confirmation keywords, answer connectors, and the contextual semantic relationship between adjacent dialogue rounds, question-and-answer matching is performed on the text transcription results to obtain question-and-answer pairs consisting of the question content and the corresponding answer content. Semantic relation extraction is performed on each question-answer pair to identify the subject semantic item, predicate semantic item, and object semantic item in the question-answer pair and generate core semantic triples. Each core semantic triple is marked as a clarifying question-answer anchor point, and each clarifying question-answer anchor point is bound to the timestamp, speaker identifier, and business semantic relation information of the corresponding question-answer pair.

5. The method for automatically generating business research reports based on multimodal semantic fusion according to claim 1, characterized in that, The execution confirmation, supplementation, and rewriting operations specifically include: Confirmation: When the question and answer content is consistent with the business semantics of the current business node and there is no new information, the corresponding node will be marked as confirmed. Additional information: When the Q&A content mentions new data items or business rules that are not covered by the current node but fall within the scope of the node's responsibilities, add the corresponding data fields or rule constraints to the node; Rewrite: When the question and answer content conflicts with the business semantics of the current node or explicitly requires modification of the node definition, update the node's business name, input and output data items, or processing logic based on the question and answer content.

6. The method for automatically generating business research reports based on multimodal semantic fusion according to claim 1, characterized in that, S21 specifically includes: Read the survey anchor chain and obtain the sequential dependencies between each anchor according to the order in the survey anchor chain; Read multiple standard business node templates from the business component library. The standard business node template includes a node name, node function description, input data items, output data items, node rule constraints, and node connection rules. Extract the business semantic features corresponding to each anchor point and extract the business semantic features corresponding to each standard business node template. Semantic matching is performed between the business semantic features corresponding to the anchor point and the business semantic features corresponding to each standard business node template to determine the semantic consistency between business actions, business objects, input and output data items, and rule constraints. From multiple standard business node templates, the standard business node template that is closest to the business semantics of the current anchor point is selected as the candidate business node. The mapping relationship between the anchor point and the candidate business nodes is established, and the anchor point business node mapping set is generated according to the order of dependencies of the anchor point in the survey anchor point chain.

7. The method for automatically generating business research reports based on multimodal semantic fusion according to claim 1, characterized in that, S3 specifically includes: S31: Extract the data item definition carried by each anchor point from the survey anchor point chain. The data item definition includes the data item name, data type, value range and business stage information. Match each data item to the corresponding business node in the clarification and revision business workflow according to the business stage. If multiple data items are attached to the same business node, a list of data items in the node is established. S32: Obtain the predefined financial attribute caliber rule base, traverse each business node with attached data items, retrieve the corresponding list of financial attribute fields to be filled from the rule base according to the business type of the current node, and check whether the data items attached to the corresponding business node have included each financial attribute field in the list. S33: For missing necessary financial attribute fields, generate a missing marker in the data item list corresponding to the business node. The missing marker includes the field name, the reason for the missing field, and the recommended source of supplementary data. For fields that already exist but whose value format or business definition does not meet the requirements of the rule base, generate a definition exception marker. S34: According to the execution order of the nodes in the clarification and revision business workflow, summarize each business node and data item list, missing marker and abnormal caliber marker into a phased business caliber list; the list is displayed in segments by node, and each node is marked with its integrity status.

8. The method for automatically generating business research reports based on multimodal semantic fusion according to claim 7, characterized in that, Specifically, S32 includes: Read the predefined financial attribute caliber rule library, which stores financial attribute field rules corresponding to different business types; parse the business node of each attached data item and extract the business type identifier of the current business node; Match the corresponding financial attribute field rules in the financial attribute caliber rule base according to the business type identifier; retrieve the list of financial attribute fields to be filled for the current business node from the matched financial attribute field rules; Read the list of data items attached to the current business node, and extract the data item name, field type, field format and business definition information from the list of data items; compare the list of data items with the list of financial attribute fields to be filled item by item, and determine whether there is a data item in the list of data items that corresponds to each financial attribute field in the list of financial attribute fields to be filled; When a corresponding data item exists, the consistency of the field type, field format, and business definition information of the corresponding data item is checked; when a corresponding data item does not exist, the corresponding financial attribute field is identified as a missing necessary financial attribute field.

9. The method for automatically generating business research reports based on multimodal semantic fusion according to claim 1, characterized in that, The S3 also includes using the clarification and revision business workflow as the main framework of the report, embedding the corresponding phased business scope list content under each business node in the framework, and attaching overall integrity statistics, and finally generating a visual or exportable business research report.

Images