AI-based business form filling method, device, and storage medium

Abstract

The application provides an AI-based business form filling method and device and a storage medium. The method comprises: obtaining business source data required for business form filling, determining the modal type of the business source data, the business source data comprising voice data and / or document data; applying an information extraction strategy corresponding to the modal type to convert the business source data into unstructured text data; performing deep semantic analysis on the unstructured text data using a semantic analysis model to obtain structured data conforming to the business form field value standard; determining the target business form corresponding to the intent label based on an intent label and business form associated knowledge base; and mapping the business entity data to the corresponding field in the target business form to obtain the filled business form. The application realizes end-to-end processing from multi-source input to automatic form filling by fusing multi-modal information extraction and semantic analysis technology, thereby improving the filling efficiency, accuracy and automation level.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and automated form processing technology, and in particular to an AI-based business form filling method, device and storage medium. Background Technology

[0002] In government affairs, finance, and corporate administration, business form completion is a core part of daily operations. With the deepening of digital transformation, the types of business forms are becoming increasingly complex (such as tax returns, loan approval forms, travel expense reports, and personnel file forms), and the data sources involved are becoming increasingly diversified, including paper documents, electronic images, voice commands, and handwritten records. How to efficiently and accurately extract information from multi-source heterogeneous data and complete form completion has become a pressing technical problem in this field.

[0003] In related technologies, business form filling mainly relies on manual operation. Specifically, users enter business data into the corresponding fields of the form one by one through keyboard input or handwriting. After completion, the field format is manually checked item by item. After completing the above operations, the business form is considered to be filled out and can be submitted to the next stage.

[0004] However, the above-mentioned reporting methods suffer from low efficiency, low accuracy, and low level of automation. Summary of the Invention

[0005] This application provides an AI-based business form filling method, device, and storage medium to improve the problems of low filling efficiency, low accuracy, and low automation level in related technologies.

[0006] Firstly, this application provides an AI-based method for filling out business forms, including:

[0007] Obtain the business source data required for filling out business forms, and determine the modal type of the business source data, which includes voice data and / or document data;

[0008] Apply information extraction strategies corresponding to the modality type to convert business source data into unstructured text data;

[0009] A semantic parsing model is used to perform deep semantic parsing on unstructured text data to obtain structured data that conforms to the business form field value standards. The structured data includes business entity data and intent tags.

[0010] Based on the knowledge base that associates intent tags with business forms, determine the target business form corresponding to the intent tag;

[0011] The business entity data is mapped to the corresponding fields in the target business form to obtain the filled-in business form.

[0012] In one possible implementation, a semantic parsing model is used to perform deep semantic parsing on unstructured text data to obtain structured data that conforms to the value standards of business form fields, including:

[0013] Unstructured text data is input into a semantic parsing model, which performs entity extraction and intent classification to identify business entity data and intent labels in the unstructured text data. The semantic parsing model is a natural language understanding model that has been fine-tuned for the target business domain.

[0014] The business entity data is normalized and transformed to obtain normalized business entity data. The normalization and transformation includes: semantic normalization processing to convert colloquial expressions into standard written expressions, and numerical formatting processing to convert non-standard numerical expressions into standard numerical formats.

[0015] The standardized business entity data is associated with intent tags to obtain structured data.

[0016] In one possible implementation, after obtaining the completed business form, the process further includes:

[0017] Generate a preview form for the completed business form; obtain user input based on the preview form, including confirmation, editing, or abandonment; in response to the confirmation instruction, submit the completed business form to the downstream business system; in response to the editing instruction, enter the form editing state and regenerate the preview form after editing; in response to the abandonment instruction, terminate the current form filling process.

[0018] Alternatively, based on preset data flow rules, determine whether the filled-in business form meets the transfer conditions corresponding to the current node. The data flow rules include at least one process node and the transfer conditions between nodes. If the transfer conditions are met, submit the filled-in business form to the next process node corresponding to the transfer conditions. If the transfer conditions are not met, generate a prompt message and pause automatic triggering.

[0019] In one possible implementation, business entity data is mapped to corresponding fields in the target business form to obtain the filled-in business form, including:

[0020] Map the business entity data to the corresponding fields in the target business form to obtain the business form to be validated;

[0021] Based on the rule engine, the data reasonableness of the business forms to be validated is checked, and the validation results are obtained.

[0022] When the verification result indicates that the verification has passed, the business form to be verified will be used as the business form after being filled in.

[0023] In one possible implementation, the target business form corresponding to the intent tag is determined based on a knowledge base that associates intent tags with business forms, including:

[0024] The current business scenario is determined based on the business context corresponding to the intent tag and / or business source data;

[0025] Based on the current business scenario, load the corresponding target intent tag and business form associated knowledge base;

[0026] Based on the knowledge base that associates target intent tags with business forms, the target business form corresponding to the intent tag is determined.

[0027] In one possible implementation, the knowledge base associated with intent tags and business forms is dynamically updated in the following way:

[0028] Collect feedback data from the historical data entry process;

[0029] Based on the feedback data, machine learning algorithms were used to adjust the mapping relationship between intent tags in the knowledge base and business forms;

[0030] The adjustments include at least one of the following: increasing or decreasing the confidence weight of existing mapping entries, adding mapping entries between intent tags and business forms, deleting mapping entries that meet preset elimination conditions, and merging or splitting semantically similar intent tags.

[0031] In one possible implementation, when the business source data includes both voice data and document data, a semantic parsing model is used to perform deep semantic parsing on the unstructured text data to obtain structured data that conforms to the value standards of business form fields, including:

[0032] A semantic parsing model is used to perform deep semantic parsing on the unstructured text data corresponding to the voice data, so as to obtain the first intermediate structured data that meets the value standards of business form fields.

[0033] A semantic parsing model is used to perform deep semantic parsing on the unstructured text data corresponding to the document data, so as to obtain the second intermediate structured data that conforms to the value standards of the business form fields.

[0034] Cross-validation is performed on the first intermediate structured data and the second intermediate structured data. Cross-validation includes aligning the first intermediate structured data and the second intermediate structured data according to the business entity field, and comparing the values ​​of the same business entity field and the corresponding confidence scores.

[0035] Based on the preset confidence strategy, the data source with the higher confidence score is selected from the first intermediate structured data and the second intermediate structured data to obtain structured data that meets the value standards of the business form fields.

[0036] In one possible implementation, an information extraction strategy corresponding to the modality type is applied to convert the business source data into unstructured text data, including:

[0037] When the modality type is speech, streaming speech recognition technology deployed on edge computing nodes is used to perform local recognition processing on business source data and convert the business source data into unstructured text data in real time.

[0038] When the modality type is document modality, document recognition and document understanding technologies are used to parse the business source data, extract the text content and structured information in the business source data, and generate unstructured text data.

[0039] Secondly, this application provides an AI-based business form filling device, comprising:

[0040] The acquisition module is used to acquire the business source data required for filling in business forms and to determine the modal type of the business source data, which includes voice data and / or document data.

[0041] The processing module is used for information extraction strategies corresponding to modal types, converting business source data into unstructured text data;

[0042] The semantic parsing module is used to perform deep semantic parsing on unstructured text data using a semantic parsing model to obtain structured data that conforms to the business form field value standards. The structured data includes business entity data and intent tags.

[0043] The determination module is used to determine the target business form corresponding to the intent tag based on the knowledge base associated with the intent tag and the business form.

[0044] The data entry module is used to map business entity data to the corresponding fields in the target business form, resulting in the filled-in business form.

[0045] In one possible implementation, the semantic parsing module is specifically used for: inputting unstructured text data into a semantic parsing model, performing entity extraction and intent classification through the semantic parsing model to identify business entity data and intent tags in the unstructured text data; the semantic parsing model is a natural language understanding model fine-tuned for the target business domain; performing normalization transformation on the business entity data to obtain normalized business entity data, the normalization transformation including: semantic normalization processing to convert colloquial expressions into standard written expressions, and numerical formatting processing to convert non-standard numerical expressions into standard numerical formats; associating the normalized business entity data with intent tags to obtain structured data.

[0046] In one possible implementation, the processing module is further configured to:

[0047] Generate a preview form for the completed business form; obtain user input based on the preview form, including confirmation, editing, or abandonment; in response to the confirmation instruction, submit the completed business form to the downstream business system; in response to the editing instruction, enter the form editing state and regenerate the preview form after editing; in response to the abandonment instruction, terminate the current form filling process.

[0048] Alternatively, based on preset data flow rules, determine whether the filled-in business form meets the transfer conditions corresponding to the current node. The data flow rules include at least one process node and the transfer conditions between nodes. If the transfer conditions are met, submit the filled-in business form to the next process node corresponding to the transfer conditions. If the transfer conditions are not met, generate a prompt message and pause automatic triggering.

[0049] In one possible implementation, the data entry module is specifically used to: map business entity data to corresponding fields in the target business form to obtain the business form to be verified; perform data rationality verification on the business form to be verified based on the rule engine to obtain the verification result; and when the verification result indicates that the verification has passed, use the business form to be verified as the business form after data entry.

[0050] In one possible implementation, the determining module is specifically used to: determine the current business scenario based on the business context corresponding to the intent tag and / or business source data; load the corresponding target intent tag and business form association knowledge base based on the current business scenario; and determine the target business form corresponding to the intent tag based on the target intent tag and business form association knowledge base.

[0051] In one possible implementation, the knowledge base for associating intent tags with business forms is dynamically updated in the following way: feedback data is collected during the historical data entry process; based on the feedback data, a machine learning algorithm is used to adjust the mapping relationship between intent tags and business forms in the knowledge base; wherein, the adjustment includes at least one of the following: increasing or decreasing the confidence weight of existing mapping entries, adding mapping entries between intent tags and business forms, deleting mapping entries that meet preset elimination conditions, and merging or splitting semantically similar intent tags.

[0052] In one possible implementation, when the business source data includes both voice data and document data, the semantic parsing module is further configured to: employ a semantic parsing model to perform deep semantic parsing on the unstructured text data corresponding to the voice data to obtain first intermediate structured data that conforms to the business form field value standards; employ a semantic parsing model to perform deep semantic parsing on the unstructured text data corresponding to the document data to obtain second intermediate structured data that conforms to the business form field value standards; perform cross-validation on the first intermediate structured data and the second intermediate structured data, the cross-validation including aligning the first intermediate structured data and the second intermediate structured data according to business entity fields, and comparing the values ​​of the same business entity field and their corresponding confidence scores; and select the data source with the higher confidence score from the first intermediate structured data and the second intermediate structured data according to a preset confidence strategy to obtain structured data that conforms to the business form field value standards.

[0053] In one possible implementation, the processing module is further configured to: when the modality type is speech modality, use streaming speech recognition technology deployed on edge computing nodes to perform local recognition processing on the business source data, and convert the business source data into unstructured text data in real time; when the modality type is document modality, use document recognition and document understanding technology to parse the business source data, extract the text content and structured information in the business source data, and generate unstructured text data.

[0054] Thirdly, this application provides an electronic device, including: a memory and a processor;

[0055] The memory stores the instructions that the computer executes;

[0056] The processor executes computer execution instructions stored in memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0057] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible embodiments of the first aspect.

[0058] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0059] This application provides an AI-based business form filling method, device, and storage medium. By acquiring multimodal business source data including voice and / or document data, and employing corresponding information extraction strategies for different modalities, the raw data is uniformly converted into unstructured text data. This ensures that multi-source heterogeneous inputs can be automatically parsed and utilized by the system without manual transcription and input. Furthermore, a semantic parsing model is used to perform deep semantic parsing on the unstructured text data, extracting business entity data and intent tags to obtain structured data that conforms to the business form field value standards. This converts natural language expressions into standardized data that can be used for subsequent mapping, improving the mismatch between unstructured input and form field formats. Then, based on intent tags and a business form association knowledge base, the target business form is determined, and the business entity data is mapped to the corresponding fields to complete automatic filling, eliminating the need for manual form type determination and item-by-item field value entry. Therefore, in multimodal, unstructured business form filling scenarios with varying form types, the process from multi-source input to automatic filling is automated, reducing manual operations, improving filling efficiency, and reducing input errors caused by human error. Attached Figure Description

[0060] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0061] Figure 1 A flowchart illustrating the AI-based business form filling method provided in this application embodiment. Figure 1 ;

[0062] Figure 2 A flowchart illustrating the AI-based business form filling method provided in this application embodiment. Figure 2 ;

[0063] Figure 3 A schematic diagram of the structure of the AI-based business form filling device provided in the embodiments of this application;

[0064] Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0065] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0066] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application.

[0067] The terms “first,” “second,” etc., used in this application’s specification are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, products, or apparatus.

[0068] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0069] In digital business processes such as government administration, financial credit, medical services, and enterprise operations management, business form completion is a crucial foundational step connecting information collection, business review, and workflow. Related technologies are typically deployed in an application environment comprised of user terminals, business processing platforms, and form management systems. Users can input information into corresponding application forms, approval forms, medical record forms, reimbursement forms, or archive forms via computers, mobile terminals, or self-service devices. In real-world scenarios, business source data often doesn't exist directly in standard field formats but is scattered across various media such as voice descriptions, invoice images, contract texts, medical records, policy attachments, and supporting documents. However, in these technologies, business form completion primarily relies on manual operation. Specifically, users input business data item by item into the corresponding fields of the form using keyboard input or handwriting. After completion, the field format is manually checked item by item. Once this process is complete, the business form is considered finished and can be submitted to the next stage. The basic working principle of this method is that users manually input data, and then operators fill in the relevant content into preset fields based on experience. In business environments with many form types, manual template selection is often also required.

[0070] The aforementioned manual form-filling method has the following technical drawbacks: First, low efficiency. Because business source data is scattered across multiple media, users need to repeatedly consult materials from different sources and then input data item by item using a keyboard or handwriting. The average time for a single complex form can reach 15 to 30 minutes, making it difficult to meet the needs of high-frequency, high-volume business processing. Second, data accuracy is difficult to guarantee. Manual input and verification processes are easily affected by fatigue, negligence, or misunderstandings of business rules, leading to frequent problems such as incorrect field entries, formatting errors, and logical contradictions. The data error rate is approximately 5% to 10%. Third, insufficient automation. When business source data exists in unstructured forms such as voice, images, or scanned documents, the system cannot automatically extract valid information and complete field mapping. Users must manually transcribe and input the data, and also manually determine the form type and field correspondences.

[0071] As a result, the data entry process is not only time-consuming and error-prone, but also prone to problems such as incorrect or missing fields, mismatched forms, and inconsistent formats. Subsequent approval, archiving, and review processes are also affected by the unstable quality of the front-end data. Especially in scenarios with high requirements for accuracy and standardization, such as government affairs, finance, and healthcare, relying on manual labor is no longer sufficient to support the continuous, efficient, and stable operation of complex businesses.

[0072] To address the aforementioned technical challenges, this application provides an AI-based business form filling solution that constructs an automated filling mechanism. First, it acquires multimodal business source data, including voice data and / or document data, and employs corresponding information extraction strategies for different modalities to uniformly convert business information scattered across heterogeneous data sources such as voice descriptions, invoice images, and contract texts into unstructured text data. Second, it utilizes a semantic parsing model to perform deep semantic analysis on the unstructured text data, extracting business entities and intent tags to obtain structured data that meets the form field value requirements, thus completing the conversion from natural language to structured data. Finally, it automatically matches the target business form based on intent tags, completing field mapping and automatic filling. This forms an end-to-end processing flow from multi-source input to automatic filling, reducing manual intervention and improving filling efficiency and accuracy.

[0073] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0074] Figure 1 A flowchart illustrating the AI-based business form filling method provided in this application embodiment. Figure 1 ,like Figure 1 As shown, this AI-based business form filling method includes:

[0075] S101. Obtain the business source data required for filling in the business form, and determine the modal type of the business source data. The business source data includes voice data and / or document data.

[0076] In this step, the business source data serves as the input foundation for subsequent processing flows, carrying the original information content related to the business to be filled in. Business source data may include voice data, document data, or both. Voice data represents business information expressed orally by the user, while document data represents invoices, contracts, receipts, medical records, supporting documents, application attachments, screenshots, scans, or other electronic materials that reflect business facts, such as images and scans in JPEG, PNG, BMP, TIFF formats, and files in PDF, DOC, DOCX, XLS, XLSX, PPT, PPTX, TXT formats. In this embodiment, multimodal AI is used to integrate the aforementioned different modalities of business source data into the same business form filling process. Through collaborative processing of front-end acquisition, modality recognition, text conversion, semantic understanding, and form mapping, automated filling capabilities are achieved.

[0077] In specific implementation, the executing entity of this application embodiment can be a form filling service deployed in a business processing platform, or a processing unit set in a user terminal, local server, cloud server, or edge computing node.

[0078] For example, voice data can be collected through web-based recording controls, mobile terminal voice input interfaces, self-service terminal audio pickup devices, call center access voice streams, or external microphone arrays. During collection, speaking timestamps, session identifiers, user identifiers, business serial numbers, and voice segment sequence numbers can be recorded simultaneously for subsequent semantic association and multi-round content merging. In one possible embodiment, voice data is first locally cached and fragmented at edge computing nodes before being transmitted to the business processing platform to reduce privacy risks associated with the transmission of raw voice data and minimize the impact of network fluctuations on continuous recognition. For document data, it can be obtained through methods such as photo upload, file import, scanner acquisition, business system interface retrieval, email attachment parsing, or shared document directory monitoring. Document file types can include image files, portable document format files, text processing files, spreadsheet files, and composite documents containing embedded images.

[0079] To ensure the feasibility of subsequent processing, this step can also perform preliminary preprocessing on the acquired business source data. For voice data, noise reduction, echo suppression, silence detection, speech segmentation, speaker differentiation, and sampling rate unification can be performed to improve the quality of subsequent speech-to-text transcription. For document data, layout cleaning, image denoising, skew correction, resolution normalization, page segmentation, duplicate page removal, and document integrity verification can be performed to improve document recognition stability. If the same business corresponds to data from multiple sources, this step can also establish a unified data index, for example, using the business order number as the primary key, merging voice recordings, invoice images, explanatory texts, and related attachments for the same application into the same data container, thereby providing a foundation for subsequent multimodal fusion processing.

[0080] Based on the above analysis, this step addresses the problem of scattered sources, inconsistent formats, and difficulty in directly integrating raw information into the automated data entry process in complex business scenarios. By simultaneously supporting voice and document data during the data collection phase and establishing a unified data organization structure, business information originally distributed across different terminals and file formats can be incorporated into a computable processing scope. This provides complete input for subsequent information extraction and semantic analysis, thereby reducing delays, omissions, and format inconsistencies caused by manual processing of raw data.

[0081] It should be understood that the acquisition location, acquisition device, and upload method of the above-mentioned business source data can be adjusted according to the actual business environment, as long as the voice data and / or document data related to form filling can be obtained. This application embodiment does not impose any restrictions on this.

[0082] S102. Apply the information extraction strategy corresponding to the modality type to convert the business source data into unstructured text data.

[0083] In this step, the information extraction strategy is used to perform matching recognition and extraction processing for different modal types, so as to transform voice data and document data into uniformly processable unstructured text data. In this embodiment, unstructured text data refers to text results that have not yet been mapped to a field-level structure but can already express business content. It can originate from voice recognition results, document text extraction results, or from table parsing results, title content, notes, text near seals, and multi-page contextual text concatenated within documents. Because voice and documents differ significantly in their information representation, this step does not employ a single extraction method but instead sets up processing flows based on modal type.

[0084] For example, in one possible implementation, step S102 may specifically include: when the modality type is speech modality, using streaming speech recognition technology deployed on edge computing nodes to perform local recognition processing on business source data, and converting the business source data into unstructured text data in real time; when the modality type is document modality, using document recognition and document understanding technology to parse the business source data, extract the text content and structured information in the business source data, and generate unstructured text data.

[0085] Edge computing nodes are used to perform speech recognition and preliminary text generation processing close to the data generation point, reducing privacy risks and network transmission latency associated with the transmission of raw speech. Streaming speech recognition technology refers to a recognition mechanism that segments and decodes continuous speech streams and outputs the results in real time. It can continuously generate intermediate recognition results before the speech ends and form complete text after the sentence ends. Document recognition and document understanding technologies combine character recognition, layout analysis, and semantic structure parsing capabilities to extract information such as text, titles, tables, paragraphs, field relationships, signatures, and QR codes from documents such as scanned documents, photos, PDFs, or screenshots, forming unstructured text data that can be uniformly processed by the subsequent semantic parsing module.

[0086] When the source data includes voice data, the modality type is voice modality. Specifically, the system first performs endpoint detection and framing processing on the voice stream, then extracts acoustic feature vectors, such as Mel-frequency cepstral coefficients, filter bank features, or implicit acoustic representations based on neural network learning, and inputs them into the speech recognition model to generate transcribed text. Optionally, during the recognition process, a business dictionary, industry terminology glossary, entity hot words, and contextual error correction mechanisms can be combined to specifically enhance high-frequency business elements such as amount, date, department name, project name, disease name, and ID number. For example, when a user says, "I'm claiming reimbursement for 1,000 yuan in travel expenses, and the invoice was issued last Friday," the streaming speech recognition system, in addition to outputting the original transcribed text, can also retain sentence segmentation information, pause information, and suspected keyword markers so that the subsequent semantic parsing module can identify the relationship between the expression of amount and the expression of expense type.

[0087] When the source data includes document data, the modality type is document modality. Specifically, the layout analysis model can first identify the title area, body text area, table area, seal area, signature area, and attachment area in the document, and then extract the character text using an optical character recognition engine. For documents with layout structures such as tables, invoices, contract pages, medical record pages, and reimbursement lists, a key-value pair extraction model, a table structure restoration model, and a layout semantic annotation model can be combined to output the relationships between text blocks. This yields not only isolated text fragments but also the relative position of the text on the page, its region, and its semantic connections to the preceding and following text. In one possible embodiment, the system reorganizes the text blocks in the document according to the reading order, field adjacency relationships, or table logical order to generate unstructured text data that is closer to natural semantic expression, thereby reducing the contextual fragmentation caused by simple character extraction.

[0088] For scenarios involving both voice and document data, this step can also perform cross-modal text convergence processing. Specifically, the speech-to-text, extracted document text, and their source identifiers can be uniformly written into a text fusion buffer and concatenated based on temporal proximity, business topic similarity, entity co-occurrence, or user session context. For example, if a user first utters "This is my hotel invoice for a business trip to Shanghai" and then uploads a picture of the hotel invoice, the system can associate the speech-to-text with the invoice recognition text to form a unified unstructured text corpus, providing a basis for subsequent models to more accurately understand the type of accommodation fee, invoice date, and amount information. To avoid the amplification of conflicts between texts from different sources, this step can also retain text source labels, confidence scores, and original segment indexes, enabling subsequent steps to trace back to the specific modal source when ambiguity arises.

[0089] Based on the above analysis, the core of this step lies in establishing a unified text entry point for multimodal input, parsing the originally heterogeneous speech signals and document images into text representations that can be incorporated into the natural language understanding model. This not only avoids the fragmented processing of speech and documents separately in related technologies, but also improves the ability of subsequent semantic parsing to grasp the integrity of business through source identification and context reconstruction, providing a foundation for recognizing the correspondence between spoken expressions, abbreviations, invoice text, and explanatory text.

[0090] It should be understood that information extraction strategies can be replaced or combined according to the specific modal type, file format, and scenario requirements of the business source data, as long as the business source data can be converted into unstructured text data. This application's embodiments do not impose such limitations.

[0091] S103. Using a semantic parsing model, perform deep semantic parsing on unstructured text data to obtain structured data that conforms to the business form field value standards. The structured data includes business entity data and intent tags.

[0092] In this step, the standard values ​​for business form fields are used to constrain the output format of structured data, enabling it to be directly utilized by the subsequent field mapping module (i.e., the form filling module). The semantic parsing model performs deep semantic understanding processing on the unstructured text data generated in the previous step, rather than simply relying on literal matching or keyword extraction. The semantic parsing model can be a natural language understanding model pre-trained on a large-scale general corpus and fine-tuned using data from the target business domain. Its input is a unified sequence of unstructured text, a set of text fragments, or multiple text segments with source identifiers; its output is structured data that directly supports form filling. The structured data consists of business entity data and intent tags. Business entity data represents content that can be mapped to form fields, such as amount, date, person name, organization name, expense category, ID number, diagnosis name, contract number, account information, address information, and item name. Intent tags represent the business intent or category corresponding to this business process, such as travel reimbursement, outpatient registration, hospital registration, loan application, qualification application, and file supplementation.

[0093] For example, in one possible implementation, the unstructured text data generated in the aforementioned steps is input into a pre-trained language model. This language model can employ a general semantic parsing model architecture, or it can be fine-tuned based on historical data from the target business domain (such as government affairs, finance, corporate administration, etc.) to improve semantic understanding accuracy in specific scenarios. The language model performs deep semantic parsing on the input text data. This parsing process includes at least two levels of processing: first, identifying key entities with business meaning (such as amounts, dates, entity names, etc.) and the user's expressed business intent (such as reimbursement, project initiation, approval, etc.) from the text data; second, performing format validation and standardization on the identified entity values, such as converting colloquial expressions into standardized written expressions and non-standard numerical abbreviations into standard numerical formats. After the above parsing processing, the language model outputs structured data results.

[0094] In another possible implementation, deep semantic parsing is not limited to using a single semantic parsing model; it can also be achieved through multi-model collaboration. For example, a sequence labeling model for entity recognition and a text classification model for intent classification can be deployed separately. The two models process the same unstructured text data in parallel or sequentially, and then their respective outputs are correlated and fused to generate structured data containing business entity data and intent tags.

[0095] Another possible implementation is that the "compliance with the business form field value standards" in step S103 can be achieved through post-processing rules, rather than relying entirely on the language model's own output. Specifically, the language model first outputs preliminary entity recognition results and intent recognition results, and then an independent rule engine formats the entity recognition results. For example, it converts the recognized date format to "year-month-day", keeps two decimal places for monetary values, and parses "last Friday" into a specific date based on the current session time, thereby obtaining data that meets the form field requirements.

[0096] The above implementation methods can be used individually or in combination, depending on the performance requirements, data scale, deployment environment, and other factors of the specific application scenario. For example, in scenarios with high real-time requirements, a lightweight model combined with a post-processing rule engine can be used; in scenarios with high accuracy requirements, multi-model collaboration or generative models with a large number of parameters can be used.

[0097] Based on the above analysis, this step converts unstructured natural language text into structured data that meets the requirements of business form field values. Compared with solutions that only perform optical character recognition (OCR) or simple transcription, this step significantly improves the ability to recognize the true meaning of business entities and reduces the impact of semantic breaks between spoken expressions, screenshot text, and scanned content on the accuracy of data entry, thereby providing high-quality input for subsequent target form matching and field filling.

[0098] Furthermore, the semantic parsing model in step S103 can be flexibly selected based on the actual deployment environment. When deployed in the cloud, a general model with a large number of parameters can be used to obtain higher semantic understanding accuracy; when deployed at the edge or locally, a lightweight model with a small number of parameters can be used, or the general model can be knowledge distilled and compressed to balance processing speed and recognition accuracy. Regardless of the model form used, its core function is to convert unstructured text data into structured data that conforms to the value standards of business form fields.

[0099] S104. Based on the knowledge base that associates intent tags with business forms, determine the target business form corresponding to the intent tag.

[0100] In this step, the intent tag and business form association knowledge base are used to establish the association between business intents and form templates, form versions, applicable scope, field sets, and constraint rules. The target business form is the form object that ultimately receives the business entity data. It can be a single form template or a form version selected after matching and filtering multiple candidate templates. Because there are many form types and business rules are updated rapidly in real-world business environments, relying solely on manual template selection or fixed string matching can easily lead to mismatches. Therefore, this step uses the linkage between intent tags and the knowledge base to determine the target business form.

[0101] In practice, the knowledge base for associating intent tags with business forms can pre-store multi-dimensional association information, including the mapping relationship between intent tags and form identifiers, applicable departments, applicable business stages, applicable material types, mandatory field rules, historical usage frequency, version effective time, and mutual exclusion conditions. When S103 outputs an intent tag, the system can first retrieve the set of forms directly associated with that intent tag from the knowledge base, and then filter them based on the current business context. The business context can include the user's organization, processing area, terminal type, current process node, type of uploaded materials, session time, and historical processing records. For example, the intent "reimbursement application" might be associated with official travel reimbursement forms in an administrative unit environment, expense reimbursement forms in a corporate environment, but might not trigger such forms in a hospital environment. Through contextual constraints, the system can filter out the target business form that best suits the scenario from multiple candidate forms with the same or similar intents.

[0102] In one possible implementation, when an intent tag corresponds to multiple candidate forms, they can be sorted based on the association strength in the knowledge base. Association strength can be calculated based on historical hit counts, manual confirmation results, field coverage, entity matching ratio, or the degree to which scenario conditions are met. For example, a comprehensive matching score M can be constructed, where M is formed by weighting intent matching weight, context matching weight, and field adaptation weight. The higher the matching score, the more consistent the corresponding form is with the current business request. This comprehensive score quantifies the degree of adaptation between different candidate forms and the current intent tag, thereby avoiding misselection when business names are similar or forms coexist. If multiple forms have similar scores, the system can also retain the candidate form list and recommendation order for subsequent secondary confirmation or further judgment based on field mapping results.

[0103] Based on the above analysis, this step addresses the problem of numerous form types, frequent business changes, and the difficulty of correctly selecting forms using only manual methods or fixed rules in complex business environments. By jointly matching the semantic understanding result of intent tags with a knowledge base, the system can determine the form's affiliation from a business semantic perspective, rather than relying solely on filenames, template numbers, or human experience. Therefore, it improves form adaptation capabilities in new scenarios, cross-departmental scenarios, and scenarios with similar business meanings, and reduces subsequent rejections, re-entry, and approval failures caused by mismatched forms.

[0104] S105. Map the business entity data to the corresponding fields in the target business form to obtain the filled-in business form.

[0105] In this step, the corresponding field refers to the location in the target business form where specific business entity data is entered. This can be a text field, numeric field, date field, enumeration field, boolean field, table cell, attachment index, or composite field. The completed business form refers to the previewable, submitable, transferable, or archiveable form result formed after field mapping and data writing. Because target business forms typically have differences in field names, field format constraints, required fields, and cross-field validation rules, business entity data cannot be simply copied item by item; instead, mapping must be performed according to field semantics and constraint relationships.

[0106] For example, in one possible implementation, the system first reads the field definition information of the target business form, including field identifier, field name, field alias, data type, length limit, value range, enumeration value set, required status, dependency relationship, and validation rules. Then, it establishes a mapping relationship between entities and fields based on the business entity categories in the structured data. The mapping process proceeds in the order of field name matching, field semantic matching, and field constraint validation: first, an initial candidate field set is established based on the entity category and field category; then, semantic matching is completed by combining the field description text, adjacent field relationships, and business grouping information in the form template, mapping the amount entity to the "reimbursement amount" field, the invoice date to the "invoice date" field, the organization name to the "invoicing unit" field, and the expense category derived from the intent tag to the "expense type" field; finally, field constraint validation is performed, such as checking whether the amount format retains two decimal places, whether the date meets the year-month-day format, whether the document number length conforms to the rules, and whether the enumeration value belongs to the limited options. If the target business form has synonymous field names or composite fields, the system can determine the final landing point by combining the field alias dictionary, knowledge base field relationships, and fine-grained categories of entity labels output by the model.

[0107] In another possible implementation, the system pre-configures a field mapping table, which records the correspondence between business entity types and fields in the target business form. Each entry in the mapping table contains at least the source entity identifier, target field identifier, conversion rule, and priority information. Based on the business entity type in the structured data, the system queries the corresponding target field in the mapping table and directly fills the entity value into the specified field position. This method eliminates the need for real-time semantic matching calculations and is suitable for application environments with fixed business scenarios, minimal changes in form structure, or high performance requirements. When the same business entity type corresponds to multiple candidate fields in the mapping table, the system can select the field with the highest matching degree for filling based on priority information, or generate a candidate list for user confirmation.

[0108] After completing the field mapping using any of the above methods, the system writes the business entity data into the corresponding fields of the target business form. For business entity data that cannot be matched to a field, the system can choose to ignore it, store it in an additional information field, or generate a prompt message for the user to handle, according to a preset strategy. After the above mapping and filling operations, each field in the target business form obtains the corresponding data value, thus forming a complete business form that can be directly used for subsequent processes.

[0109] Optionally, after mapping, this step can also generate form validation results and perform a completeness check on the filled-in business forms. Completeness checks can include checking for missing required fields, consistency of cross-field logic, balanced summaries, reasonable date order, and completeness of attachment types. For example, if the travel expense reimbursement form includes accommodation costs but not the travel date range, the system can mark that field as needing completion; if the admission date in the hospital registration form is later than the discharge date, an exception check will be triggered. For issues that can be automatically corrected, such as inconsistent date separators, missing decimal places in amounts, or extra spaces in the address field, this step can also perform formatting repair before outputting the final form.

[0110] This application embodiment acquires multimodal business source data including voice data and / or document data, and uses corresponding information extraction strategies for different modalities to uniformly convert the raw data into unstructured text data. This ensures that multi-source heterogeneous inputs can be automatically parsed and utilized by the system without manual transcription and input. Based on this, a semantic parsing model is further used to perform deep semantic parsing on the unstructured text data, extracting business entity data and intent tags to obtain structured data that conforms to the business form field value standards. This converts natural language expressions into standardized data that can be used for subsequent mapping, improving the mismatch between unstructured input and form field formats. Furthermore, based on the intent tags and the business form association knowledge base, the target business form is determined, and the business entity data is mapped to the corresponding fields to complete automatic filling, eliminating the need for manual judgment of form types and item-by-item field value entry. Therefore, in multimodal, unstructured business form filling scenarios with varying form types, the process from multi-source input to automatic filling is automated, reducing manual operation steps, improving filling efficiency, and reducing input errors caused by human error.

[0111] In some embodiments, a semantic parsing model is used to perform deep semantic parsing on unstructured text data to obtain structured data that conforms to the value standards of business form fields. This includes: inputting unstructured text data into the semantic parsing model, performing entity extraction and intent classification through the semantic parsing model to identify business entity data and intent tags in the unstructured text data; the semantic parsing model is a natural language understanding model finely tuned for the target business domain; performing normalization transformation on the business entity data to obtain normalized business entity data, the normalization transformation including: semantic normalization processing to convert colloquial expressions into standard written expressions, and numerical formatting processing to convert non-standard numerical expressions into standard numerical formats; associating the normalized business entity data with intent tags to obtain structured data.

[0112] In this embodiment, the semantic parsing model is a natural language understanding model fine-tuned for the target business domain. The purpose is to enable the model to adapt to differences in industry terminology, form field constraints, and business intent expression, thereby improving the accuracy of entity extraction and intent classification.

[0113] For example, unstructured text data is first segmented, sub-word encoded, positionally encoded, and source-label encoded, then input into the domain-fine-tuned language model backbone network for contextual representation learning. Subsequently, an entity extraction head and an intent classification head are set up in parallel at the model output layer. The entity extraction head can use sequence labeling, pointer networks, or generative extraction methods to identify the boundaries of segments related to business entities and entity categories in the text; the intent classification head outputs intent labels based on the overall semantic vector, multi-sentence aggregation vector, or key segment vector. For text like "Reimburse 1,000 yuan for travel expenses, invoice date is May 10, 2024," the model can extract the amount entity "1000 yuan," the expense type entity "travel expenses," and the date entity "2024-05-10," while identifying the intent label as "travel reimbursement application." For text like "Zhang San reimburses 800 yuan for travel expenses," the model can extract the reimbursement person entity "Zhang San," the expense type entity "travel expenses," and the amount entity "800 yuan."

[0114] To ensure the parsed results conform to the business form field value standards, further normalization transformation is performed. Semantic normalization, based on a business thesaurus, synonym mapping table, and domain rule base, converts colloquial, abbreviated, or informal expressions into standardized written expressions suitable for form filling. For example, "approximately one thousand" is normalized to a specific numerical value, "First Municipal Hospital" to the full name of the medical institution, "early this month" to a standard date, and "what to do" to a standardized operational expression. Numerical formatting standardizes Chinese numerals, mixed numerals, colloquial numerals, and expressions with units. For example, "one thousand yuan," "twenty-five thousand," and "3.5w" are converted to unified Arabic numerals and standard unit representations to ensure that field values ​​can be directly used for form mapping and system validation. After normalization, the normalized business entity data is associated with intent tags to form structured data containing both field content and business semantics, providing reliable input for subsequent target form determination and field filling.

[0115] In terms of working principle, the semantic parsing model first performs semantic understanding on the input text, then standardizes the identified entities, and finally binds the entity results with intent tags, thereby transforming the original unstructured text into a structured result oriented towards form processing. This approach can reduce the impact of spoken language, non-standard numbers, and differences in business terminology on field entry, enabling the system to maintain high parsing consistency even in complex business scenarios.

[0116] In this embodiment of the application, by accurately extracting and standardizing business information from unstructured text, the correspondence between business entity data and business intent becomes clearer, thereby improving the accuracy and automation level of form field values, reducing the workload of manual verification and repetitive data entry, and enhancing the system's adaptability to different business domains and different expression methods.

[0117] In some embodiments, after obtaining the completed business form, the process further includes:

[0118] Generate a preview form for the completed business form; obtain user input based on the preview form, including confirmation, editing, or abandonment; in response to the confirmation instruction, submit the completed business form to the downstream business system; in response to the editing instruction, enter the form editing state and regenerate the preview form after editing; in response to the abandonment instruction, terminate the current form filling process.

[0119] Alternatively, based on preset data flow rules, determine whether the filled-in business form meets the transfer conditions corresponding to the current node. The data flow rules include at least one process node and the transfer conditions between nodes. If the transfer conditions are met, submit the filled-in business form to the next process node corresponding to the transfer conditions. If the transfer conditions are not met, generate a prompt message and pause automatic triggering.

[0120] In other words, after receiving the completed business form, the system supports either manual confirmation mode or automatic trigger mode to achieve closed-loop business management.

[0121] The preview form is a visual representation of the completed business form, presented to the user before submission. This allows the user to verify field content, identify discrepancies, and determine subsequent processing methods, such as automatically generating a visual representation through a rendering engine. Operation commands are input by the user based on the preview form, typically triggered by confirmation, edit, or abandon controls on the interface, but can also be generated by touch events, keyboard input, or voice confirmation commands on a mobile terminal. The form editing state allows users to modify previously filled content. In this state, the system retains the original field mappings and re-generates the preview for the modified content, ensuring the displayed result matches the latest submission. The downstream business system receives the confirmed submission results and continues processing subsequent business. Its interface can be implemented using message queues, API calls, or database synchronization. In practical applications, the system can also choose other receiving methods, which are not limited in this embodiment.

[0122] When using data flow rules to control workflow, the system pre-configures at least one process node and the transition conditions between nodes, and matches the currently filled business form with the constraints of the current node. If the form meets the transition conditions, the system automatically submits it to the corresponding next process node to drive the approval, review, or archiving process to continue. If the form does not meet the transition conditions, the system generates a prompt message containing missing items, format errors, or inconsistencies, and pauses automatic triggering until the user corrects the errors before re-entering the judgment process. Data flow rules are used to standardize the complete workflow path required for a business form from completion to final archiving or approval. A process node represents a processing step in the business process, such as "Completion of Form," "Department Head Approval," "Financial Review," and "Archiving." Transition conditions refer to the judgment rules required to transition from the current node to the next node, typically based on field values ​​in the form, business context, or external events, such as "Reimbursement Amount Less Than or Equal to 1000 Yuan," "High Urgency Level," and "All Validations Passed."

[0123] For example, taking a corporate travel expense reimbursement scenario, the data flow rules configured in the system are as follows: The current node is "Reimbursement Completed," and there are two transfer paths starting from this node. The transfer condition for the first path is "reimbursement amount less than or equal to 1000 yuan," and when this condition is met, it is automatically submitted to the "Department Head Approval" node. The transfer condition for the second path is "reimbursement amount greater than 1000 yuan," and when this condition is met, it is automatically submitted to the "Financial Initial Review" node. When an employee submits an expense report with an amount of 800 yuan, the system extracts the amount field from the form, determines that it meets the first transfer condition, automatically submits the expense report to the department head approval node, generates an approval task in the supervisor's to-do list, and sends an approval notification to the supervisor. When the same employee submits an expense report with an amount of 1500 yuan, the system determines that it meets the second transfer condition and automatically submits the expense report to the financial initial review node. If the amount field in the expense report submitted by the employee is empty or contains non-numeric characters, the system determines that it does not meet any transfer conditions and generates a prompt message: "The current form lacks valid amount data, and the next approval node cannot be determined. Please check and resubmit or manually specify the approval path." At the same time, the automatic triggering process is paused, and the expense report is transferred to the manual processing queue, waiting for the user to correct it and then participate in the judgment again.

[0124] In this embodiment, after the business form is filled out, it supports either manual confirmation mode or automatic triggering mode, so that the filling results can be submitted securely after manual confirmation or automatically flow under rule constraints, thereby improving the accuracy of form submission, the continuity of flow and the efficiency of business processing, and realizing an end-to-end closed loop from data collection to process flow.

[0125] In some embodiments, mapping business entity data to corresponding fields in a target business form to obtain a filled-in business form includes: mapping business entity data to corresponding fields in a target business form to obtain a business form to be verified; performing data validity verification on the business form to be verified based on a rule engine to obtain a verification result; and using the business form to be verified as the filled-in business form when the verification result indicates that the verification has passed.

[0126] The business entity data consists of structured business information extracted from semantic parsing results and processed through normalization. This typically includes data items directly usable for form filling, such as name, amount, date, item type, and document information. The target business form is a standard form template matching the current business intent, pre-defined with field names, field types, value ranges, and validation constraints. The business form to be validated represents the intermediate form result after field mapping, used for subsequent reasonableness checks. The rule engine is a validation module built based on pre-defined business rules, field dependencies, and constraints, used to identify situations such as out-of-range filling, inconsistencies, missing items, or formatting errors. The validation result characterizes the rule engine's conclusion regarding the business form to be validated, typically including at least two states: pass and fail.

[0127] For example, after identifying the target business form, the system first performs semantic matching and field alignment between the business entity data and the form fields. For instance, it maps "reimbursement amount" to the amount field, "invoice date" to the date field, and "ID number" to the document number field. After mapping, a business form to be verified is generated. Next, the system performs data validity checks on the business form based on a rule engine. The rule engine has built-in pre-defined business rules, field dependencies, and constraints. Verification content may include, but is not limited to: whether a field is empty, whether a value exceeds a preset threshold, whether the date order is reasonable, whether the enumerated value belongs to the allowed set, and whether related fields meet business logic constraints. The rule engine iterates through each field in the business form to be verified, checking each item to see if it meets the above rule conditions, and generates corresponding verification results based on the check results. Finally, the system determines the final business form based on the verification results. When the verification result indicates that the verification has passed, meaning that all fields in the business form to be verified meet the preset rules, the system directly confirms the business form to be verified as the completed business form. When the verification result indicates that the verification failed, the system outputs the failed verification result and generates a corresponding prompt message. For example, when the daily accommodation cost for a business trip exceeds 600 RMB, the system prompts "Accommodation cost exceeds the standard, additional materials are required." The prompt message can be sent via email, voice call, SMS, pop-up window, or one or more other methods, depending on the urgency of the application scenario and user configuration.

[0128] In practical applications, this rule engine can be implemented in various ways. For example, it can use a built-in rule table to store business rules in tabular form, facilitating configuration and maintenance by business personnel; it can use an expression parser to support the dynamic writing and parsing of rule expressions, adapting to more complex judgment logic; or it can use a scripted rule configuration method, describing the rule logic through a scripting language to meet highly customized business needs. This application's embodiments do not limit this approach.

[0129] Taking a corporate travel expense reimbursement scenario as an example, after the system completes field mapping, it generates an expense form to be verified, where the value of the accommodation expense field is 800 yuan. The default verification rule in the rule engine is "accommodation expenses per day shall not exceed 600 yuan". The rule engine reads this field value, compares it with the preset threshold, determines that 800 yuan is greater than 600 yuan, triggers a rule conflict, outputs a failed verification result, and generates a prompt message: "Accommodation expenses exceed the standard, additional materials are required", which is displayed to the user in a pop-up window. After the user supplements the materials or modifies the amount according to the prompt, the system re-executes the verification until all fields meet the rule requirements. Finally, the business form to be verified is confirmed as the filled-in business form for subsequent approval or archiving processes.

[0130] In this embodiment, by adding a rule engine validation step after field mapping, the system can complete form validity screening before downstream submission, avoiding erroneous submissions due to entity mismatch, missing fields, or abnormal values. Since the business form to be validated is only recognized as the submitted business form after passing validation, the accuracy, consistency, and business usability of form data can be significantly improved, reducing manual review costs and enhancing the stability of automated submission in complex business scenarios.

[0131] In some embodiments, determining the target business form corresponding to an intent tag based on a knowledge base associated with the intent tag and the business form includes: determining the current business scenario based on the business context corresponding to the intent tag and / or business source data; loading the corresponding knowledge base associated with the target intent tag and the business form based on the current business scenario; and determining the target business form corresponding to the intent tag based on the knowledge base associated with the target intent tag and the business form.

[0132] The business context is used to characterize the business background information implicit in the business source data and the business semantic scope corresponding to the intent tags. It can be comprehensively determined by keywords, business objects, time elements, location elements, amount elements, or item names in the business source data, and is used to assist in determining the current business scenario. The current business scenario represents the business environment in which the system operates. When the business source data comes from different business types such as reimbursement, application, medical treatment, or contract processing, the system selects the matching knowledge base accordingly to avoid calling the wrong form mapping relationship in different scenarios. The target intent tag and business form association knowledge base is used to store the correspondence between business scenarios, intent tags, and business forms. The knowledge base can be organized using a table structure, key-value index structure, or graph structure, and different versions can be configured according to the business domain to facilitate quick switching when the business type changes.

[0133] For example, after obtaining the intent tag and business source data, the system first performs semantic parsing on the business source data to extract contextual features that reflect the business background. Then, these contextual features, along with the intent tag, are input into the scenario determination unit to output the current business scenario. For instance, if the business source data contains expense reports, invoices, or travel descriptions, the current business scenario can be identified as a travel expense reimbursement scenario; if it contains medical records, examination reports, or registration information, the current business scenario can be identified as an outpatient visit scenario. Subsequently, based on the current business scenario, the system retrieves the corresponding target intent tag and business form association knowledge base from the storage medium and loads this knowledge base into memory or a cache for subsequent fast querying. The loaded knowledge base records the correspondence between intent tags and target business forms; for example, "travel expense reimbursement" corresponds to an expense reimbursement form, and "outpatient follow-up visit" corresponds to an outpatient medical record. The system then matches and searches for the intent tag based on this knowledge base, outputs the corresponding target business form, and uses this form as the carrier for subsequent field mapping. In practical applications, the knowledge base can also adopt a dynamic update mechanism to maintain it synchronously when the form version is adjusted, business rules are changed, or new business items are added. This application embodiment does not limit this.

[0134] In this embodiment, by first determining the business scenario and then loading the corresponding knowledge base, the form selection of intent tags can be limited to the current business context, thereby reducing the probability of cross-scenario mismatch. Since the knowledge base pre-stores scenario-based intent tag and form mapping relationships, the system does not need to traverse all form templates one by one when determining the target business form, improving retrieval efficiency and enabling rapid adaptation to various business scenarios such as government affairs application, financial credit approval, medical record entry, and corporate human resources administration, without requiring repeated development for a single scenario. Furthermore, for scenarios with multiple business source data, this method can also utilize contextual information to compensate for potential ambiguities or incomplete expressions in a single intent tag, making the form selection results more consistent with the actual business semantics, thereby improving the accuracy and consistency of subsequent field entry.

[0135] Understandably, business scenarios and form types are not static. Over time, new business requirements may arise, rendering the original mapping between intent tags and forms inapplicable. Alternatively, users may develop new expression habits and form selection preferences during actual use. If the mapping in the knowledge base remains static, it will gradually deviate from actual business needs, leading to a decrease in the accuracy of intent tag matching, increasing the workload for users to manually correct form types, and even affecting the efficiency of the entire form filling process.

[0136] In view of this, in some embodiments, the knowledge base associated with intent tags and business forms is dynamically updated in the following ways: collecting feedback data from the historical data entry process; and adjusting the mapping relationship between intent tags and business forms in the knowledge base using machine learning algorithms based on the feedback data. The adjustment includes at least one of the following: increasing or decreasing the confidence weight of existing mapping entries, adding mapping entries between intent tags and business forms, deleting mapping entries that meet preset elimination conditions, and merging or splitting semantically similar intent tags.

[0137] In this embodiment, feedback data from the historical data entry process is used to reflect the deviation between the actual data entry results and the knowledge base recommendation results. Feedback data can consist of user modification records of the system's recommended forms, manual correction records, field validation failure records, and downstream system return records. This data is written to the feedback database after completing one data entry loop as a subsequent learning sample. Machine learning algorithms can employ classification models based on sample weight updates, incremental learning models, or clustering analysis models to extract features from the feedback data, generate mapping correction values, and update the correspondence between intent tags and business forms accordingly.

[0138] Specifically, for existing mapping entries, the system adjusts their confidence weight based on feedback consistency, hit rate, and validation pass rate. When the mapping relationship between a certain intent tag and the business form is frequently confirmed by users or passes validation, the system increases the confidence weight of that mapping entry; conversely, when the mapping relationship is frequently modified by users or causes validation failure, the system decreases its confidence weight. The confidence weight is directly used for ranking and recommendation when the same intent tag corresponds to multiple candidate forms during subsequent data entry, giving higher priority to high-frequency and stable mapping relationships.

[0139] Regarding adding new mapping entries, when feedback data indicates that a new intent tag has formed a stable correspondence with a specific business form, the system adds a new mapping entry to the knowledge base and sets an initial weight for that entry.

[0140] Regarding the deletion of mapping entries, the system determines whether an entry needs to be removed based on preset elimination criteria. These criteria include one or more of the following: the mapping entry has not been hit within a preset period or its hit frequency is below a set threshold; the confidence weight of the mapping entry remains below a preset lower threshold; the proportion of user rejections or modifications in the recommended form corresponding to the mapping entry exceeds a preset upper threshold; or the business form associated with the mapping entry has been marked as obsolete or removed. If any of these conditions are met, the system automatically deletes the entry to maintain the simplicity and effectiveness of the knowledge base.

[0141] For multiple semantically similar intent tags, the system can merge semantic vector similarity calculations with clustering results to reduce redundant expressions and simplify the knowledge base structure. In cases where a single intent tag has too broad a coverage, leading to ambiguity in form selection, the system can split it into multiple sub-tags based on the contextual differences of the feedback samples and associate them with different business forms, thereby improving recognition accuracy and adaptability.

[0142] In terms of storage architecture, the feedback database and knowledge base can be stored separately. Mapping entries in the knowledge base should include at least the intent tag identifier, form identifier, confidence weight, and update time field to facilitate quick retrieval and rollback after updates.

[0143] In this embodiment, the aforementioned dynamic update mechanism enables the knowledge base to be continuously revised as business policies change, form templates are adjusted, and user filling habits evolve, thereby making the correspondence between intent recognition results and target forms more stable and reliable. Adopting this approach reduces the probability of mismatches caused by new business categories, changes in semantic expression, or invalid historical rules, improves the accuracy of automatic form allocation, and reduces the workload of manual rule maintenance.

[0144] In real-world business scenarios, users often provide data from multiple sources simultaneously when filling out forms. For example, users might verbally describe their travel expenses while uploading corresponding invoice images or scanned copies of contracts. However, extracting information from a single source can have accuracy bottlenecks: speech recognition may suffer from inaccuracies due to accents, noise, or colloquial expressions, while document recognition may lead to OCR errors due to image quality, complex layouts, or obstructions from seals. If semantic parsing relies solely on one data source, any deviation in the recognition results from that source will directly affect the accuracy of subsequent form filling, and may even result in incorrect field mapping or data entry. Therefore, how to comprehensively utilize information from different sources to improve overall recognition accuracy in scenarios with multiple data sources is one of the technical problems that this application aims to solve.

[0145] In some embodiments, when the business source data includes both voice data and document data, a semantic parsing model is used to perform deep semantic parsing on the unstructured text data to obtain structured data that conforms to the business form field value standards. This includes: using a semantic parsing model to perform deep semantic parsing on the unstructured text data corresponding to the voice data to obtain first intermediate structured data that conforms to the business form field value standards; using a semantic parsing model to perform deep semantic parsing on the unstructured text data corresponding to the document data to obtain second intermediate structured data that conforms to the business form field value standards; performing cross-validation on the first and second intermediate structured data, which includes aligning the first and second intermediate structured data according to business entity fields and comparing the values ​​of the same business entity field and their corresponding confidence scores; and selecting the data source with the higher confidence score from the first and second intermediate structured data according to a preset confidence strategy to obtain structured data that conforms to the business form field value standards.

[0146] The system consists of two intermediate structured data streams: a first stream for receiving deep semantic parsing results from voice data and a second stream for receiving deep semantic parsing results from document data. Both streams use business entity fields, intent tags, and confidence scores as their structured representations. Business entity fields represent semantic units that can be mapped to form fields, such as name, amount, date, institution name, and medical record information. Confidence scores indicate the reliability of the semantic parsing model's results for the corresponding fields. Cross-validation is used to determine the consistency between the two parsing results and retain the more reliable data source. The two parsing processes are independent and do not interfere with each other.

[0147] For example, in a specific implementation, speech data is first converted into corresponding unstructured text data by a speech recognition unit, while document data is extracted into unstructured text data by an OCR recognition or text parsing unit. These are then input into a semantic parsing model fine-tuned for the specific business domain. When the model receives the text corresponding to the speech, it performs joint analysis on entity boundaries, semantic references, and business intent, outputting first intermediate structured data with field names, field values, intent labels, and confidence scores. Similarly, when the model receives the text corresponding to the document, it outputs second intermediate structured data. After both results are generated, the system aligns them according to a pre-configured business entity field mapping table, merging identical fields into the same comparison dimension, and then jointly compares whether the field values ​​are consistent, whether the standardized forms of the values ​​are consistent, and the confidence scores. When a field appears in both sets of data but the content differs, the system selects the set with the higher confidence level as the final value for that field based on a preset confidence level strategy. When a field appears only in one set of data, it can be retained or reviewed based on the confidence score corresponding to that field, thereby forming structured data that meets the business form field value standards. When the values ​​of the same field are inconsistent in the two sets of intermediate structured data and the confidence scores are both below a preset threshold, the field is marked as requiring manual confirmation, and a prompt message is generated.

[0148] It should be noted that the confidence strategy can be flexibly configured according to the business scenario. For example, a high confidence priority strategy (directly select the result with the higher score), a threshold filtering strategy (only used when the confidence of a certain source exceeds the preset threshold, otherwise marked as pending manual confirmation), or a weighted fusion strategy (merge the two results by weighting according to confidence).

[0149] Taking an employee travel expense reimbursement scenario as an example, a user verbally states, "Accommodation fee 800 yuan," and simultaneously uploads an invoice image showing an accommodation fee of 850 yuan. After parsing the voice data, the system outputs "800 yuan" for the "Accommodation fee" field with a confidence score of 0.85; after parsing the document data, the same field outputs "850 yuan" with a confidence score of 0.92. The system aligns the two intermediate structured data sets by field and compares the value and confidence score of the "Accommodation fee" field in both datasets. Based on a pre-defined high-confidence priority strategy, the system selects the document parsing result "850 yuan," which has a higher confidence score, as the final recognition result for this field, thus obtaining structured data that conforms to the business form field value standards.

[0150] In this embodiment, by parsing both voice and document sources into intermediate structured results and performing cross-validation at the field level, the system can leverage the complementarity of voice and document information simultaneously, reducing errors caused by single-source recognition bias. Since the final output data is selected based on a confidence strategy, the consistency and reliability of field values ​​are improved, thereby increasing form completion accuracy, reducing the frequency of manual review, and enhancing the system's adaptability in multi-source heterogeneous business scenarios.

[0151] Based on the foregoing embodiments, by Figure 2 This document provides a detailed explanation of AI-based business form completion methods. This example uses the scenario of R&D personnel filling out an external technical service fee form in a car manufacturer's R&D project expense declaration as a case study.

[0152] like Figure 2 As shown, this AI-based business form filling method includes:

[0153] S201. Obtain the business source data required for filling in the business form and determine the modal type of the business source data.

[0154] The business source data includes at least one of voice data and document data. Correspondingly, the modal type of the voice data is voice modality, and the modal type of the document data is document modality.

[0155] For example, after completing an outsourced testing project for the battery management system of a new energy vehicle, R&D personnel Zhang verbally stated on his mobile terminal, "This month's external technical service fee is 50,000 yuan. The service provider is Company A, and I have received the invoice." He also simultaneously uploaded a scanned copy of the invoice for this fee from his phone's photo album.

[0156] S202. When the source data of the business is voice data, the modality type is voice modality. Streaming speech recognition technology deployed on edge computing nodes is used to perform local recognition processing on the voice data, converting the continuously input voice content into unstructured text data in real time.

[0157] For example, Zhang's dictated speech was recognized in real time by edge nodes and converted into unstructured text: "This month's external technical service fee is 50,000 yuan, the service provider is A, and the invoice has been received." Because local edge computing processing was used, the speech data was not uploaded to the cloud, avoiding the risk of privacy leaks of corporate financial data.

[0158] S203. When the source data is document data, the modality type is document modality. Document recognition and document understanding technologies are used to parse the document data, extract the text content and structured information in the document data, and generate unstructured text data.

[0159] For example, extract text content such as "Amount: 50,000 yuan", "Invoice date: March 15, 2023", "Buyer: A certain car company", and "Seller: A" from the invoice, and at the same time identify the position of the amount field, QR code information and the official seal of the invoice in the invoice format, and integrate this information into unstructured text data.

[0160] S204. Input unstructured text data into a semantic parsing model, perform entity extraction and intent classification through the semantic parsing model to identify business entity data and intent tags in the unstructured text data; perform normalization transformation on the business entity data to obtain normalized business entity data; associate the normalized business entity data with intent tags to obtain structured data.

[0161] For example, unstructured text data is input into a natural language understanding model fine-tuned for the automotive R&D field. The model performs entity extraction, identifying the following business entities: the fee type is "external technical service fee," the amount is "50,000 yuan," and the service provider is "A." Intent classification is then performed, identifying the user intent as "R&D expense application." The extracted business entity data is then normalized: the colloquial "50,000 yuan" is converted to the standard number "50,000," and "this month" combined with the current system time (March 2025) is converted to the specific date range "March 1, 2025 to March 31, 2025." The normalized business entity data is then associated with the intent tag "R&D expense application" to obtain structured data.

[0162] If the source data includes both voice data and document data (both in this embodiment), deep semantic parsing is performed on the unstructured text data corresponding to the voice data to obtain the first intermediate structured data, such as the fee type "external technical service fee", amount "50000", service provider "A", and intent "R&D expense declaration", with a confidence score of 0.88. Deep semantic parsing is performed on the unstructured text data corresponding to the document data to obtain the second intermediate structured data, such as the invoice amount "50000", invoice date "March 15, 2025", seller "A", and invoice number "12345678", with a confidence score of 0.95. Then, the first and second intermediate structured data are aligned according to the business entity fields. The value of the "amount" field in the two data streams (both are 50000) and the confidence scores (one is 0.88, and the other is 0.95) are compared. According to the preset high confidence priority strategy, the document parsing result with higher confidence is selected as the source of the final structured data. At the same time, the "fee type" obtained from voice parsing is combined with the "invoice date" and "invoice number" obtained from document parsing to form complete structured data.

[0163] S205. Determine the current business scenario based on the business context corresponding to the intent tag and / or business source data, and load the corresponding target intent tag and business form association knowledge base according to the current business scenario.

[0164] For example, the intent label is "R&D Expense Declaration". Based on the user's identity (R&D department employee) and business context (uploaded technical service fee invoice), the current business scenario is determined to be "Automotive R&D Project Expense Declaration Scenario". The system loads the knowledge base corresponding to this scenario. This knowledge base stores the mapping relationship between intent labels and business forms in R&D scenarios. For example, the intent "External Technical Service Fee" maps to "R&D Project External Technical Service Fee Declaration Form", the intent "Materials Fee" maps to "R&D Project Materials Requisition Form", and the intent "Travel Expenses" maps to "R&D Project Travel Expense Reimbursement Form", etc.

[0165] S206. Based on the loaded knowledge base that associates target intent tags with business forms, determine the target business form corresponding to the intent tag.

[0166] For example, the target business form is the "Application Form for External Technical Service Fees for R&D Projects". This form has the following pre-set fields: Project Name, Service Item, Service Provider Name, Contract Amount, Invoice Number, Invoice Date, Application Amount, and Remarks.

[0167] S207. Map the business entity data to the corresponding fields in the target business form to obtain the business form to be validated.

[0168] For example, the system maps each business entity in the structured data to the corresponding field in the target business form. After the mapping is completed, the system generates a business form to be validated.

[0169] S208. Based on the rule engine, perform data validity checks on the business forms to be validated and obtain the validation results.

[0170] For example, the rule engine has pre-defined verification rules for R&D expense declarations, including invoice number format verification, amount logic verification, and mandatory field verification. For instance, the declared amount cannot exceed the contract amount; the invoice date cannot be later than the declaration date; the invoice number must be 12 digits; the difference between the declared amount and the invoice amount cannot exceed 1 yuan, etc. The rule engine checks each item on the form to be verified: the declared amount of 50,000 yuan matches the invoice amount of 50,000 yuan, so it passes; the invoice date of March 15, 2025 is earlier than the current declaration date of March 20, 2025, so it passes; the invoice number 12345678 is 8 digits, which does not conform to the 12-digit rule, triggering a verification failure. The system generates a verification result, prompting "Invoice number format is incorrect, please check and re-enter."

[0171] S209. When the verification result indicates that the verification has passed, the business form to be verified shall be used as the business form after being filled in.

[0172] S210. Generate a preview form for the filled-in business form, obtain the operation instructions entered by the user based on the preview form, and perform the corresponding operation according to the operation instructions.

[0173] Specifically, in response to a confirmation command, the completed business form is submitted to the downstream business system; in response to an edit command, the form is entered into edit mode, and a preview form is regenerated after editing; in response to an abandon command, the current form filling process is terminated.

[0174] S211. Based on the preset data flow rules, determine whether the business form after being filled in meets the transfer conditions corresponding to the current node.

[0175] For example, the system is pre-configured with approval data flow rules for R&D expense applications. The current process node is "R&D personnel submit application". This node has two transfer paths: Path 1's transfer condition is "application amount less than or equal to 20,000 yuan", which automatically submits to the "department head approval" node when met; Path 2's transfer condition is "application amount greater than 20,000 yuan", which automatically submits to both the "department head approval" and "R&D director approval" nodes when met. In this example, the application amount is 50,000 yuan, which is greater than 20,000 yuan, so the system determines that the transfer condition for Path 2 is met.

[0176] If the transfer condition is met, then execute S212; if the transfer condition is not met, then execute S213.

[0177] S212. Submit the completed business form to the next process node corresponding to the transfer condition.

[0178] For example, the system automatically submits the "Application Form for External Technical Service Fees for R&D Projects" filled out by Zhang to the "Department Head Approval" and "R&D Director Approval" nodes, generates an approval task in the to-do list of the corresponding approver, and sends an approval notification message at the same time.

[0179] S213. Generate a prompt message and pause automatic triggering.

[0180] Through the above steps, this embodiment takes the application for external technical service fees by R&D personnel of car companies as a complete example, demonstrating the end-to-end intelligent filling process from multi-source data collection, semantic parsing, cross-validation, form mapping, rule verification, user confirmation to automatic process triggering, realizing the automated processing of R&D expense application forms and the intelligent connection of approval processes.

[0181] Figure 3This is a schematic diagram of the structure of the AI-based business form filling device provided in the embodiments of this application, as shown below. Figure 3 As shown, the AI-based business form filling device 30 provided in this embodiment includes:

[0182] The acquisition module 31 is used to acquire the business source data required for filling in the business form and to determine the modal type of the business source data. The business source data includes voice data and / or document data.

[0183] Processing module 32 is used to apply the information extraction strategy corresponding to the modality type to convert the business source data into unstructured text data;

[0184] The semantic parsing module 33 is used to perform deep semantic parsing on unstructured text data using a semantic parsing model to obtain structured data that conforms to the business form field value standards. The structured data includes business entity data and intent tags.

[0185] Module 34 is used to determine the target business form corresponding to the intent tag based on the knowledge base associated with the intent tag and the business form.

[0186] The data entry module 35 is used to map business entity data to the corresponding fields in the target business form to obtain the filled business form.

[0187] In one possible implementation, the semantic parsing module 33 is specifically used to: input unstructured text data into a semantic parsing model, perform entity extraction and intent classification through the semantic parsing model to identify business entity data and intent tags in the unstructured text data; the semantic parsing model is a natural language understanding model finely tuned for the target business domain; perform normalization transformation on the business entity data to obtain normalized business entity data, the normalization transformation including: semantic normalization processing to convert colloquial expressions into standard written expressions, and numerical formatting processing to convert non-standard numerical expressions into standard numerical formats; associate the normalized business entity data with intent tags to obtain structured data.

[0188] In one possible implementation, the processing module 32 is further configured to: generate a preview form for the filled-in business form; obtain user input based on the preview form, the operation instructions including confirm, edit, or abandon; in response to the confirm instruction, submit the filled-in business form to the downstream business system; in response to the edit instruction, enter the form editing state and regenerate the preview form after editing; in response to the abandon instruction, terminate the current filling process; or, according to preset data flow rules, determine whether the filled-in business form meets the transfer conditions corresponding to the current node, the data flow rules including at least one process node and transfer conditions between nodes; if the transfer conditions are met, submit the filled-in business form to the next process node corresponding to the transfer conditions; if the transfer conditions are not met, generate a prompt message and pause automatic triggering.

[0189] In one possible implementation, the filling module 35 is specifically used to: map business entity data to corresponding fields in the target business form to obtain the business form to be verified; perform data rationality verification on the business form to be verified based on the rule engine to obtain the verification result; and when the verification result indicates that the verification has passed, use the business form to be verified as the filled business form.

[0190] In one possible implementation, the determining module 34 is specifically used to: determine the current business scenario based on the business context corresponding to the intent tag and / or business source data; load the corresponding target intent tag and business form association knowledge base based on the current business scenario; and determine the target business form corresponding to the intent tag based on the target intent tag and business form association knowledge base.

[0191] In one possible implementation, the knowledge base for associating intent tags with business forms is dynamically updated in the following way: feedback data is collected during the historical data entry process; based on the feedback data, a machine learning algorithm is used to adjust the mapping relationship between intent tags and business forms in the knowledge base; wherein, the adjustment includes at least one of the following: increasing or decreasing the confidence weight of existing mapping entries, adding mapping entries between intent tags and business forms, deleting mapping entries that meet preset elimination conditions, and merging or splitting semantically similar intent tags.

[0192] In one possible implementation, when the business source data includes both voice data and document data, the semantic parsing module 33 is further configured to: employ a semantic parsing model to perform deep semantic parsing on the unstructured text data corresponding to the voice data to obtain first intermediate structured data that conforms to the business form field value standards; employ a semantic parsing model to perform deep semantic parsing on the unstructured text data corresponding to the document data to obtain second intermediate structured data that conforms to the business form field value standards; perform cross-validation on the first intermediate structured data and the second intermediate structured data, the cross-validation including aligning the first intermediate structured data and the second intermediate structured data according to the business entity fields, and comparing the values ​​of the same business entity fields and their corresponding confidence scores; and select the data source with the higher confidence score from the first intermediate structured data and the second intermediate structured data according to a preset confidence strategy to obtain structured data that conforms to the business form field value standards.

[0193] In one possible implementation, the processing module 32 is further configured to: when the modality type is speech modality, use streaming speech recognition technology deployed on edge computing nodes to perform local recognition processing on the business source data and convert the business source data into unstructured text data in real time; when the modality type is document modality, use document recognition and document understanding technology to parse the business source data, extract the text content and structured information in the business source data, and generate unstructured text data.

[0194] The AI-based business form filling device provided in this embodiment can execute the methods provided in the above method embodiments. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0195] Figure 4 A schematic diagram of the structure of the electronic device provided in this application. Figure 4 As shown, the electronic device 40 provided in this embodiment includes at least one processor 401 and a memory 402. Optionally, the electronic device 40 further includes a communication component 403. The processor 401, memory 402, and communication component 403 are connected via a bus 404.

[0196] In a specific implementation, at least one processor 401 executes computer execution instructions stored in memory 402, causing at least one processor 401 to perform the above-described method.

[0197] The specific implementation process of processor 401 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0198] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0199] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0200] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0201] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0202] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0203] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0204] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0205] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0206] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0207] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0208] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0209] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0210] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A business form filling method based on AI, characterized in that, include: Obtain the business source data required for filling out the business form, and determine the modal type of the business source data, which includes voice data and / or document data; The information extraction strategy corresponding to the modality type is applied to convert the business source data into unstructured text data; A semantic parsing model is used to perform deep semantic parsing on the unstructured text data to obtain structured data that conforms to the business form field value standards. The structured data includes business entity data and intent tags. Based on the knowledge base that associates intent tags with business forms, the target business form corresponding to the intent tag is determined; The business entity data is mapped to the corresponding fields in the target business form to obtain the filled-in business form.

2. The AI-based business form filling method according to claim 1, characterized in that, The semantic parsing model is used to perform deep semantic parsing on the unstructured text data to obtain structured data that conforms to the business form field value standards, including: The unstructured text data is input into the semantic parsing model, which performs entity extraction and intent classification to identify business entity data and intent tags in the unstructured text data; the semantic parsing model is a natural language understanding model finely tuned for the target business domain. The business entity data is normalized to obtain normalized business entity data. The normalization includes: semantic normalization processing to convert colloquial expressions into standard written expressions, and numerical formatting processing to convert non-standard numerical expressions into standard numerical formats. The standardized business entity data is associated with the intent tag to obtain the structured data.

3. The AI-based business form filling method according to claim 1, characterized in that, Following the completed business form, the following is also included: Generate a preview form for the filled-in business form; obtain the user's operation instructions based on the preview form, the operation instructions including confirm, edit, or abandon; in response to the confirm instruction, submit the filled-in business form to the downstream business system; in response to the edit instruction, enter the form editing state, and regenerate the preview form after editing is completed; in response to the abandon instruction, terminate the current filling process; Alternatively, based on preset data flow rules, determine whether the filled-in business form meets the transfer conditions corresponding to the current node. The data flow rules include at least one process node and transfer conditions between nodes. If the transfer conditions are met, submit the filled-in business form to the next process node corresponding to the transfer conditions. If the transfer conditions are not met, generate a prompt message and pause automatic triggering.

4. The AI-based business form filling method according to any one of claims 1 to 3, characterized in that, The step of mapping the business entity data to the corresponding fields in the target business form to obtain the filled-in business form includes: The business entity data is mapped to the corresponding fields in the target business form to obtain the business form to be verified. Based on the rule engine, the data rationality of the business form to be verified is validated to obtain the verification result. When the verification result indicates that the verification is successful, the business form to be verified is used as the filled-in business form.

5. The AI-based business form filling method according to any one of claims 1 to 3, characterized in that, The step of determining the target business form corresponding to the intent tag based on the knowledge base associated with the intent tag and the business form includes: The current business scenario is determined based on the intent tag and / or the business context corresponding to the business source data; Based on the current business scenario, load the corresponding target intent tag and business form association knowledge base; Based on the knowledge base that associates the target intent tag with the business form, the target business form corresponding to the intent tag is determined.

6. The AI-based business form filling method according to any one of claims 1 to 3, characterized in that, The knowledge base linking intent tags and business forms is dynamically updated in the following ways: Collect feedback data from the historical data entry process; Based on the feedback data, a machine learning algorithm is used to adjust the mapping relationship between intent tags and business forms in the knowledge base; The adjustments include at least one of the following: increasing or decreasing the confidence weight of existing mapping entries, adding mapping entries between intent tags and business forms, deleting mapping entries that meet preset elimination conditions, and merging or splitting semantically similar intent tags.

7. The AI-based business form filling method according to any one of claims 1 to 3, characterized in that, When the business source data includes both voice data and document data, a semantic parsing model is used to perform deep semantic parsing on the unstructured text data to obtain structured data that conforms to the business form field value standards, including: Using the semantic parsing model, deep semantic parsing is performed on the unstructured text data corresponding to the voice data to obtain the first intermediate structured data that conforms to the business form field value standards; Using the semantic parsing model, deep semantic parsing is performed on the unstructured text data corresponding to the document data to obtain second intermediate structured data that conforms to the business form field value standards; Cross-validation is performed on the first intermediate structured data and the second intermediate structured data. The cross-validation includes aligning the first intermediate structured data and the second intermediate structured data according to the business entity field, and comparing the values ​​of the same business entity field and the corresponding confidence scores. According to the preset confidence strategy, the data source with the higher confidence score is selected from the first intermediate structured data and the second intermediate structured data to obtain the structured data that meets the business form field value standards.

8. The AI-based business form filling method according to any one of claims 1 to 3, characterized in that, The information extraction strategy corresponding to the application modality type converts the business source data into unstructured text data, including: When the modality type is speech modality, streaming speech recognition technology deployed on edge computing nodes is used to perform local recognition processing on the business source data and convert the business source data into unstructured text data in real time. When the modality type is document modality, document recognition and document understanding technologies are used to parse the business source data, extract the text content and structured information in the business source data, and generate unstructured text data.

9. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes the computer execution instructions stored in the memory, causing the processor to perform the AI-based business form filling method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, are used to implement the AI-based business form filling method as described in any one of claims 1 to 8.

Images