System for automatically generating writing instructions based on field context analysis in a non-integrated template environment

By analyzing template files for contextual and structural clues, the system generates writing instructions for input fields, addressing the challenge of non-integrated environments and enhancing the accuracy and efficiency of electronic record creation.

KR103044472B1Active Publication Date: 2026-11-06PUZZLE AI CO LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
KR1020260086143
Authority / Receiving Office
KR · KR
Patent Type
Patents
Current Assignee / Owner
Filing Date
2026-05-13
Publication Date
2026-11-06
Estimated Expiration
2046-05-13

AI Technical Summary

Technical Problem

Conventional technologies struggle to automatically determine the meaning and input conditions of input fields in electronic record templates without direct access to internal metadata, leading to inaccurate structured output generation by generative AI, especially in non-integrated environments.

Method used

A system that analyzes template files to identify input target fields and generates writing instructions based on contextual and structural information, such as hierarchical relationships, positional relationships, and semantic clues within the template, without relying on application program interfaces.

Benefits of technology

Enables accurate and flexible generation of structured output by generative AI, reducing manual mapping tasks and improving the accuracy and efficiency of electronic record creation across different institutions and medical departments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 112026057789312-PAT00001_ABST
    Figure 112026057789312-PAT00001_ABST
Patent Text Reader

Abstract

The present invention relates to a system for automatically generating writing instructions based on field context analysis in a non-linked template environment. The problem to be solved is to analyze the structural and contextual information of the template file itself without linking with an application program interface with an external electronic record system to determine the meaning and input conditions of the input target fields, and to automatically generate field-specific writing instructions and structured output data based thereon. For example, a system for automatically generating instructions based on field context analysis in a non-linked template environment is disclosed, comprising: a template field identification unit that analyzes a template file used in a non-linked template environment to identify an input target field and generates field structure information corresponding to the identified input target field; a field context extraction unit that extracts context information for field interpretation corresponding to the input target field within the template file based on the field structure information generated by the template field identification unit; and an instruction generation unit that determines the input meaning and input conditions of the input target field based on the context information for field interpretation extracted by the field context extraction unit, and generates a writing instruction corresponding to the input target field based on the determined input meaning and input conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technology Field

[0001] Embodiments of the present invention relate to the fields of electronic record systems, generative artificial intelligence, and automatic template analysis technology. More specifically, the invention relates to a system for automatically generating writing instructions based on field context analysis in a non-linked template environment, which analyzes structural and contextual information of input fields included in a template file itself without linking with an application program interface with an external electronic record system, and automatically generates writing instructions corresponding to each input field based on the analyzed field context. Background Technology

[0002] Recently, various institutions, including medical facilities, have been widely utilizing electronic record systems to electronically create and manage patient information, medical records, consultation records, test results, prescription information, and administrative records. In particular, the medical field manages various records generated during the patient's treatment process in the form of electronic documents through electronic medical record systems or electronic health record systems; within these systems, different types of templates are registered and used depending on the medical department, institution, type of work, or purpose of the record.

[0003] These templates can generally be created based on markup languages ​​or screen composition languages ​​such as XML, HTML, or CS, and may include multiple input fields where users can directly enter values ​​or external input values ​​can be inserted. For example, a medical record template may include input fields corresponding to patient name, date of birth, gender, reason for visit, chief complaint, medical history, vital signs, diagnosis, prescription details, progress notes, etc., and each input field may be defined according to the template's document structure, screen layout, data binding structure, or internal identifier.

[0004] Meanwhile, with the recent advancements in speech recognition technology and large-scale language model-based generative AI, technologies are emerging in which a speech recognition engine converts medical consultation details spoken by medical staff into text data, and a generative AI interprets this data to generate structured output values ​​that fit the templates of electronic medical record systems. This technology has high potential for application in clinical settings, as it reduces the burden of manual data entry for medical staff and enables the automatic creation of medical consultation details in natural language format according to electronic record templates.

[0005] However, for generative AI to generate structured output suitable for electronic record templates, field-specific guidelines are required regarding the meaning of each input field, the data format in which values ​​should be entered, the range of allowed values, and the required format. For example, the meaning and format of the output value that the generative AI must generate may vary depending on whether a single input field is intended to record the patient's chief complaint, medical history, systolic blood pressure, or the physician's comprehensive opinion.

[0006] Conventionally, to understand the meaning of these input fields, methods involving querying field names, data types, allowed values, mandatory input status, coding schemes, and metadata through application program interfaces provided by electronic record systems could be considered. However, not all electronic record systems provide sufficient application program interfaces for external systems, and even when such interfaces are provided, there are issues where external systems find it difficult to freely access field metadata due to security policies, access rights, patient information protection regulations, system closed systems, and vendor-specific data structure differences.

[0007] In particular, since electronic record systems in the medical field handle sensitive patient and medical information, direct access by external generative AI systems or voice recognition-based record assistance systems to internal field definition information may be restricted. Consequently, in actual operating environments, external systems often cannot deeply integrate with the electronic record system and must implement automatic recording functions by utilizing only the template files or screen configuration information used within the system in a limited manner.

[0008] However, in non-integrated environments where only template files are provided, each input field is often displayed using identifiers arbitrarily assigned by internal developers, numerical identifiers, abbreviation-based identifiers, or internal system identifiers. For example, if an input field is defined solely by an arbitrary code or an opaque ID value rather than a field name that intuitively conveys a specific meaning, it is difficult for external systems to directly determine what content the field is actually intended to receive.

[0009] In addition, even input fields with the same meaning may have different field names, identification values, layout structures, attribute values, or display methods depending on the hospital, medical department, or template. As a result, it is difficult to universally process various electronic record templates using only a single fixed mapping table or predefined rules, and there is a problem in that separate manual mapping or customization work is required whenever a new template is added or the structure of an existing template is changed.

[0010] Meanwhile, template files may contain various clues that allow the meaning of input fields to be inferred. For example, labels surrounding the input field, guidance text, input guidance attributes, accessibility attributes, comments, data attributes, validation attributes, the title of the parent area, sibling fields within the same area, hierarchical relationships between objects, positional relationships on the screen, and semantic units included in field identifiers or field name attributes can be utilized to determine the purpose and input conditions of the input field.

[0011] For example, if a specific input field is placed within an area related to vital signs and display text or sibling fields related to systolic blood pressure, diastolic blood pressure, and heart rate exist nearby, that input field can be inferred as a field for receiving specific vital sign measurements. Additionally, if a specific input field is located within the patient basic information area and labels related to name, date of birth, and gender are placed nearby, that input field can be interpreted as a field for receiving values ​​related to patient basic information.

[0012] Nevertheless, conventional technology had limitations in automatically inferring the meaning of input fields and input requirements by comprehensively analyzing these structural, contextual, and semantic clues contained within the template file itself. Conventional automatic input technologies often relied primarily on explicit field names, pre-configured mapping information, or metadata provided by electronic record systems, and failed to adequately provide a system for generating field-specific writing guidelines by organically utilizing surrounding context, hierarchical structures, relationships between adjacent fields, annotations, and attribute information present within the template file.

[0013] Furthermore, even when generating structured output using generative AI, output quality may degrade if clear instructions are not provided for each input field. Although generative AI can extract various information from natural language record data, if the meaning and input conditions of each input field are unclear, it may duplicate the same content into different fields, generate content that does not match the purpose of the field, or output values ​​that do not conform to the required data format.

[0014] Furthermore, medical record data generated by speech recognition processing may contain natural language expressions, omitted context, irregularities in utterance order, medical abbreviations, numerical expressions, and negative expressions. Therefore, simply mapping speech recognition results sequentially to the input fields of a template makes it difficult to achieve accurate structured output; instead, field-specific writing instructions are required to specify what content to extract for each input field, in what format values ​​to generate, and according to what writing standards to record them.

[0015] As such, conventional technology fails to adequately provide a system for automatically determining the meaning and input conditions of target input fields using structural, contextual, and semantic information inherent in the template file itself, and converting this into writing instructions usable for generating structured outputs of generative artificial intelligence, in non-integrated template environments where application program interface integration with external electronic record systems is limited. Therefore, there is a need for technology capable of automatically generating writing instructions for each target input field by analyzing the template file itself, even without deep integration with electronic record systems. Prior art literature

[0016] Published Patent Application No. 10-2025-0106150 (Publication Date: July 9, 2025) Published Patent Application No. 10-2023-0006240 (Publication Date: January 10, 2023) Registered Patent Application No. 10-2827595 (Registration Date: June 30, 2025) Registered Patent Application No. 10-2728657 (Registration Date: November 13, 2024) The problem to be solved

[0017] An embodiment of the present invention provides an automatic instruction generation system based on field context analysis in a non-linked template environment, which analyzes a template file used in a non-linked template environment without linking an application program interface with an external electronic record system to identify an input target field, extracts context information for field interpretation corresponding to the identified input target field, and determines the input meaning and input conditions of the input target field based on the extracted context information for field interpretation.

[0018] In addition, even if the input target field is defined by an opaque field identifier, the present invention provides a system for automatically generating writing instructions based on field context analysis in a non-linked template environment, which infers the meaning and input conditions of the input target field by utilizing at least one of hierarchical relationships, positional relationships, surrounding text, input guide attributes, accessibility attributes, comment information, and semantic unit information within the template file, and automatically generates writing instructions for each input target field based on the inference results.

[0019] In addition, by linking generated writing instructions with template schema information and generating structured output data for each input target field from speech recognition record data or natural language-based input source data, it provides an automatic writing instruction generation system based on field context analysis in a non-linked template environment applicable to various electronic record templates, thereby reducing the user's manual field mapping work. means of solving the problem

[0020] A system for automatically generating instructions based on field context analysis in a non-linked template environment according to an embodiment of the present invention comprises: a template field identification unit that analyzes a template file used in a non-linked template environment to identify an input target field and generates field structure information corresponding to the identified input target field; a field context extraction unit that extracts context information for field interpretation corresponding to the input target field within the template file based on the field structure information generated by the template field identification unit; and an instruction generation unit that determines the input meaning and input conditions of the input target field based on the context information for field interpretation extracted by the field context extraction unit, and generates a writing instruction corresponding to the input target field based on the determined input meaning and input conditions.

[0021] Additionally, the template field identification unit may include: a template file acquisition unit that acquires a template file used in a non-linked template environment or template access information for accessing said template file; a template structure conversion unit that converts said template file into hierarchical structure data used for field identification based on the file format and document structure of the template file acquired by the template file acquisition unit; an input field detection unit that detects an inputable object from the hierarchical structure data converted by the template structure conversion unit, into which a user input value can be inserted or mapped to an external input value, and identifies it as an input target field; and a field structure information generation unit that generates field structure information corresponding to said input target field by matching at least one of the field identification value, field location information, field attribute information, and inter-field relationship information of the input target field identified by the input field detection unit.

[0022] Additionally, the field context extraction unit may include: a context range setting unit that sets a context extraction range corresponding to the input target field within the template file based on field structure information generated by the template field identification unit; a context element extraction unit that extracts context element information used for interpretation of the input target field from the context extraction range set by the context range setting unit; and a context information configuration unit that corresponds the context element information extracted by the context element extraction unit with the input target field to form context information for field interpretation.

[0023] Additionally, the context range setting unit may set the context extraction range as a range that includes at least one of a hierarchically related object, a positionally adjacent object, and an attributeally connected object based on the input target field, based on at least one of the field location information and inter-field relationship information included in the field structure information generated by the template field identification unit.

[0024] Additionally, the context element extraction unit can extract at least one of the hierarchical relationships between objects included in the context extraction range set by the context range setting unit, relationships between surrounding input objects, belonging area information, display text, input guide attributes, accessibility attributes, comment information, semantic unit tokens included in the field identification value, and semantic unit tokens included in the field attribute information, and convert the extracted context element information into semantic candidate information used for determining field semantics.

[0025] Additionally, the instruction generation unit may include: a context input data generation unit that generates context input data by sorting context information for field interpretation extracted by the field context extraction unit according to the input target field and configuring context element information included in the sorted context information for field interpretation into an input format used for determining field meaning; an inference request data generation unit that converts the context input data generated by the context input data generation unit into inference request data that can be input into a language model; an input condition determination unit that determines an input condition including at least one of the input meaning, input format, input range, whether input is required, and writing criteria of the input target field based on the inference request data generated by the inference request data generation unit; and a field-specific instruction generation unit that generates a writing instruction including at least one of a field meaning description, data format guidelines, value generation criteria, and writing guide corresponding to the field identification value of the input target field based on the input meaning and input condition determined by the input condition determination unit.

[0026] In addition, it may further include a structured output generation unit that generates structured output data corresponding to the input target field from the input source data using the writing instructions generated by the instruction generation unit.

[0027] Additionally, the structured output generation unit may include: a template schema generation unit that generates template schema information including at least one of a field identification value, input meaning, input condition, and a writing instruction for each input target field by correlating a writing instruction generated by the instruction generation unit with field structure information generated by the template field identification unit; an input source data acquisition unit that acquires input source data to be processed together with the template schema information generated by the template schema generation unit; a field value generation unit that generates an output value for each input target field from the input source data acquired by the input source data acquisition unit based on the writing instruction included in the template schema information generated by the template schema generation unit; and a structured data generation unit that generates structured output data by correlating the output value for each input target field generated by the field value generation unit with the field identification value of the input target field.

[0028] Additionally, the input source data acquisition unit acquires record data generated by speech recognition processing as the input source data, and the field value generation unit may include: a record data division unit that divides the record data acquired by the input source data acquisition unit into utterance content units or sentence units to generate record unit data; a field correspondence relationship calculation unit that calculates a correspondence relationship between the record unit data and the input target field by comparing the record unit data generated by the record data division unit with the writing instructions included in the template schema information generated by the template schema generation unit; and an output value generation unit that generates an output value for each input target field from the record unit data based on the correspondence relationship calculated by the field correspondence relationship calculation unit.

[0029] Additionally, it may further include a field semantics correction unit for correcting the accuracy of context information for field interpretation extracted from the field context extraction unit or the creation instruction generated from the instruction generation unit.

[0030] Additionally, the field semantic correction unit may include: a semantic candidate value acquisition unit that acquires semantic candidate values ​​used for determining the semantics of an input target field from context information for field interpretation extracted by the field context extraction unit; a domain semantic matching unit that calculates domain semantic information corresponding to semantic candidate values ​​by comparing the semantic candidate values ​​acquired by the semantic candidate value acquisition unit with pre-set domain term information; a semantic consistency judgment unit that determines the semantic consistency of the input target field by comparing the domain semantic information calculated by the domain semantic matching unit with field-to-field relationship information included in the context information for field interpretation extracted by the field context extraction unit; and a field interpretation correction unit that corrects at least one of the input semantics, input conditions, and writing instructions of the input target field according to the judgment result of the semantic consistency judgment unit.

[0031] In addition, the semantic consistency determination unit can compare the domain semantic information calculated by the domain semantic matching unit with the field relationship information included in the field interpretation context information extracted by the field context extraction unit to determine at least one of the arrangement relationship of input semantics, the mutual correspondence relationship of input conditions, and whether there is duplication of field semantics among a plurality of input target fields within the same context extraction range. Effects of the invention

[0032] According to the present invention, since input target fields can be identified by analyzing the template file itself without linking with an application program interface with an external electronic record system and context information for field interpretation for each input target field can be extracted, template-based writing instructions can be automatically generated even in environments where it is difficult to directly access the internal field metadata of the electronic record system.

[0033] In addition, since the input meaning and input conditions of the input target field can be determined by utilizing not only the field identification value of the input target field but also at least one of the hierarchical relationships, positional relationships, surrounding text, input guide attributes, accessibility attributes, comment information, and semantic unit information within the template file, the field meaning that is difficult to grasp with only the opaque field identification value can be interpreted more accurately.

[0034] In addition, writing instructions for each input target field are automatically generated based on the determined input meaning and input conditions, and the generated writing instructions can be used to generate structured output, thereby inducing the generative AI to generate output values ​​that match the purpose and format of the input target fields.

[0035] In addition, since it reduces reliance on manual mapping tasks by template or integration with separate application interfaces for each electronic record system, the generation of writing instructions and structured output can be flexibly applied even in environments where template structures differ by institution, task, or medical department.

[0036] In addition, since record data generated by speech recognition processing or natural language-based input source data can be converted into structured output data by linking it with writing instructions for each input target field, the burden of manual input by the user during the electronic record creation process can be reduced, and the accuracy, consistency, and processing efficiency of template-based record creation can be improved. Brief explanation of the drawing

[0037] FIG. 1 is a schematic diagram showing the overall hardware-based operation process of a system for automatically generating instructions based on field context analysis in a non-linked template environment according to an embodiment of the present invention. FIG. 2 is a block diagram showing the overall configuration of a system for automatically generating instructions based on field context analysis in a non-linked template environment according to an embodiment of the present invention. FIG. 3 is a block diagram showing the detailed configuration of a template field identification unit according to an embodiment of the present invention. FIG. 4 is a block diagram showing the detailed configuration of a field context extraction unit according to an embodiment of the present invention. FIG. 5 is a block diagram showing the detailed configuration of an instruction generation unit according to an embodiment of the present invention. FIG. 6 is a block diagram showing the detailed configuration of a field meaning correction unit according to an embodiment of the present invention. FIG. 7 is a block diagram showing the detailed configuration of a structured output generation unit according to an embodiment of the present invention. FIG. 8 is a block diagram showing the detailed configuration of a field value generation unit according to an embodiment of the present invention. Specific details for implementing the invention

[0038] The terms used in this specification will be briefly explained, and the invention will be described in detail.

[0039] The terms used in this invention have been selected based on currently widely used general terms, taking into account their functions within the invention; however, these terms may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been arbitrarily selected by the applicant, and in such cases, their meanings will be described in detail in the relevant description of the invention. Therefore, the terms used in this invention should be defined not merely by their names, but based on their meanings and the overall content of the invention.

[0040] When a part of a specification is described as "comprising" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as "...part" or "module" as used in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware or software, or as a combination of hardware and software.

[0041] Embodiments of the present invention are described below with reference to the attached drawings so that those skilled in the art can easily implement them. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0042] FIG. 1 is a schematic diagram showing the overall hardware-based operation process of a system for automatically generating instructions based on field context analysis in a non-linked template environment according to an embodiment of the present invention.

[0043] Referring to FIG. 1, the automatic creation instruction generation system (1000) based on field context analysis in a non-linked template environment according to the present embodiment can be implemented by including a user terminal (10), an external electronic record system server (20), a template database (30), and an artificial intelligence processing server (40), and can form a non-linked processing flow that identifies an input target field using the template file itself without application program interface linkage between the external electronic record system server (20) and the artificial intelligence processing server (40), extracts context information for field interpretation corresponding to the input target field, and automatically generates a creation instruction for each input target field based on the extracted context information for field interpretation.

[0044] The above user terminal (10) may be implemented as an input / output device for an electronic record creator or electronic record user to view a template-based record creation screen and perform necessary input or verification actions. For example, the user terminal (10) may be implemented as at least one of a desktop computer, a laptop computer, a tablet terminal, a medical workstation, or a terminal device on which an electronic record creation application is installed. The user terminal (10) may display a template creation screen provided through an external electronic record system server (20) and may be used as an input interface for the user to create or verify medical records, consultation records, examination records, or work records. Additionally, the user terminal (10) may be used as an input means for converting the user's speech content into record data by linking with a voice input device such as a microphone or a voice recognition module, and may display to the user the process of creation instructions or structured output data generated by the artificial intelligence processing server (40) being reflected in the input target fields of the template creation screen.

[0045] Additionally, the user terminal (10) may provide an overlay interface so that the user can check at least one of the input meaning, input condition, creation instruction, buffer text, or structured output value of the input target field. The overlay interface may be implemented as a guide area that is superimposed on the existing template creation screen without changing it, and the display position may be adjusted based on at least one of the screen position of the input target field, the screen scroll position, the size of the display area, whether it overlaps with surrounding objects, and the user pointer position. For example, if automatic injection for a specific input target field fails or the input target field is not in a state where it can be entered immediately, the user terminal (10) may display buffer text, creation instruction, and re-injection guide information near the input target field, and may allow the output value to be reflected back into the input target field according to the user's confirmation input, modification input, or re-injection command.

[0046] The external electronic record system server (20) may be implemented as a server device that provides a template-based electronic record creation environment. For example, the external electronic record system server (20) may be implemented as at least one of an electronic medical record system server, an electronic health record system server, a consultation record management server, a work record management server, or an internal electronic document management server of an institution. This external electronic record system server (20) may provide a record creation screen, a document selection screen, or a template usage environment to a user terminal (10), and may retrieve a template file corresponding to a document type, medical department type, work type, or record type selected by the user terminal (10) from a template database (30). Additionally, the external electronic record system server (20) may enable the AI ​​processing server (40) to analyze the template file itself by providing a template file or template access information for accessing the template file, even if it does not directly provide internal field metadata to the AI ​​processing server (40) via an application program interface.

[0047] The above template database (30) may be implemented as a data storage device that stores and provides template files used by an external electronic record system server (20). For example, the template database (30) may be implemented as at least one of a relational database server, a file storage server, an object storage, a document management database, a cloud-based storage, or an internal storage connected to an external electronic record system server (20). This template database (30) may store template files written in at least one format among XML, HTML, CS, markup documents, screen configuration files, or data binding definition files. Additionally, the template database (30) may store at least one of a template identification value, document type information, creation purpose information, version information, screen configuration information, or field attribute information corresponding to the template file, and may provide template files or template access information upon request from the external electronic record system server (20).

[0048] The artificial intelligence processing server (40) described above may be implemented as a central processing unit that analyzes a template file to generate writing instructions for each input target field. For example, the artificial intelligence processing server (40) may be implemented as at least one of a server computer including one or more processors and memory, a cloud-based computing instance, a GPU-based artificial intelligence computing server, a distributed processing server cluster, or a computing module called in the form of a software development kit. This artificial intelligence processing server (40) may receive a template file or template access information provided from an external electronic record system server (20) or a template database (30), and may determine the input meaning and input conditions of the input target field by analyzing the document structure, object attributes, surrounding context, and semantic clues included in the template file, even without querying the internal field metadata of the external electronic record system server (20) in an application program interface manner. Additionally, the artificial intelligence processing server (40) may generate writing instructions corresponding to the input target field based on the determined input meaning and input conditions, and may generate structured output data for each input target field from record data generated by speech recognition processing or natural language-based input source data as needed.

[0049] The overall operation process of the automatic instruction generation system (1000) based on field context analysis according to the present embodiment is as follows.

[0050] First, the user can access the external electronic record system server (20) using the user terminal (10), and the external electronic record system server (20) can provide the user terminal (10) with a non-linked template creation environment or a record creation environment. At this time, the user can select at least one of a document type, a medical department type, a work type, or a record type through the user terminal (10), and the external electronic record system server (20) can transmit a template file request or a query request to the template database (30) to specify a template file corresponding to the user's selection.

[0051] Subsequently, the template database (30) may provide the external electronic record system server (20) with a template file corresponding to a selected record type or template access information for accessing the template file in response to a request from the external electronic record system server (20). At this time, the template file may be configured in at least one format among XML, HTML, CS, a markup document, a screen configuration file, or a data binding definition file, and the template file may include at least one of an input object into which a user input value can be inserted, a data binding object that can be mapped to an external input value, a field identification value, display text, an input guide attribute, an accessibility attribute, comment information, or field attribute information.

[0052] Next, the external electronic record system server (20) may provide the template file or template access information to the artificial intelligence processing server (40) without directly providing the internal field metadata of the external electronic record system server (20) via an application program interface. Accordingly, the artificial intelligence processing server (40) can analyze the template file in a non-linked manner that does not directly link to the internal data structure of the external electronic record system server (20), and can generate writing instructions for each input target field using structural information and contextual information inherent in the template file itself.

[0053] Subsequently, the artificial intelligence processing server (40) can check the file format and document structure of the provided template file and convert the template file into analyzable hierarchical structure data. Additionally, the artificial intelligence processing server (40) can detect inputtable objects in the converted hierarchical structure data into which user input values ​​can be inserted or mapped to external input values, and can identify the detected inputtable objects as input target fields. Furthermore, the artificial intelligence processing server (40) can generate field structure information corresponding to the input target field by matching at least one of the field identification value, field location information, field attribute information, and inter-field relationship information of the input target field.

[0054] Next, the artificial intelligence processing server (40) can set a context extraction range corresponding to the input target field within the template file based on the generated field structure information. At this time, the context extraction range may be set to include at least one of a hierarchically related object, a positionally adjacent object, or an attributeally connected object based on the input target field. For example, the parent area of ​​the input target field, a sibling field within the same container, display text connected to the input target field, an input guide attribute, or an accessibility attribute may be included in the context extraction range.

[0055] Subsequently, the artificial intelligence processing server (40) can extract context element information used for interpreting the input target field from the set context extraction range. For example, the artificial intelligence processing server (40) can extract at least one of the following: hierarchical relationships between objects, relationships between surrounding input objects, belonging area information, display text, input guide attributes, accessibility attributes, comment information, semantic unit tokens included in the field identification value, or semantic unit tokens included in the field attribute information. Additionally, the artificial intelligence processing server (40) can correspond the extracted context element information with the input target field to form context information for field interpretation.

[0056] Next, the artificial intelligence processing server (40) sorts context information for field interpretation by input target field and converts the sorted context information for field interpretation into inference request data that can be input into a language model. Subsequently, the artificial intelligence processing server (40) can determine the input meaning and input conditions of the input target field based on the inference request data. At this time, the input conditions may include at least one of an input format, an input range, whether input is required, or a writing criterion. Furthermore, the artificial intelligence processing server (40) can generate a writing instruction that includes at least one of a field meaning description, data format guidelines, value generation criteria, or a writing guide corresponding to the field identification value of the input target field.

[0057] Additionally, the artificial intelligence processing server (40) can compare semantic candidate values ​​derived from context information for field interpretation with domain term information to increase the accuracy of the generated writing instruction, and can produce domain semantic information corresponding to the semantic candidate values ​​based on the comparison result. Additionally, the artificial intelligence processing server (40) can compare the domain semantic information with the relationship information between fields to determine at least one of the arrangement relationship of input semantics, the mutual correspondence relationship of input conditions, or whether there is duplication of field semantics among multiple input target fields within the same context extraction range. Subsequently, the artificial intelligence processing server (40) can correct at least one of the input semantics, input conditions, or writing instruction based on the result of determining semantic consistency.

[0058] Next, the artificial intelligence processing server (40) can generate template schema information including at least one of a field identification value, input meaning, input condition, and a writing instruction for each input target field by correlating the writing instruction or the corrected writing instruction with field structure information. Additionally, the artificial intelligence processing server (40) can obtain record data or natural language-based input source data generated by speech recognition processing, and can generate output values ​​for each input target field from the input source data based on the writing instruction included in the template schema information.

[0059] Subsequently, the artificial intelligence processing server (40) can generate structured output data by correlating the generated output value for each input target field with the field identification value of the input target field. The generated structured output data may be composed of at least one of JSON, key-value pair data, field mapping data, or return data format available from the external electronic record system server (20). Additionally, the artificial intelligence processing server (40) can provide the generated creation instructions or structured output data to the external electronic record system server (20) or the user terminal (10), and the user terminal (10) can display the result in which the structured output data is reflected in each input target field of the template creation screen.

[0060] In this way, the field context analysis-based automatic instruction generation system (1000) according to the present embodiment can operate in such a manner that, in response to a record type selected by a user terminal (10), an external electronic record system server (20) queries a template file from a template database (30), an artificial intelligence processing server (40) analyzes the template file itself without linking internal field metadata through an application program interface of the external electronic record system server (20), and sequentially performs input target field identification, extraction of context information for field interpretation, generation of instruction, correction of instruction, and generation of structured output data.

[0061] Meanwhile, the automatic generation system for writing instructions based on field context analysis in a non-linked template environment according to the present invention (1000) may be implemented as a computing device having one or more processors and a memory that stores programs executed by one or more processors. Here, the computing device may be composed of at least one of a single server, a cloud server, an internal hospital server, an external processing device linked with a user terminal, or an execution environment called in the form of a software development kit. Instructions for analyzing a template file, interpreting the surrounding context of an input target field included in the template file, and generating field-specific writing criteria based on the interpretation results may be stored in the memory. By executing the instructions stored in the memory, one or more processors can determine the meaning of a field and input conditions using structural information, display information, attribute information, and surrounding context information included in the template file itself, without directly accessing the internal field metadata of an external electronic record system.

[0062] The computing device described above can convert a template file into an analyzable internal data structure even when the template file is provided in a different format, such as XML, HTML, CS, a markup document, a screen configuration file, or a data binding definition file. This internal data structure may include the location of input objects, hierarchical relationships, surrounding text, accessibility attributes, input guidance attributes, comment information, and data binding representations as interconnected information. Accordingly, one or more processors can identify objects within the template file that can receive actual values ​​or be mapped to external output values, and analyze the surrounding information and domain semantic information of said objects together to construct decision data necessary for generating authoring instructions.

[0063] Furthermore, the computing device can be linked with generative artificial intelligence or language models to infer the meaning, data format, allowable value range, mandatory input status, and writing criteria of the input target field. For example, even if a specific input field within a template file contains only opaque internal identifiers, the computing device can analyze the input field's parent area, adjacent labels, placeholders, sibling fields, accessibility descriptions, and annotation information to determine which category—chief complaint, present medical history, medical history, vital signs, or prescription details—the input field is closest to. Subsequently, based on the determined meaning and input conditions, the computing device can generate writing instructions indicating what content should be written in the corresponding field and in what format.

[0064] Additionally, the computing device may store the generated writing instructions by correlating them with the field identifiers of a template, or process them together with record data generated by speech recognition processing or natural language-based input source data. Accordingly, one or more processors can extract necessary content by field from the record data in natural language form, generate output values ​​according to the writing criteria of each field, and configure the generated output values ​​into structured output data corresponding to the field identifiers of the original template. For example, if a medical professional's utterance includes current symptoms, medical history, and blood pressure readings, the computing device can separate each piece of content according to the corresponding input items of the template and generate output values ​​in date, numeric, multiple-choice, or natural language descriptive formats.

[0065] Additionally, the computing device may be linked via a network with a user terminal (10) or an external electronic record system server (20) and may transmit or receive at least one of a template file, template access information, voice recognition result, creation instruction, structured output data, or user verification result. At this time, even if the computing device is not deeply linked with the internal application program interface of the external electronic record system, it may determine whether the output value can be reflected by using the template file itself and the display screen or input state of the user terminal (10). Additionally, if the input target field is not in a state where it can be immediately entered, the generated output value may be temporarily stored, and the stored output value may be reflected again when the field is displayed again or switched to a state where it can be entered.

[0066] Hereinafter, with reference to the attached drawings, the detailed configuration of the automatic instruction generation system based on field context analysis in a non-linked template environment according to an embodiment will be described in more detail.

[0067] FIG. 2 is a block diagram showing the overall configuration of a system for automatically generating instructions based on field context analysis in a non-linked template environment according to an embodiment of the present invention, FIG. 3 is a block diagram showing the detailed configuration of a template field identification unit according to an embodiment of the present invention, FIG. 4 is a block diagram showing the detailed configuration of a field context extraction unit according to an embodiment of the present invention, FIG. 5 is a block diagram showing the detailed configuration of an instruction generation unit according to an embodiment of the present invention, FIG. 6 is a block diagram showing the detailed configuration of a field semantics correction unit according to an embodiment of the present invention, FIG. 7 is a block diagram showing the detailed configuration of a structured output generation unit according to an embodiment of the present invention, and FIG. 8 is a block diagram showing the detailed configuration of a field value generation unit according to an embodiment of the present invention.

[0068] Referring to FIG. 2, the automatic instruction generation system (1000) based on field context analysis in a non-linked template environment according to the present embodiment may include at least one of a template field identification unit (100), a field context extraction unit (200), an instruction generation unit (300), a field meaning correction unit (400), and a structured output generation unit (500).

[0069] The above template field identification unit (100) can identify an input target field by analyzing a template file used in a non-linked template environment and generate field structure information corresponding to the identified input target field. To this end, the template field identification unit (100) may include at least one of a template file acquisition unit (110), a template structure conversion unit (120), an input field detection unit (130), and a field structure information generation unit (140), as shown in FIG. 3.

[0070] The above template file acquisition unit (110) can acquire a template file used in a non-linked template environment or template access information for accessing the template file.

[0071] More specifically, the template file acquisition unit (110) may receive a template file to be analyzed in a non-linked template environment from an external electronic record system server (20) or a template database (30), or acquire path information, identification information, storage location information, or access permission information for calling the said template file. Here, the non-linked template environment may refer to an environment where the artificial intelligence processing server (40) cannot directly query the internal field metadata of the external electronic record system server (20) in an application program interface manner, and the template file acquisition unit (110) may acquire input data so that subsequent analysis can be performed using only the template file itself or minimal information accessible to the template file in such an environment. For example, when the external electronic record system server (20) queries a template file selected according to a medical department, document type, or recording purpose from the template database (30), the template file acquisition unit (110) may acquire the original file of the said template file, markup data for rendering, screen configuration file, data binding definition file, or file download address.

[0072] Additionally, the template file acquisition unit (110) can verify the format and source of the template file to be acquired and transmit it in an input form that the template structure conversion unit (120) can subsequently process. For example, if the template file is provided in any one of the formats of XML, HTML, CS, markup document, or screen composition script, the template file acquisition unit (110) can verify the file extension, MIME type, header information, character encoding information, file version information, or template identification value together. This information can be used when the template structure conversion unit (120) converts the corresponding template file into a document object model, hierarchical data, or abstract syntax tree form.

[0073] Additionally, the template file acquisition unit (110) can operate even when only access information is provided and the template file is not directly delivered. For example, if the external electronic record system server (20) does not immediately deliver the original template file due to security policies and only provides a temporary access URL, file identifier, document type code, storage path, or session-based access token, the template file acquisition unit (110) can request a template file from the template database (30) or a designated storage location using the access information. At this time, the template file acquisition unit (110) can prevent an incorrect template or a previous version of the template from being used as an analysis target by checking the access time, request authority, file integrity, template version, or whether the same template has been received multiple times.

[0074] Additionally, the template file acquisition unit (110) may be configured to acquire data based on template definition information for analyzing the structure and context of input fields, rather than patient information or actual record data included in the template file. Accordingly, the template file acquisition unit (110) can provide field identification values, display text, input guidance attributes, accessibility attributes, comment information, data binding expressions, or screen layout information included in the template file to subsequent processing, even without directly accessing the internal database of the external electronic record system server (20). For example, even if an input field in a medical record template of a specific hospital is displayed only as an opaque identification value such as "template-name-3029111", the template file acquisition unit (110) can acquire the entire file containing the identification value or a part thereof, thereby enabling the input field detection unit (130) and the field context extraction unit (200) to analyze surrounding labels, parent areas, sibling fields, and input guidance phrases.

[0075] The above template structure conversion unit (120) can convert the template file into hierarchical structure data used for field identification based on the file format and document structure of the template file obtained from the template file acquisition unit (110).

[0076] More specifically, the template structure conversion unit (120) can convert the template file obtained through the template file acquisition unit (110) into an internal analysis structure suitable for subsequent input target field detection. To this end, the template structure conversion unit (120) can first determine whether the format of the template file corresponds to XML, HTML, CS, markup document, screen configuration file, data binding definition file, or script-based screen definition file, and can select different parsing methods depending on the confirmed file format. For example, in the case of an HTML-based template file, the template structure conversion unit (120) can organize tags, attributes, text nodes, input objects, and container objects into a document object model, and in the case of an XML-based template file, it can separate element names, attribute values, nesting relationships, and data binding expressions into node units to form hierarchical structure data. At this time, the hierarchical structure data may not be a simple list of strings, but may be composed of tree-type data or graph-type data in which the parent object, child object, sibling object, position order, attribute value, and display text of each object are connected to each other.

[0077] Additionally, the template structure conversion unit (120) can preserve not only the input object but also the labels, section titles, instructions, comments, accessibility attributes, and data binding expressions surrounding the input object so that structural clues necessary to interpret the meaning of the actual input target field within the template file are not lost. For example, a specific input object In the case where display text and input objects corresponding to systolic blood pressure, diastolic blood pressure, and heart rate are placed within the same area, the template structure conversion unit (120) can reflect in the hierarchical structure data which parent area the input object belongs to, which sibling field it is placed with, and what distance relationship it has with surrounding display text. Accordingly, the input field detection unit (130) can perform field identification by not only finding input tags such as 'input', 'textarea', and 'select', but also by referring to which document area and context the input target field belongs to within the template.

[0078] Additionally, the template structure conversion unit (120) can normalize dynamic rendering structures or unstructured field representations included in the template file into an analyzable form. For example, if the template file contains data binding representations such as {{field_name}}, iterative rendering areas, conditional display areas, user-defined components, or input objects based on internal identifiers, the template structure conversion unit (120) can classify the representations into input object candidates, display object candidates, or context object candidates and include them in the hierarchical structure data. At this time, even if the internal identifier consists of opaque values ​​such as "template-name-3029111", the template structure conversion unit (120) can provide base data necessary for subsequent field semantic inference by connecting the attributes of the tag to which the identifier belongs, surrounding text, belonging container, screen placement order, and relationship with adjacent objects.

[0079] Additionally, the template structure conversion unit (120) can convert original representations by file format into a standardized internal data model in order to process various template structures used in different hospitals, medical departments, or types of work into a common analysis unit. For example, HTML placeholder, XML attribute value, CS-based screen definition input attribute, accessibility attribute aria-label, and data-* value of data attribute can be integrated into attribute information used for input field interpretation, even if they are expressed in different syntaxes. The hierarchical structure data generated in this way can be used as a common input for detecting input target fields, setting context extraction ranges, determining field meanings, and generating writing instructions, and the same subsequent processing flow can be applied even if the original format of the template file changes.

[0080] The input field detection unit (130) can detect an inputable object from the hierarchical structure data converted by the template structure conversion unit (120) into which a user input value can be inserted or mapped to an external input value, and identify it as an input target field.

[0081] More specifically, the input field detection unit (130) can select an object from the hierarchical structure data generated by the template structure conversion unit (120) into which a user input value can be entered or an output value generated by the artificial intelligence processing server (40) can be inserted as an inputable object candidate. To this end, the input field detection unit (130) can analyze the tag type, attribute value, data binding expression, whether input events can be received, and the correspondence with the screen display object of each node included in the hierarchical structure data. For example, in an HTML-based template, an object with the attributes 'input', 'textarea', 'select', 'option', or 'contenteditable' can be detected as an inputable object candidate, and in an XML or CS-based template, a node corresponding to a specific field identification value, data binding key, value insertion location, user-defined input component, or repetitive input area can be detected as an inputable object candidate. At this time, the input field detection unit (130) can determine not only whether an input tag exists, but also whether the object is structured to accept user input values ​​or external output values ​​in the actual writing screen.

[0082] Additionally, the input field detection unit (130) can distinguish whether an object included in the hierarchical data is a direct input object, a data binding object to which a value is indirectly mapped, or a simple display object. For example, nodes corresponding to items subject to record creation, such as patient name, date of birth, chief complaint, medical history, vital signs, diagnosis, and prescription details, can be identified as input target fields, while section titles, descriptive phrases, guidance text, simple separators, or fixed display values ​​can be excluded from input target fields. However, even if a specific object is not displayed as a general input window on the screen, if it includes data binding expressions such as '{{field_name}}', '{patient.cc}', or 'data-bind="chiefComplaint"' and is defined as a location that can be mapped to an external output value, the input field detection unit (130) can identify the object as an input target field.

[0083] Additionally, the input field detection unit (130) can refine the input target field to be analyzed by checking the attribute status of the inputable object candidate. For example, if a specific object has a read-only attribute, an inactive attribute, or a display-only attribute set, the input field detection unit (130) can exclude the object from automatic creation or classify it as a candidate for manual verification. Conversely, if a required input attribute, maximum input length, input pattern, selectable list, default value, 'placeholder', accessibility attribute, or data binding attribute exists, the input field detection unit (130) can determine that the object is highly likely to be an actual input target field. This determination result can be used as reference data for the subsequent field context extraction unit (200) to interpret the meaning and input conditions of the input target field.

[0084] Additionally, when a single input item is divided into multiple sub-objects, the input field detection unit (130) can identify each sub-object as an individual input target field while also managing them as mutually related field groups. For example, when a blood pressure item is separated into a systolic blood pressure input field and a diastolic blood pressure input field, or a date item is separated into year, month, and day input fields, or an address item is separated into postal code, basic address, and detailed address input fields, the input field detection unit (130) can identify the relationships between each input target field based on the same container, sibling relationship, display text, and arrangement order. Accordingly, the subsequent instruction generation unit (300) can reflect not only the writing guidelines for individual fields but also the criteria for generating values ​​and format relationships between detailed fields within the same item.

[0085] Additionally, the input field detection unit (130) can identify the input target field by utilizing surrounding structure and attribute information together, even when the name or identification value of the input field is opaquely defined for each template. For example, even if the identification value of a specific input object does not directly indicate the field meaning, such as 'template-name-3029111', if the object is placed in a sub-node of the patient basic information area, the adjacent label is displayed as "Name" or "Patient Name," and the phrase "Please enter patient name" is included in the 'placeholder', the input field detection unit (130) can identify the object as a candidate for the patient name input target field. In this way, the input field detection unit (130) can detect inputable objects within the template file and organize each inputable object into an input target field that serves as the basic unit for subsequent field meaning determination and creation instruction generation.

[0086] Additionally, the input field detection unit (130) may operate in a manner that identifies an input target field based on hierarchical structure data and specifies an input target field based on a pointer event occurring at the user terminal (10). Here, the pointer event may include at least one of a mouse click, touch input, pen input, drag, hovering, cursor entry, focus movement, or coordinate selection event. For example, when a user clicks a specific input field on a template creation screen, the input field detection unit (130) may acquire the corresponding click coordinates and convert the acquired screen coordinates into the rendering coordinate system of the template file or the object coordinate system of the hierarchical structure data. Subsequently, the input field detection unit (130) may search for an inputable object containing the converted coordinates and specify an input target field corresponding to the pointer event by referring to the field identification value, field location information, and field attribute information of the searched inputable object.

[0087] Additionally, the input field detection unit (130) can specify an input target field by distinguishing between cases where a pointer event occurs on a direct input object and cases where it occurs on a surrounding display object. For example, if an input cursor is created after a user directly clicks a specific input target field, that input target field can be designated as a direct input target field, and if a user clicks a specific label or display text but an input object connected to that label exists in an adjacent area, that input object can be designated as an indirect input target field. Additionally, if a user places a pointer in a specific area for a certain period of time or longer, the input field detection unit (130) can calculate the priority of inputable objects within that area and specify an input target field with high relevance by considering context information for field interpretation together with the user pointer position.

[0088] The field structure information generation unit (140) can generate field structure information corresponding to the input target field by matching at least one of the field identification value, field location information, field attribute information, and inter-field relationship information of the input target field identified by the input field detection unit (130) with each other.

[0089] More specifically, the field structure information generation unit (140) can organize reference information necessary for subsequent analysis into a single data unit for each object identified as an input target field by the input field detection unit (130). Here, the field identification value may include 'id', 'name', data binding key, internal code value, user-defined attribute value, or object path value assigned to distinguish the input target field in the template file, and the field location information may include information regarding which parent area the input target field belongs to in the hierarchical data structure, in what order it is placed within the same parent area, and in which coordinate or area on the screen it is rendered. Accordingly, the field structure information generation unit (140) can not only record the fact that the input target field simply exists, but also store the location and structural context of the input target field within the template document.

[0090] Additionally, the field structure information generation unit (140) can manage attribute information of an input target field by correlating it with field identification values ​​and location information. For example, if a specific input target field contains attributes or selection list information such as 'placeholder, required, maxlength, pattern, readonly, disabled, aria-label, aria-describedby, title, data-*', the field structure information generation unit (140) can extract these attributes as field attribute information and include them in the field structure information. This attribute information can subsequently be used by the field context extraction unit (200) and the instruction generation unit (300) to determine the meaning, input format, whether input is required, the range of allowed values, and the writing criteria of the input target field. For example, if an input target field has a maxlength="3" and a numeric input pattern and the surrounding area is interpreted as a vital signs section, the field may be more likely to be determined as an input target field related to numeric medical record items such as blood pressure, pulse, and body temperature.

[0091] Additionally, the field structure information generation unit (140) can construct relationships between input target fields. The relationship information between fields may include sibling field relationships contained in the same container, inclusion relationships between a parent section and a child input field, connection relationships between a label and an input object, group relationships where multiple detailed fields constitute a single item, such as a date or address, and chronological relationships based on the display order. For example, if multiple input target fields corresponding to systolic blood pressure, diastolic blood pressure, and heart rate are arranged side by side within a single vital signs area, the field structure information generation unit (140) can manage each input target field as an individual field while also connecting it to a related field group belonging to the same medical measurement section. Accordingly, in subsequent steps, in addition to independently inferring the meaning of individual fields, confusion regarding the meaning of the fields can be reduced by utilizing arrangement relationships and mutual correspondence relationships with surrounding fields.

[0092] Additionally, the field structure information generation unit (140) can organize the generated field structure information into structured data that can be referenced for each input target field. For example, for a single input target field, the field identification value, source node path, parent section name, sibling field list, connection label, screen coordinate, input attribute, data binding expression, and state attribute can be bundled and stored as a single record, and multiple records can be organized into a field structure table or field structure graph in a template unit. In this way, the generated field structure information can serve as a criterion for the context range setting unit (210) to determine up to what range surrounding objects to analyze, and can be used as a criterion for the structured output generation unit (500) to determine which field identification value of the source template to correspond to the generated output value.

[0093] The field context extraction unit (200) can extract context information for field interpretation corresponding to an input target field within a template file based on field structure information generated by the template field identification unit (100). To this end, the field context extraction unit (200) may include at least one of a context range setting unit (210), a context element extraction unit (220), and a context information configuration unit (230), as shown in FIG. 4.

[0094] The context range setting unit (210) can set a context extraction range corresponding to an input target field within a template file based on field structure information generated by the template field identification unit (100). That is, the context range setting unit (210) can set a range that includes at least one of a hierarchically related object, a positionally adjacent object, and an attributeally connected object based on the input target field, based on at least one of field location information and inter-field relationship information included in the field structure information generated by the template field identification unit (100).

[0095] More specifically, the context range setting unit (210) can determine which surrounding information to analyze together to interpret the meaning of the input target field. Since the input target field, when viewed alone, may have only an opaque identification value such as 'template-name-3029111' or appear only in the form of a simple input box, the context range setting unit (210) can determine the range for collecting interpretation clues around the input target field by using location information and relationship information included in the field structure information. For example, if the input target field is included within a specific parent container, section, table row, paragraph block, or input group, the context range setting unit (210) can determine that parent area and its sub-objects as hierarchically related objects, and can include labels, guide text, other input fields, and display text placed in the same parent area as the input target field in the context extraction range.

[0096] Additionally, the context range setting unit (210) may determine that objects close to the input target field on the screen or in the document structure are positionally adjacent objects. For example, if display text such as "Chief Complaint," "Past Medical History," "Systolic Blood Pressure," or "Prescription Details" exists immediately before the input target field, or if a section title such as "Vital Signs" is placed at the top of the input target field, the context range setting unit (210) may determine that such display text and section title are likely surrounding contexts that explain the meaning of the input target field. At this time, the positionally adjacent range may be set based on criteria that fit the template structure, such as the same row, same column, the preceding text node, the following text node, display objects within a certain distance, the same table row, or the same card area relative to the input target field.

[0097] Additionally, the context scope setting unit (210) may include attribute-linked objects explicitly connected to the input target field in the context extraction scope. For example, if the input target field is connected to the 'for' attribute of a specific 'label' tag, or is connected to a separate description object or accessibility object through attributes such as 'aria-label', 'aria-describedby', 'title', 'placeholder', or 'data-*', the context scope setting unit (210) may prioritize these attribute connection relationships over simple surrounding placement. This is because even if the input target field and the description text are placed far apart on the screen, they may be semantically connected to each other through attribute values ​​within the template file. Therefore, the context scope setting unit (210) may set the context extraction scope by considering not only the distance on the screen but also the reference relationships, attribute connection relationships, and data binding relationships between document objects.

[0098] For example, in a specific medical record template, if one input target field is located within a parent area called "vital-signs," and display text or sibling fields such as "SBP," "DBP," and "HR" are placed side by side within the same area, and the 'placeholder' of the input target field contains phrases such as "mmHg" or "number input," the context range setting unit (210) can group the parent area, the label of the same row, adjacent sibling fields, input guide attributes, and unit display text into a single context extraction range. The context extraction range set in this way can be used as a standard by the subsequent context element extraction unit (220) to extract all hierarchical relationships, surrounding input object relationships, belonging area information, display text, input guide attributes, and accessibility attributes necessary for the interpretation of the input target field.

[0099] Additionally, the context range setting unit (210) can gradually expand or contract the context extraction range when the structure of the template is complex or nested. For example, if a clear label exists right next to the input target field, a narrow range centered on the same row or adjacent node can be set first, and if there is no clear text around the input target field or the field identification value is opaque, the range can be extended to the parent section, upper container, sibling field group, or repeated input area. Conversely, if different medical items are densely packed on a single screen, setting too wide a range may incorrectly include labels or descriptive text of other items; therefore, the context range setting unit (210) can limit the analysis range based on hierarchy boundaries, group boundaries, screen layout boundaries, or data binding boundaries. In this way, the context range setting unit (210) can adjust the context extraction range so as not to include objects of low relevance while sufficiently securing the surrounding context necessary for determining the meaning of the input target field.

[0100] The context element extraction unit (220) can extract context element information used for interpreting an input target field from the context extraction range set by the context range setting unit (210). That is, the context element extraction unit (220) can extract at least one of the hierarchical relationship between objects included in the context extraction range set by the context range setting unit (210), the relationship between surrounding input objects, the area of ​​membership information, the display text, the input guide attribute, the accessibility attribute, the comment information, the semantic unit token included in the field identification value, and the semantic unit token included in the field attribute information, and can convert the extracted context element information into semantic candidate information used for determining the field semantics.

[0101] More specifically, the context element extraction unit (220) can separate and extract clues that can be used to infer the meaning of the input target field within the analysis range determined by the context range setting unit (210). At this time, the context element extraction unit (220) can not only simply collect text existing around the input target field but also determine what relationship the text has with the input target field. For example, if the text node immediately preceding the input target field is "Juohoso" and the input target field is placed in the same row or the same container, the context element extraction unit (220) can extract the corresponding display text as candidate information representing the name or meaning of the input target field. Conversely, even if text exists on the same screen, text belonging to a different section or included in a separate input group can be classified as information with low relevance to the input target field.

[0102] Additionally, the context element extraction unit (220) can extract the hierarchical structure to which the input target field belongs and the arrangement relationship of surrounding input objects together. For example, if a specific input target field is included in a sub-node of the "patient basic information" area and input target fields corresponding to name, date of birth, and gender are sequentially arranged within the same area, the context element extraction unit (220) can configure the area of ​​membership information and the sibling field relationship as semantic candidate information. Furthermore, if a specific input target field is arranged together with systolic blood pressure, diastolic blood pressure, and heart rate within the "vital signs" area, the context element extraction unit (220) can extract not only the individual display text of the input target field but also the relationship that all surrounding input objects are fields related to biometric measurements. This relationship information can be used by the subsequent input condition judgment unit (330) to more reliably determine the meaning of the field when the name of the input target field is unclear or when it is difficult to understand the meaning solely from the field identification value.

[0103] Additionally, the context element extraction unit (220) can extract natural language or abbreviation information contained in input guidance attributes, accessibility attributes, comment information, and data attributes. For example, if "mm / dd / yyyy" is included in the 'placeholder' of the input target field, it can be converted into candidate information indicating a date format input condition, and if "Please enter patient name" is included, it can be converted into semantic candidate information regarding the patient name input field. In addition, if the description or use of the input target field is included in attributes such as aria-label, aria-describedby, title, data-label, data-field-type, etc., the context element extraction unit (220) can distinguish the corresponding attribute value from general display text and process it as semantic candidate information with high reliability. In the case of comment information as well, within the template file "<!-- 환자 기본 정보 --> ,<!-- chief complaint section --> If a developer comment such as "is included, the context element extraction unit (220) can extract it as context information that complements the belonging area or field purpose.

[0104] Additionally, the context element extraction unit (220) can separate strings included in field identification values ​​or field attribute values ​​into semantic unit tokens. For example, if a field identification value such as pt_chief_complaint_3029111 exists, the context element extraction unit (220) can separate tokens such as pt, chief, and complaint, and parts that are difficult to directly contribute to semantic judgment, such as numbers or arbitrary code values, can be processed as information with low weight. In contrast, tokens with relatively clear domain meaning, such as 'chief_complaint', can be composed of semantic candidate information related to the chief complaint or reason for visit. Additionally, field attribute information such as 'maxlength', 'pattern', 'required', 'min', 'max', and selection list information can be converted into candidate information for inferring the length, format, required status, allowed range, or selectable value of the input value.

[0105] For example, even if the identification value of a specific input target field is opaque like 'template-name-3029111', if the parent area of ​​the field is "medical record", the adjacent display text is "chief complaint", the 'placeholder' contains the phrase "enter the main symptoms the patient complains of", and the "present medical history" and "past medical history" fields are placed together in the same area, the context element extraction unit (220) can synthesize this hierarchical information, display text, input guidance phrase, and sibling field relationships to convert them into candidate semantic information related to the chief complaint. At this time, the candidate semantic information may be configured in a structured form that includes at least one of a candidate meaning, an evidence element supporting the candidate meaning, the source of the evidence element, the type of relationship with the input target field, and a confidence value, rather than being a simple text bundle.

[0106] Additionally, the context element extraction unit (220) can resolve duplicates or conflicts among the extracted context element information. For example, if the display text is extracted as "blood pressure" but the 'placeholder' contains "Please enter systolic blood pressure" and one of the surrounding sibling fields is displayed as "diastolic blood pressure," the context element extraction unit (220) can determine that the "systolic blood pressure" candidate has a more specific meaning than the simple "blood pressure" candidate. Conversely, if text from different sections exists together around a single input target field, the less relevant candidate can be excluded or set to a lower priority based on hierarchical boundaries, distance, attribute connection status, and sibling field relationships. In this way, the context element extraction unit (220) can provide core context elements necessary for determining the meaning of the input target field as refined semantic candidate information so that subsequent language model-based inference is not affected by unnecessarily broad or ambiguous contexts.

[0107] The above context information configuration unit (230) can be configured as context information for field interpretation by matching the context element information extracted from the context element extraction unit (220) with the input target field.

[0108] More specifically, the context information configuration unit (230) can rearrange the hierarchical relationships, surrounding input object relationships, belonging area information, display text, input guide attributes, accessibility attributes, comment information, and semantic unit tokens extracted individually through the context element extraction unit (220) by input target field, thereby configuring them into context information for field interpretation that can be directly used for determining the meaning of each input target field. For example, even if the field identification value for a single input target field appears opaque as 'template-name-3029111', if the input target field is located in a sub-area of ​​the "Medical Record" section, the adjacent display text is "Chief Complaint," the phrase "Enter the main symptoms complained of by the patient" is included in the 'placeholder', and the "Present Illness" and "Past Illness" input fields are placed together within the same container, the context information configuration unit (230) can group these multiple context elements into a single context information for field interpretation corresponding to the input target field.

[0109] Additionally, the context information configuration unit (230) may not store the extracted context element information in the form of a simple text list, but may configure it in a structured form that includes the type of relationship with the input target field, the source of information, candidate values ​​for meaning, candidate values ​​for input conditions, and basis information necessary for determining reliability. For example, "date of birth" extracted from an adjacent label may be stored as a field meaning candidate, "YYYY-MM-DD" extracted from 'placeholder' may be stored as an input format candidate, and the 'required' attribute may be stored as a required input condition candidate. The context information for field interpretation configured by the context information configuration unit (230) in this manner may be used as input data for the subsequent instruction generation unit (300) to determine the input meaning, input format, allowable value range, and writing criteria of the input target field.

[0110] The instruction generation unit (300) can determine the input meaning and input conditions of an input target field based on the field interpretation context information extracted from the field context extraction unit (200), and generate a writing instruction corresponding to the input target field based on the determined input meaning and input conditions. To this end, the instruction generation unit (300) may include at least one of a context input data generation unit (310), an inference request data generation unit (320), an input condition determination unit (330), and a field-specific instruction generation unit (340), as shown in FIG. 5.

[0111] The above context input data generation unit (310) can generate context input data by sorting the field interpretation context information extracted from the field context extraction unit (200) by input target field and configuring the context element information included in the sorted field interpretation context information into an input format used for determining the meaning of the field.

[0112] More specifically, the context input data generation unit (310) can organize the context information for field interpretation generated by the field context extraction unit (200) into input target fields and convert it into an input data form that is easy for a language model or rule-based judgment logic to process. At this time, the context input data generation unit (310) can arrange the field identification value, original node path, affiliated area, adjacent label, 'placeholder', accessibility attribute, comment information, sibling field relationship, and semantic unit token of each input target field into a single field-specific data bundle so that context information for multiple input target fields is not mixed. For example, if a specific template includes multiple input target fields related to patient name, date of birth, chief complaint, current medical history, and vital signs, the context input data generation unit (310) can arrange the "information of the field itself," "contextual information surrounding the field," "relationship information between the field and other fields," and "attribute information capable of estimating input conditions" separately for each input target field.

[0113] Additionally, the context input data generation unit (310) can reconstruct various clues included in context information for field interpretation into a format suitable for determining the meaning of the field. For example, adjacent display text can be classified as field name candidates, 'placeholder' or input guide phrases as input format candidates, attributes such as 'required', 'maxlength', 'pattern', 'min', 'max' as input condition candidates, and parent section names and sibling field relationships as context correction information. At this time, the context input data can be configured in at least one form among natural language sentences, key-value pair data, structured data in JSON form, or block data for prompt insertion. For example, if the identification value of the input target field is 'template-name-3029111', the adjacent label is "chief complaint", the parent area is "medical record", and the 'placeholder' is "input the main symptoms the patient complains of", the context input data generation unit (310) can structure the information so that it is possible to determine that "the field meaning candidate is chief complaint, the input value is likely a natural language record regarding the patient's main symptoms, and must be distinguished from the current medical history and past medical history fields of the same area".

[0114] Additionally, the context input data generation unit (310) may include only the information necessary for determining the meaning of the input target field in the context input data, so that the language model does not unnecessarily interpret the entire original template. For example, tags for screen decoration, style attributes, simple layout information, meaningless random internal codes, or text from other sections that are not related to the input field may be excluded or treated as low-priority information. Conversely, label connection information, accessibility descriptions, data binding keys, parent section names, sibling field arrays, and input validation conditions that directly affect the determination of the field's meaning may be included as high-priority information. Accordingly, the context input data generation unit (310) can convert the complex structure of the original template into refined input data suitable for field-specific interpretation so that the subsequent inference request data generation unit (320) can reliably determine the meaning and input conditions of the input target field.

[0115] The above inference request data generation unit (320) can convert the context input data generated by the context input data generation unit (310) into inference request data that can be input into a language model.

[0116] More specifically, the inference request data generation unit (320) can reconstruct the context input data organized by field by the context input data generation unit (310) into a request format that can be interpreted by a large-scale language model (LLM). At this time, the inference request data generation unit (320) can convert the field identification value, membership area, adjacent label, input guide phrase, accessibility attribute, annotation information, sibling field relationship, semantic unit token, and input validation condition for each input target field into the input prompt, message object, key-value pair data, or structured query data of the language model. For example, if the context input data includes "field identification value: template-name-3029111", "domain: medical record", "adjacent label: chief complaint", "placeholder: input main symptoms complained of by the patient", and "sibling field: present medical history, past medical history", the inference request data generation unit (320) can be configured as inference request data in the form of a question or command so that the language model can infer the meaning of the input target field as the chief complaint input field and determine the criteria for writing the value to be entered into the field.

[0117] Additionally, the inference request data generation unit (320) may also specify the format of the judgment result that the language model must output. For example, the inference request data may include an output schema that returns at least one of the following: a semantic description of the input target field, a data type, an input format, a range of allowed values, whether input is required, a writing standard, a standard for distinguishing it from surrounding fields, and a basis for judging reliability. In this case, the inference request data generation unit (320) can reduce unnecessary contextual confusion by not inputting the entire original template as is, but by selecting and providing only the core context necessary for judging the field's semantics to the language model. Furthermore, when processing multiple input target fields at once, the separator for each field, the field order, relationship information, and arrangement information within the same section may be included together so that the language model does not judge the semantics of adjacent fields redundantly or interpret them interchangeably.

[0118] The above input condition determination unit (330) can determine an input condition including at least one of the input meaning, input format, input range, required input, and writing criteria of an input target field based on the inference request data generated by the inference request data generation unit (320).

[0119] More specifically, the input condition determination unit (330) can determine whether the input target field is a field intended to receive certain content, what format the value should be generated in, and whether there is a range or constraint on the inputable value. For example, if the adjacent label of the input target field is "date of birth", the placeholder is "YYYY-MM-DD", and the date format is set in the 'pattern' attribute, the input condition determination unit (330) can determine the input meaning of the input target field as date of birth and determine the input format as date type data. In addition, if the adjacent label is "Chief Complaint" and the "Current Medical History" and "Past Medical History" fields are placed together in the same area, the input condition judgment unit (330) can determine that the input target field is a field for recording the main symptoms currently complained of by the patient or the reason for visiting in natural language, and can distinguish the writing criteria for each field so that the time of onset, course, and related circumstances of the symptoms are recorded in the current medical history field, and existing diseases or surgical history are recorded in the past medical history field.

[0120] Additionally, the input condition determination unit (330) can determine the input range, whether input is required, and the writing criteria by considering the explicit attributes included in the template file and the surrounding context. For example, if a specific input target field includes the 'required' attribute, it may be determined as a required input field; if 'maxlength="50"' is included, it may be determined as a field where the maximum length of the input value is limited to 50 characters; and if a 'select' object or a selectable list exists, it may be determined as a field where one of the predetermined selected values ​​must be entered rather than a free-form descriptive value. Additionally, if an input target field included in the vital signs area has a label "systolic blood pressure," a unit display "mmHg," and a numeric input pattern, the input condition determination unit (330) may determine the input meaning of the field as systolic blood pressure, determine the input format as a numeric value, and set the writing criteria as numeric-centered input excluding unit characters.

[0121] Additionally, when the meaning of the input target field is not clear from a single context alone, the input condition judgment unit (330) can use hierarchical relationships, sibling field relationships, and domain term information together to make a complementary determination of the input condition. For example, even if the identification value of the input target field consists only of an opaque internal code value, if the field is placed in the "Patient Basic Information" area, and the fields for name, gender, and date of birth exist in order around it, and a specific input object is configured in the form of a selection list, the input condition judgment unit (330) can determine the possibility that the input target field is a selection type field related to gender or patient classification items. In this case, the input condition judgment unit (330) does not forcefully calculate only one confirmed value, but generates meaning candidates, judgment grounds, input format candidates, and whether correction is necessary together, so that the subsequent field-specific instruction generation unit (340) or field meaning correction unit (400) can generate or correct more stable writing instructions.

[0122] The field-specific instruction generation unit (340) can generate a writing instruction that includes at least one of a field meaning description, data format guidelines, value generation criteria, and writing guide corresponding to the field identification value of the input target field, based on the input meaning and input conditions determined by the input condition determination unit (330).

[0123] More specifically, the field-specific instruction generation unit (340) can convert the meaning of the input target field and the input conditions determined by the input condition determination unit (330) into a form of writing instructions that the generative artificial intelligence can use to generate actual structured output. At this time, the writing instructions are not merely at the level of changing the field name into natural language, but may include what the input target field is intended to record, from which source data the value should be extracted, in what format the value should be generated, and by what criteria the content should be distinguished from other input target fields. For example, even if the field identification value of the input target field is opaque, such as 'template-name-3029111', if the input condition determination unit (330) determines that the field is a main complaint input field, the field-specific instruction generation unit (340) can generate a writing instruction corresponding to the field identification value, such as "write the main symptoms currently complained of by the patient or the reason for visit in a concise natural language sentence, and do not include past medical history or treatment history."

[0124] Additionally, the field-specific instruction generation unit (340) can configure the data format of the input target field and the value generation criteria together. For example, for the date of birth field, it can be specified to generate the date value in the 'YYYY-MM-DD' format, and for the systolic blood pressure field, it can be specified to extract only the numeric value corresponding to the systolic blood pressure from voice recognition record data or natural language-based input source data and exclude unit characters such as mmHg. Additionally, for optional fields, the writing criteria can be set to convert the free sentence into one of the values ​​in a predefined list of selectable options without generating a free sentence, and for required input fields, if it is not directly verified in the source data, it can be set to be processed as a blank, a verification required indicator, or a user verification request value.

[0125] Additionally, the field-specific instruction generation unit (340) may include instructions for distinguishing the writing scope between each field when input target fields with similar meanings exist within the same context range. For example, if the "Chief Complaint," "Current Illness," and "Past History" fields are placed in the same medical record area, different writing guides may be generated to write the core symptom that the patient currently describes as most uncomfortable in the Chief Complaint field, the time of onset, progression, and factors of aggravation or alleviation of the symptom in the Current Illness field, and information regarding existing diseases, surgical history, or long-term medications in the Past History field. Accordingly, the structured output generation unit (500) can generate output values ​​by separating the content extracted from a single natural language medical record data according to the purpose of each input target field.

[0126] Additionally, the field-specific instruction generation unit (340) may store the generated instruction by correlating it with the field identification value of the original template, or provide it in the form of structured data that can be used by the template schema generation unit (510). For example, a single instruction may be composed of a record including at least one of a field identification value, a field meaning description, a data type, an input format, an allowable value range, whether it is required, a value generation criterion, an exclusion criterion, a criterion for distinguishing it from related fields, and a basis for judgment. The instruction composed in this way can be used as reference information when converting voice recognition record data or natural language-based input source data into structured output data for each input target field in a subsequent step.

[0127] The field semantic correction unit (400) can correct the accuracy of context information for field interpretation extracted from the field context extraction unit (200) or the writing instruction generated from the instruction generation unit (300). To this end, the field semantic correction unit (400) may include at least one of a semantic candidate value acquisition unit (410), a domain semantic matching unit (420), a semantic consistency determination unit (430), and a field interpretation correction unit (440), as shown in FIG. 6.

[0128] The above-mentioned meaning candidate value acquisition unit (410) can acquire meaning candidate values ​​used for determining the meaning of an input target field from context information for field interpretation extracted from the field context extraction unit (200).

[0129] More specifically, the semantic candidate value acquisition unit (410) can select candidate expressions that can directly or indirectly indicate the meaning of the input target field from among several clues included in the context information for field interpretation. For example, "chief complaint" in the label connected to the input target field, "enter the main symptoms complained of by the patient" in the 'placeholder', "medical record" in the parent area, "present history" and "past history" in the sibling field, and the 'chief_complaint' token within the field identification value can each be acquired as semantic candidate values ​​indicating the possibility that the input target field is a chief complaint input item. At this time, the semantic candidate value acquisition unit (410) is not limited to using strings extracted from display text, input guidance phrases, accessibility attributes, annotation information, data binding keys, field identification values, and field attribute values ​​as they are, but can organize them into candidate values ​​that the subsequent domain semantic matching unit (420) can compare by performing abbreviation separation, removal of unnecessary numbers, tokenization based on separators, conversion of English abbreviations into Korean meanings, and normalization of medical domain expressions.

[0130] Additionally, the semantic candidate value acquisition unit (410) can manage the source and relevance of each candidate value when multiple semantic candidate values ​​are derived simultaneously. For example, if "blood pressure" is extracted from adjacent labels for the same input target field, "systolic blood pressure input" is extracted from 'placeholder', and "diastolic blood pressure" and "heart rate" are identified as sibling fields, the semantic candidate value acquisition unit (410) can acquire the upper candidate value "blood pressure" and the specific candidate value "systolic blood pressure" together, but can assign a higher priority to the "systolic blood pressure" candidate value based on the relationship between the 'placeholder' and sibling fields. Conversely, values ​​that are difficult to directly contribute to semantic judgment, such as arbitrary codes or internal template management numbers included in the field identification value, may be classified as having a lower priority or excluded from the candidate values. Accordingly, the semantic candidate value acquisition unit (410) can provide candidate semantics, candidate sources, relationship types, and priority information that the field semantic correction unit (400) can use when correcting the semantics of the input target field.

[0131] The above domain semantic matching unit (420) can calculate domain semantic information corresponding to the semantic candidate value by comparing the semantic candidate value obtained from the semantic candidate value acquisition unit (410) with pre-set domain term information.

[0132] More specifically, the domain semantic matching unit (420) can compare candidate expressions obtained through the semantic candidate value acquisition unit (410) with standard terms, abbreviations, synonyms, code names, or item classification information used in the fields of medical, counseling, administrative records, or electronic document creation. For example, if a semantic candidate value is extracted as "CC," "chief complaint," "main symptom," "reason for visit," or "content that the patient complained of as most uncomfortable," the domain semantic matching unit (420) can search for a standard meaning corresponding to this in the pre-set domain term information and produce domain semantic information corresponding to "chief complaint" or "main reason for visit." In addition, even if a semantic candidate value is extracted in the form of an abbreviation such as BP, SBP, DBP, HR, or BT, the domain semantic matching unit (420) can convert this into domain semantic information such as blood pressure, systolic blood pressure, diastolic blood pressure, heart rate, body temperature, etc., respectively.

[0133] Additionally, the domain semantic matching unit (420) can determine the correspondence between semantic candidate values ​​and domain term information by combining not only a simple string matching method but also morphological analysis, abbreviation dictionary matching, synonym matching, similarity comparison, superordinate / subordinate concept comparison, or context-based semantic comparison. For example, even if candidate values ​​are extracted as different expressions such as "past history," "previous history," "past medical history," and "past_history," the domain semantic matching unit (420) can determine that these expressions refer to identical or similar medical record items and normalize them into common domain semantic information. Conversely, if only a superordinate concept is extracted, such as "blood pressure," the surrounding sibling fields, unit information, input patterns, or label detailed expressions can be referenced together to distinguish whether it is systolic blood pressure, diastolic blood pressure, or a simple blood pressure item; if it is difficult to distinguish, multiple domain semantic candidates and reliability information for each candidate can be calculated together.

[0134] Additionally, the domain semantic matching unit (420) can provide the generated domain semantic information in a structured form so that the subsequent semantic consistency judgment unit (430) can utilize it. For example, for a single input target field, information including at least one of a domain standard name, domain classification, synonym list, data type, general input format, related field, and matching basis may be generated. Accordingly, even if the field identification value or surrounding text is expressed differently by hospital, medical department, or template, the domain semantic matching unit (420) can organize different candidate expressions into a common domain semantic system so that the semantic expression of the writing instruction and the input criteria are generated consistently.

[0135] The semantic consistency determination unit (430) can determine the semantic consistency of an input target field by comparing the domain semantic information calculated by the domain semantic matching unit (420) with the field relationship information included in the field interpretation context information extracted by the field context extraction unit (200). That is, the semantic consistency determination unit (430) can determine at least one of the arrangement relationship of input semantics, the mutual correspondence relationship of input conditions, and whether there is duplication of field semantics among multiple input target fields within the same context extraction range by comparing the domain semantic information calculated by the domain semantic matching unit (420) with the field relationship information included in the field interpretation context information extracted by the field context extraction unit (200).

[0136] More specifically, the semantic consistency judgment unit (430) can verify whether the domain semantic information of each input target field calculated through the domain semantic matching unit (420) does not contradict the actual arrangement structure within the template. For example, if input target fields corresponding to "systolic blood pressure," "diastolic blood pressure," and "heart rate" are sequentially arranged within a single context extraction range, and each input target field is composed of sub-fields of the same vital sign area, the semantic consistency judgment unit (430) can determine whether the domain semantic information of each field has a consistent arrangement relationship with the detailed measurement values ​​of the vital sign items. At this time, even if the semantic candidate of a specific input target field is calculated as simply "blood pressure," the domain semantic information of adjacent fields, unit display, input order, and sibling field relationships can be compared together to determine whether the input target field corresponds to systolic blood pressure or diastolic blood pressure.

[0137] Additionally, the semantic consistency judgment unit (430) can determine whether the meanings of multiple input target fields included in the same parent area or the same input group are duplicated or reversed. For example, if the "Chief Complaint," "Current Illness," and "Past Medical History" fields are placed in the same medical record area, but both input target fields are matched to "Chief Complaint," the semantic consistency judgment unit (430) can determine whether there is a semantic duplication by comparing surrounding display text, field arrangement order, 'placeholder', and sibling field relationships again. Additionally, in a patient basic information area where the date of birth field and the gender field are placed together, if a field with a selection list is determined to be a date type field, or if a field with a date pattern is determined to be a gender field, the semantic consistency judgment unit (430) can determine that the mutual correspondence relationship of the input conditions is incorrect.

[0138] Additionally, the semantic consistency judgment unit (430) may also examine whether the input semantics and input conditions logically correspond to each other. For example, if the input target field, for which the domain semantic information is calculated as "body temperature," has a numerical input pattern, a decimal point allowance condition, and a "℃" unit display, it may be determined that the semantics and input conditions are consistent. Conversely, even if the domain semantic information is calculated as "comprehensive findings," if the input target field has a very short maximum input length, a numerical-only pattern, or a selection list attribute, it may be determined that there is a possibility of error in the semantic judgment or input condition judgment. In such cases, the semantic consistency judgment unit (430) may mark the input target field as a field requiring correction, or set a candidate that corresponds better to the input conditions among multiple semantic candidates as a priority candidate.

[0139] Additionally, the semantic consistency judgment unit (430) can refine the criteria for distinguishing instructions for each field by utilizing the arrangement relationship of field semantics within the same context extraction range. For example, if the fields "Symptom," "Time of Occurrence," "Duration," "Aggravating Factor," and "Relief Factor" are arranged in order in a single medical history template, the semantic consistency judgment unit (430) can determine that each input target field shares the same context regarding the patient's symptoms but receives different detailed information. Accordingly, the subsequent field interpretation correction unit (440) can reflect the name of the symptom in the symptom field, the start time in the time of occurrence field, the duration information in the duration field, and the condition of symptom change in the aggravating factor and relief factor fields, respectively, so that the writing criteria for each field do not overlap.

[0140] The field interpretation correction unit (440) can correct at least one of the input meaning, input conditions, and writing instructions of the input target field according to the judgment result of the meaning consistency judgment unit (430).

[0141] More specifically, the field interpretation correction unit (440) can modify the interpretation result of an input target field based on semantic overlap, semantic conflict, input condition mismatch, or arrangement relationship error identified by the semantic consistency judgment unit (430). For example, if two of the multiple input target fields placed in the same medical record area are both interpreted as "chief complaint," but there is a display text of "present illness" and a 'placeholder' regarding the course of symptoms around one of the fields, the field interpretation correction unit (440) can correct the input meaning of the field to "present illness." Additionally, if the domain meaning of an input target field is determined to be "date of birth," but there is a selection list in the actual field attributes and the surrounding label is placed closer to "gender," the field interpretation correction unit (440) can modify the input meaning to the gender item, or at least mark the field as a target requiring correction.

[0142] Additionally, the field interpretation correction unit (440) can correct not only the input meaning but also the input format, input range, whether input is required, and writing criteria. For example, if a specific input target field is interpreted as "systolic blood pressure" but the writing instruction includes a natural language sentence writing method, the field interpretation correction unit (440) can correct the writing instruction by modifying it to focus on numeric values, excluding unit characters, or processing them separately to match the template requirement format. Additionally, if attribute information such as 'maxlength', 'pattern', 'required', and a selectable list is not sufficiently reflected in the input condition judgment result, the field interpretation correction unit (440) can re-correct the original field attribute information and domain semantic information to add input value length, format restrictions, required input processing methods, or selection value conversion criteria to the writing instruction.

[0143] Additionally, the field interpretation correction unit (440) can organize the correction results into a final writing instruction corresponding to the field identification value of the input target field so that the subsequent structured output generation unit (500) can use them immediately. For example, the "Chief Complaint" field can be corrected to write only the core expression of the current symptom, the "Present Illness History" field can be corrected to write the time of onset, progression, and factors of aggravation or alleviation of the symptom, and the "Past Medical History" field can be corrected to write existing diseases, surgical history, or long-term medications. In this way, the field interpretation correction unit (440) organizes the field meaning and input conditions calculated in the preceding step to match the actual field relationships within the template, thereby reducing situations where the same content is duplicated in different input target fields or output values ​​different from the field purpose are generated.

[0144] The structured output generation unit (500) can generate structured output data corresponding to an input target field from input source data using a creation instruction generated by the instruction generation unit (300). To this end, the structured output generation unit (500) may include at least one of a template schema generation unit (510), an input source data acquisition unit (520), a field value generation unit (530), and a structured data generation unit (540), as shown in FIG. 7. Additionally, the field value generation unit (530) may include at least one of a record data splitting unit (531), a field correspondence relationship calculation unit (532), and an output value generation unit (533), as shown in FIG. 8.

[0145] The above template schema generation unit (510) can generate template schema information including at least one of a field identification value, input meaning, input condition, and a creation instruction by matching the creation instruction generated by the instruction generation unit (300) with the field structure information generated by the template field identification unit (100).

[0146] More specifically, the template schema generation unit (510) can correspond the field identification value, original node path, data binding key, on-screen location information, and relationship information between fields included in the field structure information with the writing instruction generated in the previous step in order to link the writing instruction to the input target field of the original template. For example, if the field identification value of a specific input target field is 'template-name-3029111' and the field is determined to be a "main complaint" item in the writing instruction, the template schema generation unit (510) can configure the field identification value, the input meaning of "main complaint," natural language descriptive input conditions, current symptom-centered writing criteria, and exclusion criteria into a single field schema item.

[0147] Additionally, the template schema generation unit (510) can construct a data structure used for generating structured output by reflecting the data format of the input target field and the relationships between fields. For example, the date of birth field can be configured as a date format schema, the systolic blood pressure field can be configured as a numeric input schema, and the chief complaint or current medical history field can be configured as a natural language descriptive schema. Furthermore, if the "chief complaint," "current medical history," and "past medical history" fields are included in the same medical record area, the template schema generation unit (510) can generate each field as an individual schema item while also connecting it to a related field belonging to the same context group. Accordingly, the subsequent field value generation unit (530) can distribute the content extracted from the input source data according to the purpose of each input target field, and can reduce the duplication of the same content in multiple fields or the swapping of values ​​in different fields.

[0148] The above input source data acquisition unit (520) can acquire input source data to be processed together with template schema information generated by the template schema generation unit (510). That is, the input source data acquisition unit (520) can acquire record data generated by voice recognition processing as input source data.

[0149] More specifically, the input source data acquisition unit (520) may acquire at least one of the following as basic data for generating output values ​​for each input target field according to template schema information: the content of a user's speech, speech recognition result text, natural language record, consultation memo, medical draft, or external input document. For example, if a medical professional speaks through a microphone connected to a user terminal (10), such as "The patient has complained of headache and fever since yesterday and has a history of hypertension," the input source data acquisition unit (520) may receive text data converted by a speech recognition module or speech recognition server as input source data.

[0150] Additionally, the input source data acquisition unit (520) can preprocess the acquired input source data so that it can be processed together with the template schema information. For example, if the speech recognition result contains unnecessary interjections, repetitive utterances, sentence breaks, recognition errors, medical abbreviations, numerical expressions, or negative expressions, the input source data acquisition unit (520) can perform sentence cleanup, utterance segment merging, special character normalization, unit expression cleanup, numerical expression conversion, or removal of unnecessary spaces. Accordingly, the subsequent field value generation unit (530) can generate output values ​​corresponding to each input target field by referencing the cleaned input source data and the writing instructions included in the template schema information together.

[0151] The above field value generation unit (530) can generate output values ​​for each input target field from the input source data obtained by the input source data acquisition unit (520) based on the creation instructions included in the template schema information generated by the template schema generation unit (510).

[0152] More specifically, the field value generation unit (530) can calculate the value to be entered into each input target field by comparing the natural language expression included in the input source data with the input meaning, input condition, and writing criteria included in the template schema information. For example, even if current symptoms, past medical history, blood pressure values, and prescription details are included together in a single sentence flow in the input source data, the field value generation unit (530) can separate the current symptoms into the main complaint field, the existing disease information into the past medical history field, and the blood pressure values ​​into the vital signs field, respectively, according to the writing instructions.

[0153] The above record data division unit (531) can generate record unit data by dividing the record data obtained from the input source data acquisition unit (520) into utterance content units or sentence units.

[0154] More specifically, the record data splitting unit (531) can divide the voice recognition result text or natural language-based record data into small processing units suitable for generating field values. At this time, the record data splitting unit (531) may not separate the record data based only on periods or line breaks, but may also consider the connection relationships of speech time intervals, sentence boundaries, semantic units, medical item transition expressions, numerical expressions, and negative expressions. For example, if record data such as "The patient has complained of headache and fever since yesterday. There is a history of hypertension, and today's blood pressure is 150 / 90" is obtained, the record data splitting unit (531) may split the content regarding "complaining of headache and fever since yesterday," the content regarding "history of hypertension," and the content regarding "blood pressure 150 / 90" into different record unit data.

[0155] The field correspondence relationship calculation unit (532) can calculate the correspondence relationship between the record unit data and the input target field by comparing the record unit data generated by the record data division unit (531) with the creation instructions included in the template schema information generated by the template schema generation unit (510).

[0156] More specifically, the field correspondence calculation unit (532) can compare the core expressions, domain terms, numerical values, time information, and negative expressions included in the record unit data with the writing instructions to determine which input target field the record unit data should be reflected. For example, if the template schema information includes writing instructions such as “the main complaint field writes the core symptoms currently complained of by the patient,” “the history field writes the history of existing diseases or surgeries,” and “the systolic blood pressure field writes only the first numerical value of the blood pressure expression,” the field correspondence calculation unit (532) can correspond the record unit data “complaining of headache and fever” to the main complaint field, the record unit data regarding “history of hypertension” to the history field, and the record unit data regarding “blood pressure 150 to 90” to the systolic blood pressure and diastolic blood pressure fields, respectively.

[0157] Additionally, when one record unit data is related to multiple input target fields, the field correspondence relationship calculation unit (532) can separate and calculate the correspondence relationship by referring together to the creation instruction, domain semantic information, relationship information between fields, and input format information. For example, if there is record unit data such as "blood pressure is 90 and pulse is 150 beats", the field correspondence relationship calculation unit (532) can separate the record unit data into detailed values ​​corresponding to the systolic blood pressure field, diastolic blood pressure field, and heart rate field, respectively.

[0158] The above output value generation unit (533) can generate output values ​​for each input target field from the recording unit data based on the correspondence relationship calculated by the field correspondence relationship calculation unit (532).

[0159] More specifically, the output value generation unit (533) can organize values ​​according to the input meaning, data format, creation criteria, and exclusion criteria of each input target field, rather than copying the record unit data as is. For example, a concise expression centered on core symptoms, such as "headache and fever," can be generated in the chief complaint field, information on past diseases, such as "history of hypertension," can be generated in the past history field, and numeric output values, such as "150" in the systolic blood pressure field and "90" in the diastolic blood pressure field, can be generated. Additionally, in fields with a specified date format, the date expression included in the input source data can be converted into the format required by the template schema information, and in optional fields, the natural language expression can be converted into one of the predefined optional values ​​or code values.

[0160] The above structured data generation unit (540) can generate structured output data by matching the output value for each input target field generated by the field value generation unit (530) with the field identification value of the input target field.

[0161] More specifically, the structured data generation unit (540) can configure the field identification value of the original template and the output value into a key-value pair, JSON, object array, field mapping data, or a return data format acceptable to the external electronic record system server (20). For example, if the original identification value of the input target field is 'template-name-3029111' and the field is interpreted as a primary complaint field and the output value is generated as "headache and fever," the structured data generation unit (540) can correspond the field identification value and the output value to a single structured item. Additionally, the structured data generation unit (540) may include at least one of the record unit data that is the source of each output value, the applied writing instruction, the generation reliability, whether it is undetermined, or whether user verification is required.

[0162] Additionally, the structured data generation unit (540) may refer to at least one of the field structure information of the input target field, screen rendering state information, focus state information, attribute state information, and user interaction state information to determine whether the generated structured output data can be reflected on the actual template creation screen. For example, if the input target field is identified as an input-capable object in the hierarchical structure data, is rendered in a visible area on the display screen of the user terminal (10), is capable of receiving focus events, and does not have read-only or inactive attributes set, the structured data generation unit (540) may classify the output value corresponding to the input target field as an output value that can be automatically reflected. Conversely, if the input target field is not displayed on the screen, is obscured by another screen object, or has an automatic input restriction attribute, the output value may be classified as an output value corresponding to a manual verification target or an alternative injection target.

[0163] Additionally, the structured data generation unit (540) may temporarily store the output value as buffer text when an output value is generated for each input target field but the corresponding input target field is not in a state where it can be immediately entered. The buffer text may be managed in correspondence with at least one of the field identification value, output value, creation instruction, input condition, generation time, number of injection attempts, reason for the last injection failure, and re-injection waiting state of the input target field. For example, if an output value corresponding to the main field is generated but the field is not displayed on the current tab screen or does not acquire focus, the structured data generation unit (540) may not immediately discard the output value but maintain it as buffer text, and subsequently, when an input target field having the same field identification value is detected again, the output value stored in the buffer text may be corresponded to the re-detected input target field.

[0164] Additionally, the structured data generation unit (540) can switch the state of the buffer text based on whether the input target field screen is activated, whether focus is acquired, whether the same field identification value is re-detected, whether the re-rendering of the template screen is completed, or whether the correspondence between the cursor position of the user terminal (10) and the input target field is confirmed. For example, if the user moves to a screen area containing the input target field through scrolling, tab switching, or popup switching, and the input target field having the same field identification value is detected again, the structured data generation unit (540) can set the output value stored in the buffer text as the re-injection target value corresponding to the re-detected input target field. Conversely, if the attribute of the input target field is changed to read-only, if it is determined that the meaning of the input target field does not match the existing written instruction, or if the number of injection failures exceeds a threshold value, the buffer text can be switched to a manual verification waiting state.

[0165] Meanwhile, the structured output generation unit (500) classifies the input environment in which the generated structured output data is reflected on the actual template creation screen, and selects the method of injecting output values ​​for each input target field according to the classified input environment. The input environment may be classified into at least one of a web-based input environment, a desktop application-based input environment, an embedded webview-based input environment, a remote screen-based input environment, an image-based or canvas-based input environment, a clipboard-based input environment, and an accessibility interface-based input environment. For example, in a web-based input environment, methods such as accessing document object model objects, triggering script events, or setting form values ​​may be applied, and in a desktop application-based input environment, methods such as operating system input events, accessibility object control, or clipboard pasting may be applied.

[0166] Additionally, the structured output generation unit (500) can classify the input environment based on at least one of the following criteria: whether an object of an electronic record screen running on a user terminal (10) is accessible, whether a document object model is accessible, whether an accessibility tree is provided, whether input events can be received, whether a clipboard can be used, whether mapping between screen coordinates and field objects is possible, and whether automatic input is restricted by a security policy. For example, in a web-based input environment, if the document object model object access method fails, it can sequentially switch to script event-based injection, clipboard-based injection, or keyboard event-based injection, and in a desktop application-based input environment, if accessibility object control is restricted, it can inject an output value after giving focus to the input target field using a pointer event.

[0167] In this way, the field context analysis-based automatic instruction generation system (1000) according to the present embodiment can determine the meaning and input conditions of an input target field by utilizing the structure and context within a template file even in a template environment where it is difficult to understand the meaning based only on opaque field identification values, and can automatically generate a suitable instruction for each input target field.

[0168] The above description is merely one embodiment for implementing a system for automatically generating instructions based on field context analysis in a non-linked template environment according to the present invention. The present invention is not limited to the above embodiment, and the technical spirit of the present invention extends to the scope in which various modifications can be made by anyone with ordinary knowledge in the field to which the invention belongs, without departing from the gist of the invention as claimed in the following claims. Explanation of the symbols

[0169] 1000: Automatic Instruction Generation System Based on Field Context Analysis 100: Template field identifier 110: Template File Acquisition Section 120: Template Structure Conversion Section 130: Input field detection unit 140: Field Structure Information Generation Section 200: Field Context Extraction Unit 210: Context scope setting section 220: Context Element Extraction Section 230: Context Information Configuration Section 300: Instruction generation section 310: Context Input Data Generation Section 320: Inference Request Data Generation Unit 330: Input condition judgment unit 340: Instruction generation section by field 400: Field meaning correction section 410: Meaning candidate value acquisition part 420: Domain semantic matching part 430: Semantic consistency judgment part 440: Field interpretation correction unit 500: Structured Output Generator 510: Template Schema Generation Section 520: Input source data acquisition unit 530: Field value generation section 531: Record data partition 532: Field Correspondence Relationship Calculation Unit 533: Output value generation section 540: Structured Data Generation Unit 10: User terminal 20: External electronic record system server 30: Template Database 40: Artificial Intelligence Processing Server

Claims

Claim 1 A system implemented as a computing device having one or more processors and memory for storing programs controlling them, comprising: a template field identification unit that analyzes a template file used in a non-linked template environment to identify an input target field and generates field structure information corresponding to the identified input target field; and a field context extraction unit that extracts field interpretation context information corresponding to the input target field within the template file based on the field structure information generated by the template field identification unit. The invention includes an instruction generation unit that determines the input meaning and input conditions of the input target field based on the field interpretation context information extracted by the field context extraction unit, and generates a writing instruction corresponding to the input target field based on the determined input meaning and input conditions. The instruction generation unit comprises: a context input data generation unit that sorts the field interpretation context information extracted by the field context extraction unit by input target field and configures the context element information included in the sorted field interpretation context information into an input format used for determining the field meaning to generate context input data; an inference request data generation unit that converts the context input data generated by the context input data generation unit into inference request data that can be input into a language model; and an input condition determination unit that determines an input condition including at least one of the input meaning, input format, input range, whether input is required, and writing criteria of the input target field based on the inference request data generated by the inference request data generation unit. A system for automatically generating writing instructions based on field context analysis in a non-linked template environment, characterized by including a field-specific instruction generation unit that generates a writing instruction comprising at least one of a field meaning description, data format guidelines, value generation criteria, and a writing guide corresponding to a field identification value of the input target field, based on an input meaning and input conditions determined by the input condition determination unit. Claim 2 The system for automatically generating instructions based on field context analysis in a non-linked template environment according to claim 1, wherein the template field identification unit comprises: a template file acquisition unit that acquires a template file used in a non-linked template environment or template access information for accessing said template file; a template structure conversion unit that converts said template file into hierarchical structure data used for field identification based on the file format and document structure of the template file acquired by the template file acquisition unit; an input field detection unit that detects an inputable object from the hierarchical structure data converted by the template structure conversion unit into which a user input value can be inserted or mapped to an external input value, and identifies it as an input target field; and a field structure information generation unit that generates field structure information corresponding to said input target field by matching at least one of the field identification value, field location information, field attribute information, and inter-field relationship information of the input target field identified by the input field detection unit. Claim 3 A system for automatically generating instructions based on field context analysis in a non-linked template environment, wherein, in claim 1, the field context extraction unit comprises: a context range setting unit that sets a context extraction range corresponding to the input target field within the template file based on field structure information generated by the template field identification unit; a context element extraction unit that extracts context element information used for interpretation of the input target field from the context extraction range set by the context range setting unit; and a context information configuration unit that corresponds the context element information extracted by the context element extraction unit with the input target field to form context information for field interpretation. Claim 4 In claim 3, the context range setting unit sets the context extraction range to include at least one of a hierarchically related object, a positionally adjacent object, and an attributeally connected object based on the input target field, based on at least one of the field location information and inter-field relationship information included in the field structure information generated by the template field identification unit. Claim 5 A system for automatically generating instructions based on field context analysis in a non-linked template environment, wherein, in claim 3, the context element extraction unit extracts at least one of the following: a hierarchical relationship between objects included in the context extraction range set by the context range setting unit, a relationship between surrounding input objects, a membership area information, display text, an input guide attribute, an accessibility attribute, annotation information, a semantic unit token included in a field identification value, and a semantic unit token included in field attribute information, and converts the extracted context element information into semantic candidate information used for determining field semantics. Claim 6 delete Claim 7 A system for automatically generating writing instructions based on field context analysis in a non-linked template environment, characterized in that, in claim 1, it further includes a structured output generation unit that generates structured output data corresponding to the input target field from the input source data using the writing instructions generated by the instruction generation unit. Claim 8 In claim 7, the structured output generation unit comprises: a template schema generation unit that generates template schema information including at least one of a field identification value for each input target field, an input meaning, an input condition, and a creation instruction by correlating a creation instruction generated by the instruction generation unit with field structure information generated by the template field identification unit; an input source data acquisition unit that acquires input source data to be processed together with the template schema information generated by the template schema generation unit; a field value generation unit that generates an output value for each input target field from the input source data acquired by the input source data acquisition unit based on the creation instruction included in the template schema information generated by the template schema generation unit; and a structured data generation unit that generates structured output data by correlating the output value for each input target field generated by the field value generation unit with the field identification value of the input target field. Claim 9 In claim 8, the input source data acquisition unit acquires record data generated by speech recognition processing as the input source data, and the field value generation unit comprises: a record data division unit that divides the record data acquired by the input source data acquisition unit into utterance content units or sentence units to generate record unit data; a field correspondence relationship calculation unit that calculates a correspondence relationship between the record unit data and the input target field by comparing the record unit data generated by the record data division unit with the writing instructions included in the template schema information generated by the template schema generation unit; and an output value generation unit that generates output values ​​for each input target field from the record unit data based on the correspondence relationship calculated by the field correspondence relationship calculation unit. Claim 10 A system for automatically generating instructions based on field context analysis in a non-linked template environment, characterized in that, in claim 1, it further includes a field semantic correction unit for correcting the accuracy of context information for field interpretation extracted from the field context extraction unit or the writing instruction generated from the instruction generation unit. Claim 11 In claim 10, the field semantic correction unit comprises: a semantic candidate value acquisition unit that acquires semantic candidate values ​​used for determining the semantics of an input target field from context information for field interpretation extracted by the field context extraction unit; a domain semantic matching unit that calculates domain semantic information corresponding to semantic candidate values ​​by comparing the semantic candidate values ​​acquired by the semantic candidate value acquisition unit with pre-set domain term information; a semantic consistency judgment unit that determines the semantic consistency of the input target field by comparing the domain semantic information calculated by the domain semantic matching unit with inter-field relationship information included in the context information for field interpretation extracted by the field context extraction unit; and a field interpretation correction unit that corrects at least one of the input semantics, input conditions, and writing instructions of the input target field according to the judgment result of the semantic consistency judgment unit. Claim 12 In claim 11, the semantic consistency determination unit compares the domain semantic information calculated by the domain semantic matching unit with the field-to-field relationship information included in the field interpretation context information extracted by the field context extraction unit, and determines at least one of the arrangement relationship of input semantics, the mutual correspondence relationship of input conditions, and whether there is duplication of field semantics among a plurality of input target fields within the same context extraction range, characterized in that it is an automatic instruction generation system based on field context analysis in a non-linked template environment.

Citation Information

Patent Citations

  • Voice based data generation method

    KR1020220001852A

  • Control method of electronic device, electronic device, system, and computer program for generating virtual electronic medical record information for education based on actual patient electronic medical record information

    KR1020260014297A

  • Apparatus and method for automatic structured document conversion

    KR1020260046004A

  • A medical record generation display device using AI-based voice recognition technology for ophthalmology and its operation method

    KR102818821B1

  • Ai orchestration-based automated electronic medical record generation system through integration of clinical data

    KR102940957B1