Performance evaluation scheme document configuration parameter extraction method and device, electronic equipment and storage medium

Abstract

This invention provides a method, apparatus, electronic device, and storage medium for extracting configuration parameters from performance appraisal scheme documents, relating to the field of data processing technology. The method significantly improves the accuracy and reliability of extracting configuration parameters from performance appraisal documents by introducing a hybrid intelligent extraction mechanism of "rule-guided and large-model collaboration," combined with a strong constraint verification system of "precise evidence backlink" and "conflict resolution / automatic correction," while reducing the cost of manual review. Through unified coordinate mapping and evidence backlink, precise traceability and auditing of the configuration parameter source are achieved, enhancing the compliance of enterprise-level applications. Simultaneously, the conflict resolution and automatic correction mechanisms ensure the consistency and rationality of the output configuration parameters in business logic, thereby meeting enterprise-level requirements for high reliability and strong interpretability while ensuring high-efficiency automation.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, electronic device, and storage medium for extracting configuration parameters from performance appraisal scheme documents. Background Technology

[0002] Existing performance / appraisal systems typically require HR to manually transcribe information from the appraisal policy or plan document (such as indicators, weights, scoring rules, appraisal cycle, scope of appraisal targets, process nodes, and permission allocation) into the system configuration before initiating an appraisal. This process has the following problems: Manual configuration is time-consuming: it requires repeatedly flipping through documents, aligning tables, breaking down rules, and entering data into multiple pages, which is inefficient and involves a lot of repetitive work.

[0003] High risk of configuration errors: Common errors include inconsistent weight summation, mismatched rating levels, missing key nodes in the process, and omissions / incorrect scope of objects, which directly affect performance results and pose high business risks.

[0004] Lack of explainability and auditability: The system configuration is disconnected from the original source, making it impossible to quickly trace "where a certain configuration field comes from" during review, resulting in high audit and dispute resolution costs.

[0005] The input documents are complex and non-standard in format: Word / PDF / Lark documents / Excel are mixed together, and the writing structure of different HRs in the same company is inconsistent, resulting in poor generalization ability of traditional rule extraction.

[0006] Pure generative extraction is uncontrollable: relying solely on large models for generation can easily lead to "reasonable but incorrect" configurations. It lacks strong constraints and verification systems, making it difficult to meet enterprise-level reliability requirements.

[0007] In summary, the manual transcription of configuration parameters in traditional performance appraisal scheme documents suffers from technical problems such as low efficiency, high error rate, and difficulty in auditing and tracing. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a method, apparatus, electronic device and storage medium for extracting configuration parameters of performance appraisal scheme documents, so as to alleviate the technical problems of low efficiency, high error rate and difficulty in audit traceability of traditional manual transcription of configuration parameters of performance appraisal scheme documents.

[0009] In a first aspect, the present invention provides a method for extracting configuration parameters from a performance appraisal scheme document, comprising: The document analyzes and maps the coordinates of the multi-source heterogeneous performance appraisal scheme document to generate an intermediate representation containing document structure information and unified coordinate information; By intelligently extracting the intermediate representation through rule guidance and collaboration with a large model, and implementing precise evidence backlinking, a verifiable set of configuration parameters with unified coordinate representation is output. The verifiable set of configuration parameters with a unified coordinate representation is subjected to conflict resolution and automatic correction to obtain the target set of configuration parameters.

[0010] Furthermore, the performance appraisal scheme documents from multiple sources and heterogeneous structures are parsed and mapped to coordinates, including: Obtain the performance appraisal scheme document and generate a preprocessed information set of the performance appraisal scheme document, wherein the preprocessed information set includes at least a set of structural anchor points for identifying the logical structure of the document and a set of physical coordinates for recording the physical location of document elements; Based on the set of structural anchor points, the performance appraisal scheme document is parsed to generate a structured intermediate representation containing multiple document elements, wherein each document element is associated with an anchor point in the set of structural anchor points. A unified coordinate representation structure is established, wherein the unified coordinate representation structure includes: canvas identifier, normalized coordinates, logical anchor point identifier, and granularity information; Based on the physical coordinate set and the unified coordinate representation structure, the coordinates in the physical coordinate set are mapped to the unified coordinate representation corresponding to the document elements in the intermediate representation through a mapping function, thereby obtaining an intermediate representation containing document structure information and unified coordinate information.

[0011] Furthermore, the mapping function employs different mapping methods for performance appraisal scheme documents of different formats: Specifically, when the performance appraisal scheme document is in a streaming document format, the document elements in the performance appraisal scheme document are identified as block elements based on the physical coordinate set, and arranged into a block sequence according to the reading order; based on the style topology features of each block element, its virtual height is calculated, and the style topology features include at least one of font size, line spacing, indentation, and heading level; according to the order of the block sequence and the virtual height, the start and end positions of each block element are cumulatively calculated, and the start and end positions are normalized to a predetermined numerical range to serve as a unified coordinate representation of the block element on the virtual layout plane; When the performance appraisal scheme document is in the format of a layout document, the physical boundary frame coordinates recorded in the physical coordinate set are directly normalized, and the normalized coordinates and the page numbers recorded in the physical coordinate set are used as the unified coordinate representation of the corresponding document elements. When the performance appraisal scheme document is in table format, the worksheet and the row and column indices of the target cell or cell range in the performance appraisal scheme document are identified based on the physical coordinate set; the worksheet is mapped to a canvas identifier; the normalized coordinates of the cell or cell range are calculated based on the visible area size of the performance appraisal scheme document and the rendering size of the cell or cell range; and the canvas identifier, row and column indices, and normalized coordinates are used as the unified coordinate representation of the corresponding document element.

[0012] Furthermore, through rule-guided intelligent extraction of the intermediate representation in collaboration with a large model, and by implementing precise evidence backlinking, verifiable configuration parameters with unified coordinate representation are output, including: Based on the intermediate representation, a set of rule configuration candidates is obtained by rule extraction, and a feature vector is generated for each rule configuration candidate in the set. The rule confidence is calculated based on the feature vector, wherein each rule configuration candidate is associated with a position in the original text. Based on the intermediate representation and the rule configuration candidate set, structured information is extracted through a large language model to obtain an enhanced configuration parameter set. For each configuration parameter in the enhanced configuration parameter set, the large language model is constrained to output its corresponding configuration parameter value, evidence identification information indicating the source of the configuration parameter value, the character-level offset position of the configuration parameter value in the original text corresponding to the evidence identification information, the normalized signature of the configuration parameter value, and the model confidence score evaluated by the large language model itself. In the parameter evidence retrieval space determined by the evidence identification information, a matching and positioning algorithm is used to find the corresponding evidence text fragment for the configuration parameter value output by the large language model, and the position of the evidence text fragment is mapped to a unified coordinate representation. In the parameter evidence retrieval space determined by the position of the rule configuration candidate associated with it in the original text, a matching and positioning algorithm is used to find the corresponding evidence text fragment for the configuration parameter value corresponding to the rule configuration candidate, and the position of the evidence text fragment is mapped to a unified coordinate representation, thereby obtaining a verifiable configuration parameter set with a unified coordinate representation.

[0013] Furthermore, using a matching and localization algorithm, corresponding evidence text fragments are found for the configuration parameter values ​​output by the large language model, including: Based on the configuration parameter values, calculate the multi-feature similarity between it and each evidence text fragment in the parameter evidence retrieval space, and recall the Top-K candidate evidence text fragments according to the multi-feature similarity, wherein the multi-feature similarity includes: lexical similarity, semantic similarity and structural prior similarity; For each candidate evidence text fragment, a sliding window mechanism is used to scan its entire text. The optimal matching window is determined by calculating the fine alignment score between the configuration parameter value and the content of each window. The fine alignment score is a weighted sum of multiple similarities, including edit distance similarity, character sequence matching degree, and semantic similarity. If the fine alignment scores corresponding to the optimal matching windows of multiple candidate evidence text fragments all exceed the threshold, then the candidate evidence text fragment with the highest context overlap is selected as the evidence text fragment for the configuration parameter value output by the large language model.

[0014] Furthermore, conflict resolution and automatic correction are performed on the verifiable set of configuration parameters with unified coordinate representation, including: Based on the rule configuration candidate set and its rule confidence, the enhanced configuration parameter set and its model confidence, and the fine alignment score corresponding to each candidate evidence text fragment, the optimal configuration parameter value is selected for the configuration field with conflicting configuration parameter values ​​among the verifiable configuration parameters with uniform coordinate representation, and the conflict-resolved configuration parameter set is obtained. The consistency of the configuration parameter set for conflict resolution is verified based on predefined business logic constraints. When the verification fails, the configuration parameter values ​​of the relevant conflict resolution configuration parameters are automatically repaired based on the principle of minimum modification, so as to obtain the target configuration parameter set.

[0015] Furthermore, the method also includes: An editable configuration draft is generated based on the target configuration parameter set.

[0016] Secondly, the present invention also provides a device for extracting configuration parameters from a performance appraisal scheme document, comprising: The parsing and coordinate mapping unit is used to parse and map the coordinates of multi-source heterogeneous performance appraisal scheme documents, and generate an intermediate representation containing document structure information and unified coordinate information. The intelligent extraction and backlink unit is used to intelligently extract the intermediate representation in collaboration with the large model through rule guidance, and to implement precise evidence backlink, outputting a verifiable set of configuration parameters with unified coordinate representation. The conflict resolution and automatic correction unit is used to resolve conflicts and automatically correct the verifiable configuration parameter set with uniform coordinate representation to obtain the target configuration parameter set.

[0017] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the method described in the first aspect.

[0018] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the method described in the first aspect.

[0019] This invention provides a method for extracting configuration parameters from performance appraisal scheme documents, comprising: parsing and mapping multi-source heterogeneous performance appraisal scheme documents to generate an intermediate representation containing document structure information and unified coordinate information; intelligently extracting the intermediate representation through rule guidance and large model collaboration, and implementing precise evidence backlinking to output a verifiable set of configuration parameters with unified coordinate representation; and resolving conflicts and automatically correcting the verifiable set of configuration parameters with unified coordinate representation to obtain a target set of configuration parameters. As described above, the method for extracting configuration parameters from performance appraisal document documents of the present invention effectively solves the technical problems of low efficiency, high error rate, and uncontrollable pure generative extraction by introducing a hybrid intelligent extraction mechanism of "rule-guided and large-scale model collaboration" and combining it with a strong constraint verification system of "precise evidence backlink" and "conflict resolution / automatic correction". This significantly improves the accuracy and reliability of performance appraisal document configuration parameter extraction and reduces the cost of manual review. Through unified coordinate mapping and evidence backlink, the source of configuration parameters is accurately traced and audited, enhancing the compliance of enterprise-level applications. At the same time, by utilizing conflict resolution and automatic correction mechanisms, the consistency and rationality of output configuration parameters in business logic are ensured. Thus, while ensuring high efficiency and automation, it meets the enterprise-level requirements for high reliability and strong interpretability, alleviating the technical problems of low efficiency, high error rate, and difficult audit traceability of traditional manual transcription of performance appraisal document configuration parameters. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 A flowchart illustrating a method for extracting configuration parameters from a performance appraisal scheme document, as provided in an embodiment of the present invention; Figure 2 A schematic diagram of a device for extracting configuration parameters from a performance appraisal scheme document, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] The manual transcription of configuration parameters in traditional performance appraisal scheme documents is inefficient, has a high error rate, and is difficult to audit and trace.

[0024] Based on this, the method for extracting configuration parameters from performance appraisal documents in this invention introduces a hybrid intelligent extraction mechanism of "rule-guided and large-scale model collaboration," combined with a strong constraint verification system of "precise evidence backlink" and "conflict resolution / automatic correction." This effectively solves the technical problems of low efficiency, high error rate, and uncontrollable pure generative extraction in traditional manual transcription, significantly improving the accuracy and reliability of performance appraisal document configuration parameter extraction and reducing manual review costs. Through unified coordinate mapping and evidence backlink, precise traceability and auditing of the configuration parameter source are achieved, enhancing the compliance of enterprise-level applications. At the same time, the conflict resolution and automatic correction mechanisms ensure the consistency and rationality of the output configuration parameters in business logic, thereby meeting the enterprise-level requirements for high reliability and strong interpretability while ensuring high-efficiency automation.

[0025] To facilitate understanding of this embodiment, a method for extracting configuration parameters of a performance appraisal scheme document disclosed in this embodiment of the invention will first be described in detail.

[0026] Example 1: According to an embodiment of the present invention, an embodiment of a method for extracting configuration parameters of a performance appraisal scheme document is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0027] Figure 1 This is a flowchart of a method for extracting configuration parameters from a performance appraisal scheme document according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps: Step S102: Parse and coordinate map the multi-source heterogeneous performance appraisal scheme documents to generate an intermediate representation containing document structure information and unified coordinate information; Specifically, the constituent document elements in the Intermediate Representation (IR) collectively form a hierarchical structure tree. For example, a section contains multiple paragraphs, and a paragraph may contain a list. Document structure information = the set of all document elements in the IR that describe the logical organization of the document.

[0028] The aforementioned unified coordinate information answers the question, "Where exactly is each document element located?" It focuses on the physical or visual layout of document elements, essentially adding a precise "spatial map" to the document's "skeleton." Unified coordinate information refers to the range information attached to each document element (such as a paragraph or a table cell). Essentially, this range is not the original format coordinates (such as pixel coordinates in a PDF), but rather a normalized and standardized unified coordinate representation (i.e., IRRange). This allows elements from PDF, Word, or Excel to be understood and compared within the same coordinate system.

[0029] The process will be described in more detail below, and will not be repeated here.

[0030] Step S104: Through rule-guided intelligent extraction of intermediate representations in collaboration with large models, and implementation of precise evidence backlinking, output a verifiable set of configuration parameters with unified coordinate representation. Specifically, the verifiable set of configuration parameters with a unified coordinate representation includes: field name, field value, and the corresponding evidence mapping (i.e., unified coordinate representation). In other words, each field in the output configuration parameter set is associated with coordinate information (i.e., a unified coordinate representation). This coordinate information points to the precise location of the field value in the original document.

[0031] When the system outputs a field and its associated IRRange, it has already achieved "original text coordinate mapping". This is because the system or user can use the information in the IRRange to accurately locate and highlight the original text area on the corresponding document canvas (such as a PDF page).

[0032] The process will be described in detail below, and will not be repeated here.

[0033] Step S106: Conflict resolution and automatic correction are performed on the verifiable configuration parameter set with unified coordinate representation to obtain the target configuration parameter set.

[0034] This invention provides a method for extracting configuration parameters from performance appraisal scheme documents, comprising: parsing and mapping multi-source heterogeneous performance appraisal scheme documents to generate an intermediate representation containing document structure information and unified coordinate information; intelligently extracting the intermediate representation through rule guidance and large model collaboration, and implementing precise evidence backlinking to output a verifiable set of configuration parameters with unified coordinate representation; and resolving conflicts and automatically correcting the verifiable set of configuration parameters with unified coordinate representation to obtain a target set of configuration parameters. As described above, the method for extracting configuration parameters from performance appraisal document documents of the present invention effectively solves the technical problems of low efficiency, high error rate, and uncontrollable pure generative extraction by introducing a hybrid intelligent extraction mechanism of "rule-guided and large-scale model collaboration" and combining it with a strong constraint verification system of "precise evidence backlink" and "conflict resolution / automatic correction". This significantly improves the accuracy and reliability of performance appraisal document configuration parameter extraction and reduces the cost of manual review. Through unified coordinate mapping and evidence backlink, the source of configuration parameters is accurately traced and audited, enhancing the compliance of enterprise-level applications. At the same time, by utilizing conflict resolution and automatic correction mechanisms, the consistency and rationality of output configuration parameters in business logic are ensured. Thus, while ensuring high efficiency and automation, it meets the enterprise-level requirements for high reliability and strong interpretability, alleviating the technical problems of low efficiency, high error rate, and difficult audit traceability of traditional manual transcription of performance appraisal document configuration parameters.

[0035] The above provides a brief overview of the method for extracting configuration parameters from the performance appraisal scheme document of the present invention. The specific details involved are described in detail below.

[0036] In an optional embodiment of the present invention, parsing and coordinate mapping of multi-source heterogeneous performance appraisal scheme documents specifically includes the following steps: (1) Obtain the performance appraisal scheme document and generate a preprocessed information set of the performance appraisal scheme document, wherein the preprocessed information set includes at least a set of structural anchor points for identifying the logical structure of the document and a set of physical coordinates for recording the physical location of document elements; Specifically, the aforementioned performance appraisal scheme document can be a multi-source heterogeneous document, such as doc / docx, pdf, online document, xls / xlsx, text pasting, etc. The resulting preprocessed information set includes: doc_id and metadata (document type, page number, paragraph index, timestamp, version number, etc.), providing a stable index for subsequent coordinate normalization and evidence backlink. The system also constructs a document fingerprint for each document object: hash_content, hash_structure, a set of structural anchor points: Anchors={a_k} (title node, paragraph node, table node, cell node, etc.), and a set of physical coordinates: B={b_k} (PDF bbox, Word paragraph layout box, Excel cell range, etc.).

[0037] The above-described document logical structure is the document's "skeleton" or "outline." It describes how the document is composed of different logical parts, such as chapters, headings, paragraphs, lists, and tables. It focuses on organizational relationships rather than visual appearance. It is represented by a "set of structural anchors." Each structural anchor (such as Heading_1, Paragraph_5, Table_A1) is a node in the logical structure, uniquely identifying a logical component of the document.

[0038] Document elements are the basic units that constitute a "structured intermediate representation." They can be understood as the embodiment of logical structure. One structural anchor point corresponds to one document element. The document element contains all the specific information of the logical part represented by that anchor point, such as: text content, style attributes (font, color), and the finally calculated unified coordinate representation.

[0039] (2) Based on the set of structural anchor points, the performance appraisal scheme document is parsed to generate a structured intermediate representation containing multiple document elements, wherein each document element is associated with an anchor point in the set of structural anchor points; Specifically, different parsers are used for different document types, and a unified IR (i.e., intermediate representation, sections / paragraphs / tables / lists / snippets / source_map) is output, while the range is preserved for positioning.

[0040] (3) Establish a unified coordinate representation structure, which includes: canvas identifier, normalized coordinates, logical anchor point identifier and granularity information; Specifically, the aforementioned unified coordinate representation structure is the unified IR coordinate system (IR-Coordinate System). To avoid merely describing "highlights," this invention defines a unified coordinate representation, IRRange, for cross-format evidence location: .

[0041] : Unified display surface (PDF page, Word virtual page, Excel worksheet / view plane); : Normalized coordinates (normalized relative to the width and height of the canvas); Logical anchor points (structural node IDs such as paragraphs / tables / cells); Granularity (section / paragraph / table / cell / char-span).

[0042] (4) Based on the physical coordinate set and the unified coordinate representation structure, the coordinates in the physical coordinate set are mapped to the unified coordinate representation corresponding to the document elements in the intermediate representation through the mapping function, thereby obtaining the intermediate representation containing document structure information and unified coordinate information.

[0043] Specifically, the aforementioned document structure information is a higher-level, collective concept. It refers to all the structured information contained in the final "structured intermediate representation" itself. It includes all document elements, as well as the hierarchy and relationships between these elements (e.g., a paragraph belongs to a chapter). Therefore, document structure information ≈ structured intermediate representation.

[0044] The above mapping function uses different mapping methods for performance appraisal scheme documents of different formats (i.e., for different source documents). physical coordinates () Define the normalized mapping: ): (1) When the performance appraisal scheme document is in the format of a streaming document, the document elements in the performance appraisal scheme document are identified as block elements based on the physical coordinate set, and arranged into a block sequence according to the reading order; the virtual height is calculated based on the style topology features of each block element, including at least one of font size, line spacing, indentation and heading level; the start and end positions of each block element are calculated cumulatively according to the order of the block sequence and the virtual height, and the start and end positions are normalized to a predetermined numerical range as a unified coordinate representation of the block element on the virtual layout plane; Specifically, the aforementioned streaming document formats can be Word or online documents, which typically lack a stable bounding box (bbox). This invention constructs a "virtual layout plane": a. Treat paragraphs / headings / lists, etc., as block elements. And generate a block sequence in the order of reading; b. Assign virtual height to each block element ,in, Includes topological features such as font size, line spacing, indentation, and heading levels; c. Accumulate the start and end positions of the block elements. This is then normalized to ([0,1], i.e., the predetermined numerical range) as a unified coordinate representation of the block element on the virtual layout plane. Therefore, if a finer-grained char-span is needed, character offset mapping is performed within the paragraph text (see S52).

[0045] The virtual height is calculated as follows: First, determine the base height.

[0046] Base height is the starting point for calculations, usually based on the number of lines of text.

[0047] Formula: Base height = Row height × Number of rows; Line height: Usually determined by the font size. A common setting is line height = font size × a coefficient (such as 1.2 or 1.5). For example, for a 12-point font, the line height can be set to 12pt × 1.2 = 14.4pt (virtual unit).

[0048] Line Count: Calculated based on the length of the block's text content and the preset line width (such as the virtual width of the page).

[0049] After determining the base height, calculate the style modification factors.

[0050] This is key to simulating visual importance. Different style features act as multipliers, amplifying or reducing the base height.

[0051] Font-size weight: Objective: Elements with larger font sizes appear more important and should occupy more space.

[0052] Calculation: Font size factor = (current font size / base font size)^α. Where, the base font size is the common font size of the document body (such as 12pt), and α is an exponent greater than 0, used to control the strength of the amplification effect (usually taken as 1 or 2).

[0053] Heading Level Factor: Purpose: The higher the heading level (such as H1), the more important it is in terms of structure and the more visual white space it needs.

[0054] Calculation: A level weight mapping can be preset. For example, {H1:2.0,H2:1.7,H3:1.4,H4:1.2,Text:1.0}.

[0055] Paragraph Spacing Factor: Objective: To simulate the blank spacing before and after paragraphs.

[0056] Calculation: A fixed value or a percentage can be added to the base height. For example, final height = base height × other factors + pre-segment spacing + post-segment spacing.

[0057] Finally, the virtual height is synthesized: The base height is combined with various correction factors to obtain the final virtual height.

[0058] A simplified product model: Virtual height = base height × font size weight factor × heading level factor + paragraph spacing.

[0059] The above start and end positions are calculated as follows: Assume there is a simple document containing three paragraphs, which has been processed into a block sequence ([block1,block2,block3]), and the virtual height (h1,h2,h3) of each block is calculated through style analysis (such as font size, line spacing).

[0060] Initialize the cumulative variable Define a variable =0. This variable is used to track the vertical position (Y-axis coordinate) of the current block's start. It can be understood as the current position of a "pen" on a virtual layout plane.

[0061] Traverse the block sequence and compute The system processes each block element sequentially.

[0062] For the first block (block1): Starting position :that is The current value is 0.

[0063] End position It is the starting position plus the height of the block itself, i.e. +h1.

[0064] Update the pen's position: After the calculation is complete, move the "pen" to the beginning position of the next block, i.e. = .

[0065] For the second block (block2): Starting position ( ): Inherited from the end position of the previous block, i.e. (At this time equals) ).

[0066] End position ( ): +h2.

[0067] Update pen location: = .

[0068] For the third block (block3): And so on After the above steps, the absolute start and end positions of each block in the virtual space are obtained: block1: [ , ] block2: [ , ] block3: [ , ] Position normalization At this point, we have the absolute coordinates of each block, but their values ​​are arbitrary and depend on how the virtual height is calculated. To decouple them from specific dimensions, normalization is required.

[0069] Get the end position of the last block. As the total virtual height of the entire document ( ).

[0070] Divide the absolute start and end positions of each block by this total height. : Normalized coordinates of block1: [ / , / ] Normalized coordinates of block2: [ / , / ] Normalized coordinates of block3: [ / , / ] (2) When the performance appraisal plan document is in the format of a layout document, the coordinates of the physical boundary box recorded in the physical coordinate set are directly normalized, and the normalized coordinates and the page numbers recorded in the physical coordinate set are used as the unified coordinate representation of the corresponding document elements. Specifically, the above-mentioned document format can be PDF, with physical coordinates based on page number and bounding box. ,get , .

[0071] (3) When the performance appraisal scheme document is in the form of a table document, the worksheet in the performance appraisal scheme document and the row and column indices of the target cell or cell range are identified based on the physical coordinate set; the worksheet is mapped to a canvas identifier; the normalized coordinates of the cell or cell range are calculated based on the visible area size of the performance appraisal scheme document and the rendering size of the cell or cell range; the canvas identifier, row and column indices and normalized coordinates are used as the unified coordinate representation of the corresponding document elements.

[0072] Specifically, the above table document can be in Excel format, with the physical coordinates being worksheets, rows, columns, and cell ranges: Or a range of cells. Based on the visible area's width and height (... The cell rendering size estimation position is used to map the cell range to ([0,1]) coordinates; ,granularity If the field comes from a fragment within a cell, then further... Perform character offset positioning within (see S52).

[0073] In an optional embodiment of the present invention, through rule-guided intelligent extraction of intermediate representations in collaboration with large models, and by implementing precise evidence backlinks, verifiable configuration parameters with unified coordinate representations are output, including: (1) Based on the intermediate representation, a set of rule configuration candidates is obtained by rule extraction, and a feature vector is generated for each rule configuration candidate in the set of rule configuration candidates. The rule confidence is calculated based on the feature vector, wherein each rule configuration candidate is associated with a position in the original text. Specifically, rule confidence is calculated by inputting a weighted sum of feature vectors and a configurable weight vector into a sigmoid function.

[0074] The above process is a structure-first candidate extraction (rules / dictionary / templates) to obtain a set of candidate rule configurations.

[0075] Specifically, table / title / regex extraction is performed on the IR (Input / Output) module, outputting a candidate pool (i.e., a set of rule configuration candidates). Each candidate (c) (i.e., a rule configuration candidate) additionally carries a feature vector (for subsequent conflict resolution):

[0076] Rule confidence is defined as:

[0077] in, It is Sigmoid. This is a configurable weight vector.

[0078] in, For candidate feature vectors, For configurable weight vectors, This represents a weighted linear score.

[0079] For the Sigmoid function, , is used to map the score to the ([0,1]) interval as the rule confidence.

[0080] Each feature can be normalized (e.g., min-max or z-score) to avoid scoring bias caused by different units.

[0081] (2) Based on the intermediate representation and the rule configuration candidate set, the structured information is extracted through the big language model to obtain the enhanced configuration parameter set. For each configuration parameter in the enhanced configuration parameter set, the big language model is constrained to output its corresponding configuration parameter value, the evidence identification information indicating the source of the configuration parameter value, the character-level offset position of the configuration parameter value in the original text corresponding to the evidence identification information, the normalized signature of the configuration parameter value, and the model confidence level evaluated by the big language model itself. Specifically, the above process is based on structured constraint extraction from a large model. LLM outputs a schema JSON (cycle_candidates / scope / process_candidates / model_rules / issues) (assessment cycle / assessment scope / assessment process / scoring model and rules / issues and conflicts), and mandates that each field provide [specific details / requirements]. Evidence identification information.

[0082] Added: LLM outputs for each field: Model self-evaluation or results obtained based on logprob / consistency voting. : The character offset ([s,e]) in the snippet / text corresponding to the evidence (set to null if not provided) : Normalized signature of field values ​​(e.g., "70%" → 0.7, "2026Q1" → standard quarterly key).

[0083] (3) In the parameter evidence retrieval space determined by the evidence identification information, the matching and positioning algorithm is used to find the corresponding evidence text fragment for the configuration parameter value output by the large language model, and the position of the evidence text fragment is mapped to a unified coordinate representation. In the parameter evidence retrieval space determined by the position of the rule configuration candidate in the original text, the matching and positioning algorithm is used to find the corresponding evidence text fragment for the configuration parameter value corresponding to the rule configuration candidate, and the position of the evidence text fragment is mapped to a unified coordinate representation, thereby obtaining a verifiable configuration parameter set with a unified coordinate representation.

[0084] Specifically, the mapping steps are as follows: Now, let's see how to map the location of the "evidence text fragment" to the IRRange: Given: The optimal window w* has been found through sliding window matching, and it is known that: The IR element to which w* belongs (e.g., a Paragraph or TableCell).

[0085] w* represents the character-level start offset s and end offset e within the full text of this IR element.

[0086] The mapping steps are as follows: Determine the logical anchor: The IR element (such as a paragraph) to which w* belongs has a unique identifier in the IR structure tree, which is the identifier in IRRange. This is a logical container that locates text fragments.

[0087] Calculate the local offset: Normalize the character offsets [s,e] to the interval [0,1] to obtain u0 and u1 (for linear text). For table cells, two-dimensional normalized coordinates (u0,v0,u1,v1) might be used. This represents the relative position of the fragment within the logical container.

[0088] Projection to Physical Canvas: This is the most crucial step. The system utilizes the layout information acquired during the IR construction phase.

[0089] For PDFs / images: Known The physical bounding box of the corresponding paragraph or cell is (page_num, x0_anchor, y0_anchor, x1_anchor, y1_anchor). Then, based on the normalized offset (u0, v0, u1, v1), the precise physical coordinates of w* within this bounding box can be calculated by linear interpolation.

[0090] The final output combines the canvas ID (such as a PDF page number), the calculated physical coordinates (or normalized coordinates), the logical anchor ID, and the granularity information to assemble the final, unified IRRange.

[0091] The verifiable set of configuration parameters with a unified coordinate representation includes an enhanced set of configuration parameters obtained through a large language model, as well as a set of rule configuration candidates obtained through rule extraction.

[0092] The aforementioned process of evidence backtracking using a matching and location algorithm solves the problem of "where is the evidence?". Subsequent conflict resolution and automatic correction address the questions of "which to believe" and "whether the result is reasonable." The evidence backtracking yields a verifiable set of configuration parameters with a unified coordinate representation, but this set may contain contradictions and flaws. The subsequent conflict resolution and automatic correction serve as the "decision-making brain" and "quality monitoring center" of this solution, responsible for resolving two core issues: Responsibility 1: Conflict resolution (solving the question of "which to believe?"). Scenario: For a field "assessment period", a rule engine might extract "annual assessment" (high rule confidence), while an LLM might extract "first quarter of 2024" (high model confidence). Evidence backlinks can find the precise location of both candidates in the original text (e.g., one in the general principles heading, and the other in the implementation details table).

[0093] Limitations of evidence backlinks: They can prove that both candidates exist in the document, but they cannot determine which one is the correct answer to be used in this assessment.

[0094] In an optional embodiment of the present invention, a matching and localization algorithm is used to find corresponding evidence text fragments for the configuration parameter values ​​output by the large language model, including: (1) Based on the configuration parameter value, calculate its multi-feature similarity with each evidence text fragment in the parameter evidence retrieval space, and recall the Top-K candidate evidence text fragments based on the multi-feature similarity. The multi-feature similarity includes: lexical similarity, semantic similarity and structural prior similarity. (2) For each candidate evidence text fragment, a sliding window mechanism is used to scan its entire text. The optimal matching window is determined by calculating the fine alignment score between the configuration parameter value and the content of each window. The fine alignment score is a weighted sum of multiple similarities among edit distance similarity, character sequence matching degree, and semantic similarity. (3) If the fine alignment scores corresponding to the optimal matching windows of multiple candidate evidence text fragments all exceed the threshold, then the candidate evidence text fragment with the highest context overlap is selected as the evidence text fragment for the configuration parameter value output by the large language model.

[0095] Specifically, evidence back-chain mapping construction (core: precise back-chain algorithm).

[0096] In the original "field path → Based on this, the present invention adds cross-modal mapping and sliding window matching to achieve precise positioning of field-level ranges.

[0097] Step S51: Construct the cross-modal mapping function (F{map}) For each source document ( Extract the set of topological features:

[0098] Where text comes from paragraph / snippet / cell text, and IRRange comes from... .

[0099] Define cross-modal mapping:

[0100] Step S52: Text sliding window matching algorithm Given a text fragment (T) extracted by LLM (which can be a field value or a field value plus context), the system analyzes the candidate evidence set ( Perform two-phase matching on the above: Phase 1: Candidate Recall (Coarse Search) for each Calculate coarse similarity (i.e., multi-feature similarity) (feature space mapping):

[0101] : n-gram Jaccard / BM25 Vector embedding cosine similarity (or LLM embedding) : Structural a priori (header proximity, table column type matching, etc.) Selecting the Top-K candidate set .

[0102] Phase 2: Sliding window fine positioning (fine alignment) for each candidate ,exist Sliding window matching: window length Step size (p). Calculate the fine alignment score for each window segment (w):

[0103] Get the maximum score window w*:

[0104] like Projecting this onto a unified coordinate system: a confidence threshold, a successful match is determined, and the window in that system is output. Character offset in ( ).

[0105]

[0106] For PDF / Excel: it can be further mapped to a bbox sub-region (if there are text rendering coordinates); For Word: Mapped to (Offset within paragraph anchor points) to ensure positioning.

[0107] Multi-candidate decision (context overlap check): If multiple candidate pieces of evidence (e) satisfy... Define the context overlap vector:

[0108] Select The strongest evidence is the final backchain node.

[0109] Output: This makes the evidence backlink no longer "highlight a paragraph", but "field value → precise range → cross-format unified coordinates".

[0110] In the above process, the first level of decision-making (process one): competition within the individual. Decision objective: For a specific candidate evidence element (e.g., paragraph A where the LLM believes the evidence might be located), use a sliding window to scan the entire text of that paragraph and find which window has the highest matching score with the field value for that element.

[0111] Input: The full text of a candidate evidence element.

[0112] Output: The best matching window (w) for this element and its score*.

[0113] For example: Suppose there are two candidate paragraphs (element e1 and e2).

[0114] Within paragraph e1, among all the windows, the highest score is w1*, with a score of 0.9.

[0115] Within paragraph e2, among all the windows, w2* has the highest score, with a score of 0.88.

[0116] Thus, two "local champions" were obtained: (e1,w1*,0.9) and (e2,w2*,0.88).

[0117] Therefore, the process involves a competition "within" each candidate evidence element to select a representative (optimal window) for each element.

[0118] Second-level decision-making (process two): The finals between champions Decision objective: Now, we have a set of "local champions" (the optimal window for each candidate element of evidence). We need to determine a single final champion (i.e., which element the evidence ultimately comes from) from these champions.

[0119] Decision-making logic: Prerequisite check: First, check if the scores of these "local champions" are all very high (all exceeding the preset threshold). This indicates that the candidate evidence provided by LLM may be highly relevant, making it difficult to determine the winner simply by scores.

[0120] Diversion decision: Case A (Non-uniform exceeding the threshold): If one of these optimal windows has a score significantly higher than the others (i.e., it is the only high score exceeding the threshold), then its corresponding candidate evidence element is directly selected. This is the simplified case.

[0121] Scenario B (All Exceeding Thresholds): If the optimal window scores of multiple candidate evidence elements all exceed the threshold (e.g., in the example above, 0.9 and 0.88 both exceed the threshold of 0.85), a "stalemate" occurs. In this case, it is not enough to just look at the matching score; a secondary decision-making mechanism, namely "context overlap verification," needs to be initiated.

[0122] What is "context overlap check"? This is a strategy to break the tie. It argues that if a field value (such as "70%) matches well in multiple paragraphs (such as the "performance indicators paragraph" and the "data description paragraph"), then the paragraph whose context is more consistent with the semantics of the field is more likely to provide evidence.

[0123] Continuing with the previous example: the field is "weight: 70%".

[0124] The context of candidate e1 (the performance indicator paragraph) might be: "...sales revenue is weighted at 70%, and customer satisfaction is weighted at 30%..." The context of candidate e2 (data description paragraph) might be: "...70% of the above data requires financial confirmation..." Although the string "70%" matches successfully in both paragraphs, the context of e1 (filled with "weight" and "metric") has a higher semantic overlap with the "weight" field. Therefore, the system selects e1 as the final source of evidence. The position of the optimal matching window for the final evidence text fragment is mapped to a unified coordinate representation.

[0125] In an optional embodiment of the present invention, conflict resolution and automatic correction are performed on a verifiable set of configuration parameters with a unified coordinate representation, including: (1) Based on the rule configuration candidate set and its rule confidence, the enhanced configuration parameter set and its model confidence, and the fine alignment score corresponding to each candidate evidence text fragment, select the optimal configuration parameter value for the configuration field with conflicting configuration parameter values ​​in the verifiable configuration parameters with uniform coordinate representation, and obtain the configuration parameter set with conflict resolution. Specifically, the evidence backlink (solving the question "Where is the evidence?") takes the candidate data and its evidence identifiers as input from rule extraction and LLM extraction.

[0126] For example, the rule extraction result: Assessment period: "January 1, 2024 to December 31, 2024" (rule confidence level 0.95) - Evidence ID: para_1 (first paragraph) Sales volume weight: "70%" (rule confidence level 0.90) - Evidence ID: table_1_cell_1_2 (Table 1, row 1, column 2) Customer satisfaction weight: "30%" (rule confidence 0.90) - Evidence ID: table_1_cell_2_2 (Table 1, row 2, column 2) Large model extraction results: Assessment period: "2024" (LLM confidence level 0.98) - Evidence ID: para_1 (first paragraph) Sales revenue weight: "70%" (LLM confidence level 0.99) - Evidence ID: table_1_cell_1_2 Customer satisfaction weight: "30%" (LLM confidence level 0.99) - Evidence ID: table_1_cell_2_2 Evidence back-chain process: Task: Find the precise location (character-level or region-level) of each candidate value in the original text.

[0127] Taking the assessment period field as an example, the backlink process for the LLM candidate "2024": Coarse search: Locate the first segment based on evidence ID para_1.

[0128] Sliding window precise positioning: Perform a sliding window scan on the first paragraph of the full text, "The company's annual performance assessment period is from January 1, 2024 to December 31, 2024.", and match the window content with "2024".

[0129] Calculate the matching score: The window "January 1, 2024" has a high semantic similarity to "2024" but a low character sequence matching degree.

[0130] The window "2024 Year": Highest match! The character sequence matches perfectly, and the semantics are completely consistent. Assume its fine alignment score C_align = 0.99.

[0131] Coordinate projection: The character offset of the matched text fragment "2024" in the paragraph is mapped to a unified IR coordinate system to obtain its precise IRRange.

[0132] The output of the evidence backchain: Each candidate value is assigned precise coordinates. At this point, the system knows: The rule candidate "January 1, 2024 to December 31, 2024" is present in the original text, located at [X1,Y1,X2,Y2].

[0133] The LLM candidate "2024" is also mentioned in the original text, located at [X3,Y3,X4,Y4].

[0134] This concludes the evidence backlink process. It proves that both candidates "exist" in the document, and in precise locations. However, it cannot determine which one is better.

[0135] Step 2: Conflict resolution and automatic error correction output repair and consistency verification (solving the questions "Which should be believed?" and "Is the result reasonable?") The input for conflict resolution and automatic correction is the candidate set with precise coordinates and evidence matching scores output by the evidence backlink.

[0136] S61: Conflict Resolution (Which to believe?) Scenario: For the assessment period field, there are now two candidates with high confidence, but their values ​​are different.

[0137] S61 overall confidence level calculation: Rule candidate: S_rule = (rule confidence 0.95 + evidence matching score 0.8) / 2 ≈ 0.875 (assuming the matching score is slightly lower because the original text is the complete date) LLM candidate: S_llm = (LLM confidence score 0.98 + evidence matching score 0.99) / 2 = 0.985 Decision: The overall confidence score of the LLM candidate (0.985) is higher than that of the rule candidate (0.875). More importantly, the evidence matching score (0.99) provided by the evidence backlink plays a crucial role, strongly proving that the expression "2024" does indeed exist precisely in the original text. Therefore, S61 determines that the LLM candidate "2024" is the final value.

[0138] Meanwhile, for the two weight fields, the rules and LLM results were consistent, and the evidence matching scores were both very high, so they were directly adopted.

[0139] For multiple candidates for the same configuration field, the overall confidence score is a weighted combination of rule confidence, model confidence, and evidence alignment confidence (fine alignment score). Furthermore, when there are logical dependencies between multiple configuration fields, the field assignment problem is constructed as a graph search problem, with the goal of maximizing the overall confidence score of all fields while satisfying the hard constraints between fields.

[0140] The method also includes: In conflict resolution, if the overall confidence difference between the best candidate and the second-best candidate is less than a confidence difference threshold, conflict information is output and the candidate is added to the list requiring manual confirmation. If the overall confidence of a candidate is lower than a confidence absolute value threshold, it is output as a warning or directly marked as insufficient confidence based on the importance of the field.

[0141] The above process will now be described in more detail: When the same field (f) exists in both the rule candidate set and the LLM candidate set and the values ​​are inconsistent, a joint candidate set is defined: For each candidate (c\in\mathcal{C}(f)), calculate the overall confidence score (a weighted combination of rule confidence, model confidence, and evidence alignment confidence (fine alignment score)).

[0142] Heuristic Search on Candidate Graph selects a consistent solution. If there are dependencies between fields (cycle affects process, scope affects participants, etc.), the candidates are constructed as a graph search problem. State: A set of field assignments (x={(f,c_f)}) Objective: Maximize the total score while satisfying hard constraints.

[0143] Use heuristic search (such as Beam Search / A*): Heuristic function: Upper bound score estimation for unassigned fields Convergence condition: The optimal solution within the beam width is stable or reaches the iteration limit (N). Forced confirmation and blocking determination If the difference between the optimal and second-best is small: Output issues.conflict and enter the mandatory confirmation list; like If the confidence level is insufficient, output "blocked" or "warn" (depending on the importance level of the field).

[0144] (2) Perform consistency verification on the configuration parameter set for conflict resolution based on predefined business logic constraints; Specifically, business logic constraints include at least one of the following: A. Consistency constraint of weight allocation: The sum of the weight values ​​of multiple configuration items must be equal to a predetermined total value; B. Consistency of process node timing: The duration and sequence of each node in the process must satisfy topological sorting. C. Range resolvability constraint: The organizational entities specified in the configuration item must be able to be mapped to a non-empty set of entities.

[0145] Automatic correction of weight allocation consistency constraints includes: Calculate the deviation between the sum of the current weight set and the predetermined total value; if the deviation is less than the first threshold, perform normalization repair; if the deviation is greater than the first threshold, model it as a constrained programming problem and solve it with the objective of minimizing the weighted change of the initial value, where the weights are determined by the comprehensive confidence of each configuration item.

[0146] (3) When the verification fails, the configuration parameter values ​​of the relevant conflict resolution configuration parameters are automatically repaired based on the principle of minimum modification to obtain the target configuration parameter set.

[0147] Specifically, in the principle of minimum modification, the modification weight is determined based on the comprehensive confidence of the configuration parameters for each conflict resolution.

[0148] This process automatically corrects for logical consistency in the assessment, including: Consistent modeling and correction of weight allocation Let the dimension set The initial value is obtained. (Potentially missing or not equal to 1 / 100%). Define decision variables. .

[0149] Hard constraints: (or percentage system: () = 100)) Objective function (principle of minimum modification): ,in, This represents the weighted credibility coefficient.

[0150] Automatic normalization trigger logic: First calculate the deviation:

[0151] like (Small deviation threshold), perform normalization repair: And record repair_type=normalize; like However, if the number of missing items is ≤k (the repairable threshold), proceed to constraint solving: Assign variables to the missing terms and solve the above CP / L1 minimization problem; If a feasible solution exists and the minimum cost ≤ τ cost If so, it will automatically repair and record repair_type=cp_solve; If there is no feasible solution or the cost is too high Forced blocking: status=blocked, and outputs conflict evidence and suggestions for manual confirmation.

[0152] Process node SLA consistency modeling and correction Given process nodes (j=1..m), extract the node duration. (days / hours), and the priority relationship between node j and node k (topological constraints).

[0153] Hard constraints: If an overall periodic window exists ,but:

[0154] Soft constraints and repair: If slight negative values ​​or unit inconsistencies occur (such as "0.5 days" and "12 hours"), first perform unit normalization and truncation:

[0155] If the total duration exceeds the period window but the excess ratio is ≤ γ, a scaling correction can be performed (maintaining the relative ratio): And record repair_type=scale_sla; If it exceeds the severity or topological conflict (circles exist, start and end nodes are missing, key roles are missing) blocked.

[0156] Convergence criteria (automatic correction iteration): The system iterates for a maximum of (N) rounds in the order of "normalization / standardization → CP solution → constraint verification"; if the constraint set no longer changes after two consecutive rounds of correction, it is considered to have converged and the final draft is output; if it has not converged, it is blocked and manual confirmation of key fields is required.

[0157] Range Analyzable Constraints Define an organizational entity set OOO (department / sequence / job level / personnel), and a mapping function:

[0158] Constraints: include / exclude must be able to map to a non-empty set of entities; list references must be able to resolve to attachments or indicate missing items. If the mapping is empty and the field importance is high → blocked; if the mapping is empty but the user can add attachments → warning + forced confirmation. In an optional embodiment of the present invention, the method further includes the following steps: An editable configuration draft is generated based on the target configuration parameter set.

[0159] Generate draft_config, which includes configuration fields, evidence_map (including IRRange and score), validation (including constraint failure reasons / repair types), and version, and supports patch editing and diff records.

[0160] The method of the present invention will be described below through examples: Quarterly OKR document extraction Input: docx file "2026Q1 OKR Assessment Plan", which includes: Table: Dimension Weights (Goal Achievement 70%, Values ​​30%) Paragraph: Process: "Self-assessment → Supervisor assessment → Calibration → Result confirmation"; Title: "Applicable to: R&D Center P6-P8, excluding interns".

[0161] IR generation is achieved by identifying the weight table as tables[0] and the periodic segments as snippets[1].

[0162] Candidate Extraction: Extract weight candidates from the "Weight" column in the table header; Large model extraction: Output cycle_candidates[2026Q1], scope, process_candidates, model_rules, and attach evidence_ids to each field; Verifier: Weighted sum = 100% pass; All process nodes have been successfully completed; The scope can be resolved to the organization's access level; The output is draft_config + evidence_map, and the front end can click "70%" to highlight the table cell.

[0163] The implementation process of "evidence alignment verification": For the field / model_rules / dimensions / 0 / weight=70%, the system verifies that it exists in the text of the corresponding cell t_0_cell and records the evidence_span or cell coordinates; if it does not exist, the confidence level is reduced and the system enters the warning or blocked state.

[0164] Conflict cycle candidate processing The document contains both "Annual Performance" and "First Half-Year Assessment," and the system detected a period conflict. Output two cycle_candidates; The issues tag is conflict; The cycle confidence level decreased and it was added to the mandatory confirmation list; The front end displays candidates and highlighted evidence for HR to choose from.

[0165] Add `conflict_group_id="cg_01"` to the `issues[]` section and add the annotation: Evidence A comes from the paragraph below the title of the General Provisions of the System; Evidence B comes from the appendix "Assessment Arrangement Table for the First Half of the Year"; Recommended priority: If the company's general rules and regulations take precedence over the attached schedule, then "Annual Performance" is recommended; otherwise, HR needs to confirm the use of "First Half of the Year".

[0166] The core inventive points of this invention are as follows: 1. Unified IR+range mapping mechanism for heterogeneous documents: Enables field-level location-based evidence display; 2. A hybrid extraction framework combining structure-priority candidate extraction and LLM schema constraint extraction: balancing stability and generalization; 3. Forced output of field-level evidence backlinks (evidence_ids): achieving interpretability and auditability; 4. Configure domain validator-driven controllable extraction: reduce the risk of "reasonable but wrong" extraction through verification / correction / candidate management; 5. Human-machine collaborative decision-making linking conflict candidates and confidence levels: transforming "generation" into "selection" in uncertain scenarios.

[0167] 6. Field Dependency Constraint Graph + Constraint Engine Execution Order: Construct a dependency graph (cycle→process→roles→scope, etc.) for performance configuration fields, validate execution in topological order and output the reasons for constraint hit / failure.

[0168] 7. Evidence Alignment: The field output provides not only the evidence_id, but also the coordinates of the evidence span / cell; the system verifies that the field value matches the evidence text to prevent illusions.

[0169] 8. Restricted secondary correction closed loop driven by verification failure: When blocked / warned, corrections are only allowed for failed fields within constraints, and auditable patch and repair logs are generated.

[0170] 9. Candidate Graph Management (Mutual Exclusion / Dependency / Synonym Merge): The candidate pool is not a list, but a graph structure with conflict clusters and dependency edges, used for conflict interpretation and human-machine selection.

[0171] 10. Robust Location Mechanism: When online document versions change, evidence is located based on fragment hash and similarity backtracking to ensure audit traceability.

[0172] The beneficial effects of this invention are as follows: Improved configuration efficiency: Reduces "manual reading + configuration" from hours to minutes (can claim 60%~90% time savings, depending on template reuse rate); Error rate reduction: Through weight / process / scope verification and blocking mechanisms, major configuration errors can be significantly reduced (target: blocking-level errors can be intercepted before they are initiated). Improved auditability: Field evidence backlinks significantly reduce review and audit costs; Improved adaptability: Stable extraction for multi-format documents and non-standard writing habits.

[0173] Evidence location success rate ≥ X% (the percentage of fields that are successfully highlighted after being clicked); Blocking-level error interception rate ≥ X% (blocking configured before deployment); The average number of fields requiring manual confirmation decreased (from N to M); The pass rate (pass percentage) increases with each round of closed-loop correction.

[0174] This invention aims to provide a configuration parameter extraction method based on large-model structured constraints and evidence backlinks. This method enables the system to automatically extract key performance evaluation configuration parameters from heterogeneous performance evaluation scheme documents, generate editable drafts, and provide traceable evidence for each key field. Simultaneously, it significantly reduces configuration error rates and audit costs through verification and correction mechanisms. The approach is as follows: This invention employs a pipeline of "document structured IR → structure-priority candidate extraction → large-model structured constraint extraction → evidence backlinks and confidence levels → verification and correction → draft configuration generation".

[0175] To improve enterprise-level controllability and auditability, this invention introduces three types of executable algorithm mechanisms into the aforementioned pipeline: 1. Multi-source heterogeneous document coordinate normalization and evidence precise backlink algorithm: Map the original physical coordinates of PDF / Word / Excel to a unified IR coordinate system, and achieve precise alignment of "field value - original text range" through sliding window matching; 2. Conflict resolution algorithm combining heuristic rules and LLM: When there is a discrepancy between the "rule candidate pool" and the "LLM extraction result", the optimal consistent solution is selected based on weighted confidence and heuristic search. 3. Automatic correction algorithm for assessment logic consistency based on constraint programming: Mathematical modeling of weights and process SLAs is performed to trigger automatic repair or forced blocking, thereby improving the accuracy of extraction and the degree of system automation.

[0176] Example 2: This invention also provides a device for extracting configuration parameters from a performance appraisal scheme document. This device is mainly used to execute the method for extracting configuration parameters from a performance appraisal scheme document provided in Embodiment 1 of this invention. The following is a detailed description of the device for extracting configuration parameters from a performance appraisal scheme document provided in this invention.

[0177] Figure 2 This is a schematic diagram of a device for extracting configuration parameters from a performance appraisal scheme document according to an embodiment of the present invention, as shown below. Figure 2 As shown, the device mainly includes: a resolution and coordinate mapping unit 10, an intelligent extraction and backlink unit 20, and a conflict resolution and automatic correction unit 30, wherein: The parsing and coordinate mapping unit is used to parse and map the coordinates of multi-source heterogeneous performance appraisal scheme documents, and generate an intermediate representation containing document structure information and unified coordinate information. The intelligent extraction and backlink unit is used to intelligently extract intermediate representations in collaboration with large models, and to implement precise evidence backlinking, outputting a verifiable set of configuration parameters with unified coordinate representation. The conflict resolution and automatic correction unit is used to resolve conflicts and automatically correct deviations in a verifiable set of configuration parameters with a unified coordinate representation to obtain the target set of configuration parameters.

[0178] This invention provides a device for extracting configuration parameters from performance appraisal scheme documents, comprising: parsing and mapping multi-source heterogeneous performance appraisal scheme documents to generate an intermediate representation containing document structure information and unified coordinate information; intelligently extracting the intermediate representation through rule guidance and large model collaboration, and implementing precise evidence backlinking to output a verifiable set of configuration parameters with unified coordinate representation; and resolving conflicts and automatically correcting the verifiable set of configuration parameters with unified coordinate representation to obtain a target set of configuration parameters. As described above, the performance appraisal scheme document configuration parameter extraction device of the present invention, by introducing a hybrid intelligent extraction mechanism of "rule guidance and large model collaboration" and combining a strong constraint verification system of "precise evidence backlink" and "conflict resolution / automatic correction", can effectively solve the technical problems of low efficiency, high error rate and uncontrollable pure generation extraction in traditional manual transcription. It significantly improves the accuracy and reliability of performance appraisal document configuration parameter extraction and reduces the cost of manual review. Through unified coordinate mapping and evidence backlink, it realizes accurate traceability and auditing of the source of configuration parameters, enhancing the compliance of enterprise-level applications. At the same time, by using conflict resolution and automatic correction mechanisms, it ensures the consistency and rationality of output configuration parameters in business logic. Thus, while ensuring high efficiency and automation, it meets the enterprise-level requirements for high reliability and strong interpretability, alleviating the technical problems of low efficiency, high error rate and difficult auditing and traceability of traditional manual transcription of performance appraisal scheme document configuration parameters.

[0179] Optionally, the parsing and coordinate mapping unit is further configured to: acquire the performance appraisal scheme document and generate a preprocessed information set of the performance appraisal scheme document, wherein the preprocessed information set includes at least a set of structural anchor points for identifying the logical structure of the document and a set of physical coordinates for recording the physical location of document elements; based on the set of structural anchor points, parse the performance appraisal scheme document to generate a structured intermediate representation containing multiple document elements, wherein each document element is associated with an anchor point in the set of structural anchor points; establish a unified coordinate representation structure, wherein the unified coordinate representation structure includes: canvas identifier, normalized coordinates, logical anchor point identifier, and granularity information; based on the physical coordinate set and the unified coordinate representation structure, map the coordinates in the physical coordinate set to a unified coordinate representation corresponding to the document elements in the intermediate representation through a mapping function, thereby obtaining an intermediate representation containing document structure information and unified coordinate information.

[0180] Optionally, the parsing and coordinate mapping unit is also used for: when the performance appraisal plan document is in a streaming document format, identifying document elements in the performance appraisal plan document as block elements based on the physical coordinate set, and arranging them into a block sequence according to the reading order; calculating the virtual height of each block element based on its style topology features, where style topology features include at least one of font size, line spacing, indentation, and heading level; calculating the start and end positions of each block element cumulatively based on the block sequence order and virtual height, and normalizing the start and end positions to a predetermined numerical range as a unified coordinate representation of the block element on the virtual layout plane; when the performance appraisal plan document is in a layout document format, The physical bounding box coordinates recorded in the physical coordinate set are directly normalized, and the normalized coordinates and the page numbers recorded in the physical coordinate set are used as the unified coordinate representation of the corresponding document elements. When the performance appraisal plan document is in table format, the worksheet in the performance appraisal plan document and the row and column indices of the target cell or cell range are identified based on the physical coordinate set. The worksheet is mapped to a canvas identifier. Based on the visible area size of the performance appraisal plan document and the rendering size of the cell or cell range, the normalized coordinates of the cell or cell range are calculated. The canvas identifier, row and column indices, and normalized coordinates are used as the unified coordinate representation of the corresponding document elements.

[0181] Optionally, the intelligent extraction and backlink unit is also used to: obtain a rule configuration candidate set based on the intermediate representation through rule extraction, generate a feature vector for each rule configuration candidate in the rule configuration candidate set, and calculate the rule confidence based on the feature vector, wherein each rule configuration candidate is associated with a position in the original text; and extract structured information through a large language model based on the intermediate representation and the rule configuration candidate set to obtain an enhanced configuration parameter set, wherein for each configuration parameter in the enhanced configuration parameter set, the large language model is constrained to output its corresponding configuration parameter value, evidence identification information indicating the source of the configuration parameter value, and the character-level representation of the configuration parameter value in the original text corresponding to the evidence identification information. The algorithm uses the offset position, the normalized signature of the configuration parameter value, and the model confidence evaluated by the large language model itself. In the parameter evidence retrieval space determined by the evidence identification information, a matching localization algorithm is used to find the corresponding evidence text fragment for the configuration parameter value output by the large language model, and the position of the evidence text fragment is mapped to a unified coordinate representation. In the parameter evidence retrieval space determined by the position of the rule configuration candidate in the original text, a matching localization algorithm is used to find the corresponding evidence text fragment for the configuration parameter value corresponding to the rule configuration candidate, and the position of the evidence text fragment is mapped to a unified coordinate representation, thereby obtaining a verifiable set of configuration parameters with a unified coordinate representation.

[0182] Optionally, the intelligent extraction and backlink unit is also used to: calculate the multi-feature similarity between the configuration parameter value and each evidence text fragment in the parameter evidence retrieval space, and recall the Top-K candidate evidence text fragments based on the multi-feature similarity, wherein the multi-feature similarity includes: lexical similarity, semantic similarity and structural prior similarity; for each candidate evidence text fragment, a sliding window mechanism is used to scan its entire text, and the optimal matching window is determined by calculating the fine alignment score between the configuration parameter value and the content of each window, wherein the fine alignment score is the weighted sum of multiple similarities among edit distance similarity, character sequence matching degree and semantic similarity; if the fine alignment scores corresponding to the optimal matching windows of multiple candidate evidence text fragments all exceed the threshold, then the candidate evidence text fragment with the highest context overlap is selected as the evidence text fragment of the configuration parameter value output by the large language model.

[0183] Optionally, the conflict resolution and automatic correction unit is also used to: select the optimal configuration parameter value for configuration fields with conflicting configuration parameter values ​​among verifiable configuration parameters with unified coordinate representation based on the rule configuration candidate set and its rule confidence, the enhanced configuration parameter set and its model confidence, and the fine alignment score corresponding to each candidate evidence text fragment, to obtain the conflict-resolved configuration parameter set; perform consistency verification on the conflict-resolved configuration parameter set based on predefined business logic constraints; and when the verification fails, automatically repair the configuration parameter values ​​of the relevant conflict-resolved configuration parameters based on the principle of minimum modification to obtain the target configuration parameter set.

[0184] Optionally, the device is also used to: generate an editable configuration draft based on a target configuration parameter set.

[0185] The device provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0186] like Figure 3 As shown in the embodiment of this application, an electronic device 600 includes a processor 601, a memory 602, and a bus. The memory 602 stores machine-readable instructions executable by the processor 601. When the electronic device is running, the processor 601 communicates with the memory 602 via the bus. The processor 601 executes the machine-readable instructions to perform the steps of the above-described method for extracting configuration parameters from a performance appraisal scheme document.

[0187] Specifically, the memory 602 and processor 601 mentioned above can be general-purpose memory and processor, without any specific limitations. When the processor 601 runs the computer program stored in the memory 602, it can execute the above-mentioned method for extracting configuration parameters of the performance appraisal scheme document.

[0188] The processor 601 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 601 or by instructions in software form. The processor 601 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 602, and processor 601 reads the information from memory 602 and, in conjunction with its hardware, completes the steps of the above method.

[0189] Corresponding to the above-mentioned method for extracting configuration parameters of performance appraisal scheme documents, this application embodiment also provides a computer-readable storage medium storing machine-executable instructions. When the machine-executable instructions are called and run by a processor, the machine-executable instructions cause the processor to perform the steps of the above-mentioned method for extracting configuration parameters of performance appraisal scheme documents.

[0190] The device for extracting configuration parameters from the performance appraisal scheme document provided in this application embodiment can be specific hardware on the device or software or firmware installed on the device. The device provided in this application embodiment has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0191] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0192] For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0193] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0194] In addition, the functional units in the embodiments provided in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0195] If the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method for extracting configuration parameters of the performance evaluation scheme document described in various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0196] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0197] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for extracting configuration parameters from a performance appraisal scheme document, characterized in that, include: The document analyzes and maps the coordinates of the multi-source heterogeneous performance appraisal scheme document to generate an intermediate representation containing document structure information and unified coordinate information; By intelligently extracting the intermediate representation through rule guidance and collaboration with a large model, and implementing precise evidence backlinking, a verifiable set of configuration parameters with unified coordinate representation is output. The verifiable set of configuration parameters with a unified coordinate representation is subjected to conflict resolution and automatic correction to obtain the target set of configuration parameters. Specifically, through rule-guided intelligent extraction of the intermediate representation in collaboration with a large model, and by implementing precise evidence backlinking, verifiable configuration parameters with unified coordinate representation are output, including: Based on the intermediate representation, a set of rule configuration candidates is obtained by rule extraction, and a feature vector is generated for each rule configuration candidate in the set. The rule confidence is calculated based on the feature vector, wherein each rule configuration candidate is associated with a position in the original text. Based on the intermediate representation and the rule configuration candidate set, structured information is extracted through a large language model to obtain an enhanced configuration parameter set. For each configuration parameter in the enhanced configuration parameter set, the large language model is constrained to output its corresponding configuration parameter value, evidence identification information indicating the source of the configuration parameter value, the character-level offset position of the configuration parameter value in the original text corresponding to the evidence identification information, the normalized signature of the configuration parameter value, and the model confidence score evaluated by the large language model itself. In the parameter evidence retrieval space determined by the evidence identification information, a matching and positioning algorithm is used to find the corresponding evidence text fragment for the configuration parameter value output by the large language model, and the position of the evidence text fragment is mapped to a unified coordinate representation. In the parameter evidence retrieval space determined by the position of the rule configuration candidate associated with it in the original text, a matching and positioning algorithm is used to find the corresponding evidence text fragment for the configuration parameter value corresponding to the rule configuration candidate, and the position of the evidence text fragment is mapped to a unified coordinate representation, thereby obtaining a verifiable configuration parameter set with a unified coordinate representation.

2. The method of claim 1, wherein, The document analyzes and maps coordinates the multi-source heterogeneous performance appraisal scheme document, including: Obtain the performance appraisal scheme document and generate a preprocessed information set of the performance appraisal scheme document, wherein the preprocessed information set includes at least a set of structural anchor points for identifying the logical structure of the document and a set of physical coordinates for recording the physical location of document elements; Based on the set of structural anchor points, the performance appraisal scheme document is parsed to generate a structured intermediate representation containing multiple document elements, wherein each document element is associated with an anchor point in the set of structural anchor points. A unified coordinate representation structure is established, wherein the unified coordinate representation structure includes: canvas identifier, normalized coordinates, logical anchor point identifier, and granularity information; Based on the physical coordinate set and the unified coordinate representation structure, the coordinates in the physical coordinate set are mapped to the unified coordinate representation corresponding to the document elements in the intermediate representation through a mapping function, thereby obtaining an intermediate representation containing document structure information and unified coordinate information.

3. The method of claim 2, wherein, The mapping function uses different mapping methods for performance appraisal scheme documents of different formats: Specifically, when the performance appraisal scheme document is in a streaming document format, the document elements in the performance appraisal scheme document are identified as block elements based on the physical coordinate set, and arranged into a block sequence according to the reading order; based on the style topology features of each block element, its virtual height is calculated, and the style topology features include at least one of font size, line spacing, indentation, and heading level; according to the order of the block sequence and the virtual height, the start and end positions of each block element are cumulatively calculated, and the start and end positions are normalized to a predetermined numerical range to serve as a unified coordinate representation of the block element on the virtual layout plane; When the performance appraisal scheme document is in the format of a layout document, the physical boundary frame coordinates recorded in the physical coordinate set are directly normalized, and the normalized coordinates and the page numbers recorded in the physical coordinate set are used as the unified coordinate representation of the corresponding document elements. When the performance appraisal scheme document is in table format, the worksheet and the row and column indices of the target cell or cell range in the performance appraisal scheme document are identified based on the physical coordinate set; the worksheet is mapped to a canvas identifier; the normalized coordinates of the cell or cell range are calculated based on the visible area size of the performance appraisal scheme document and the rendering size of the cell or cell range; and the canvas identifier, row and column indices, and normalized coordinates are used as the unified coordinate representation of the corresponding document element.

4. The method of claim 1, wherein, Using a matching and localization algorithm, corresponding evidence text fragments are found for the configuration parameter values ​​output by the large language model, including: Based on the configuration parameter values, calculate the multi-feature similarity between it and each evidence text fragment in the parameter evidence retrieval space, and recall the Top-K candidate evidence text fragments according to the multi-feature similarity, wherein the multi-feature similarity includes: lexical similarity, semantic similarity and structural prior similarity; For each candidate evidence text fragment, a sliding window mechanism is used to scan its entire text. The optimal matching window is determined by calculating the fine alignment score between the configuration parameter value and the content of each window. The fine alignment score is a weighted sum of multiple similarities, including edit distance similarity, character sequence matching degree, and semantic similarity. If the fine alignment scores corresponding to the optimal matching windows of multiple candidate evidence text fragments all exceed the threshold, then the candidate evidence text fragment with the highest context overlap is selected as the evidence text fragment for the configuration parameter value output by the large language model.

5. The method of claim 4, wherein, Conflict resolution and automatic correction are performed on the verifiable set of configuration parameters with unified coordinate representation, including: Based on the rule configuration candidate set and its rule confidence, the enhanced configuration parameter set and its model confidence, and the fine alignment score corresponding to each candidate evidence text fragment, the optimal configuration parameter value is selected for the configuration field with conflicting configuration parameter values ​​among the verifiable configuration parameters with uniform coordinate representation, and the conflict-resolved configuration parameter set is obtained. The consistency of the configuration parameter set for conflict resolution is verified based on predefined business logic constraints. When the verification fails, the configuration parameter values ​​of the relevant conflict resolution configuration parameters are automatically repaired based on the principle of minimum modification, so as to obtain the target configuration parameter set.

6. The method of claim 1, wherein, The method further includes: An editable configuration draft is generated based on the target configuration parameter set.

7. An extraction apparatus of a performance evaluation scheme document configuration parameter, characterized by, include: The parsing and coordinate mapping unit is used to parse and map the coordinates of multi-source heterogeneous performance appraisal scheme documents, and generate an intermediate representation containing document structure information and unified coordinate information; The intelligent extraction and backlink unit is used to intelligently extract the intermediate representation in collaboration with the large model through rule guidance, and to implement precise evidence backlink, outputting a verifiable set of configuration parameters with unified coordinate representation. The conflict resolution and automatic correction unit is used to resolve conflicts and automatically correct the verifiable configuration parameter set with uniform coordinate representation to obtain the target configuration parameter set. The intelligent extraction and backlink unit is further configured to: obtain a rule configuration candidate set based on the intermediate representation through rule extraction, generate a feature vector for each rule configuration candidate in the rule configuration candidate set, and calculate the rule confidence based on the feature vector, wherein each rule configuration candidate is associated with a position in the original text; and extract structured information through a large language model based on the intermediate representation and the rule configuration candidate set to obtain an enhanced configuration parameter set, wherein for each configuration parameter in the enhanced configuration parameter set, the large language model is constrained to output its corresponding configuration parameter value, evidence identification information indicating the source of the configuration parameter value, and the position of the configuration parameter value in the original text corresponding to the evidence identification information. The system uses the character-level offset position, the normalized signature of the configuration parameter value, and the model confidence score evaluated by the large language model itself. Within the parameter evidence retrieval space determined by the evidence identification information, a matching and localization algorithm is used to find the corresponding evidence text fragment for the configuration parameter value output by the large language model, and the position of the evidence text fragment is mapped to a unified coordinate representation. Furthermore, within the parameter evidence retrieval space determined by the position of the rule configuration candidate associated with it in the original text, a matching and localization algorithm is used to find the corresponding evidence text fragment for the configuration parameter value corresponding to the rule configuration candidate, and the position of the evidence text fragment is mapped to a unified coordinate representation, thereby obtaining a verifiable set of configuration parameters with a unified coordinate representation.

8. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to perform the method of any one of claims 1 to 6.

Images