An index consistency comparison method and device and a computer storage medium

Abstract

The application provides an index consistency comparison method, an index consistency comparison device and a computer storage medium. The comparison method comprises the following steps: determining document index data of a to-be-verified document; constructing a document multi-dimensional feature vector according to the document index data; screening a first candidate public index set by using a matching result between a document index name in the document index data and a public index name of public index data; wherein the first candidate public index set comprises a plurality of first candidate public index data; calculating a comprehensive matching score by using a document multi-dimensional feature vector of the document index data and a public multi-dimensional feature vector of the first candidate public index data; determining matching public index data of the to-be-verified document according to the comprehensive matching score; and generating a consistency comparison report based on the document index data and the matching public index data. The knowledge base of the related official documents is used to determine whether the related data indexes are consistent according to semantics, and a measurement standard of the data indexes is provided.

Description

Technical Field

[0001] This application relates to the field of data verification technology, and in particular to a method for comparing index consistency, an apparatus for comparing index consistency, and a computer storage medium. Background Technology

[0002] In large enterprises such as some central and state-owned enterprises, there is a need for consistency in the reporting of certain data indicators, especially statistical and economic data, when reporting news to the outside world. However, for large enterprises, there may be differences in their historical data, statistical standards, and data from different business segments. This can easily lead to contradictions when communicating with the public due to different sources of information. Relying solely on manual review and proofreading is not only inefficient but also introduces uncontrollable human risks.

[0003] With the development of information technology, automated verification tools based on rules and keyword matching have emerged. These tools typically pre-set keyword or regular expression patterns, extract suspected indicator data from documents, and compare them with structured databases. However, this approach has significant limitations: First, indicator descriptions are flexible and varied (e.g., "Gross Domestic Product" may be expressed as "GDP," "Gross National Product," etc.), and are often embedded in complex contexts, resulting in low recall and high false positive rates with simple keyword matching. Second, they cannot deeply understand the statistical scope of indicators (e.g., whether "annual operating revenue" includes subsidiaries, the base period definition of "quarter-on-quarter growth"), and inconsistent scope is a common cause of data misuse. Third, they lack the ability to intelligently assess the timeliness and authority of data sources. Summary of the Invention

[0004] To address the aforementioned technical problems, this application proposes an indicator consistency comparison method, an indicator consistency comparison device, and a computer storage medium.

[0005] To address the aforementioned technical problems, this application proposes a method for comparing indicator consistency, the method comprising: Extract the publicly available indicator data from the knowledge base and the publicly available multi-dimensional feature vectors of the publicly available indicator data; Determine the manuscript metrics data for the document to be verified; Construct a multi-dimensional feature vector for the manuscript based on the manuscript index data; A first candidate set of public indicators is selected by matching the names of the manuscript indicators in the manuscript indicator data with the names of the public indicators in the public indicator data; wherein, the first candidate set of public indicators includes several first candidate public indicator data. A comprehensive matching score is calculated using the multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data. The matching public index data of the manuscript to be verified is determined based on the comprehensive matching score; A consistency comparison report is generated based on the document index data and the matching public index data.

[0006] The step of filtering out the first candidate set of public indicators by using the matching results between the manuscript indicator names in the manuscript indicator data and the public indicator names in the public indicator data includes: Calculate the first cosine similarity between the manuscript indicator names in the manuscript indicator data and the public indicator names in the public indicator data; The first candidate public index set is constructed based on the public index data whose first cosine similarity is higher than the dynamic similarity threshold; The dynamic similarity threshold is dynamically adjusted based on the historical matching accuracy of the domain to which the document to be verified belongs.

[0007] The step of calculating a comprehensive matching score using the multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data includes: Calculate the second cosine similarity between the manuscript multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data; Calculate the attention score between the manuscript index data and the first candidate public index data; The comprehensive matching score is calculated based on the second cosine similarity and the attention score.

[0008] The step of calculating the comprehensive matching score based on the second cosine similarity and the attention score includes: Based on the historical matching records of the first candidate publicly available indicator data, a historical consistency score is determined; The comprehensive matching score is calculated based on the historical consistency score, the second cosine similarity, and the attention score.

[0009] The step of determining the historical consistency score based on the historical matching records of the first candidate public indicator data includes: Based on the historical matching records of the first candidate public indicator data, the first candidate public indicator data is determined to be the number of times a match has been determined and the total number of matches. The historical consistency score is determined based on the ratio of the number of determined matches to the total number of matches.

[0010] The method for comparing index consistency further includes, after calculating the comprehensive matching score using the multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data: The first candidate public indicator set is filtered based on the comprehensive matching score to generate a second candidate public indicator set; the second candidate public indicator set includes several second candidate public indicator data. Construct an indicator relationship diagram between the several second candidate public indicator data and the manuscript indicator data; The node representations of the indicator relationship graph are updated and the relationships are inferred using a graph neural network model, so as to update the manuscript indicator data and / or the second candidate public indicator data; The updated manuscript index data and / or the second candidate public index data are used to recalculate the comprehensive matching score, and the matching public index data of the manuscript to be verified is determined.

[0011] The step of constructing the indicator relationship diagram between the plurality of second candidate public indicator data and the manuscript indicator data includes: Nodes are generated based on the second candidate public indicator data and the manuscript indicator data; Construct edges between adjacent nodes using any one or more of the following rules: When the semantic similarity between the manuscript indicator name of the manuscript indicator data and the public indicator name of the second candidate public indicator data is higher than a preset semantic threshold, a similarity relationship edge is added between the two nodes. When the publishing organization of the manuscript index data and the publishing organization of the second candidate public index data are the same publishing organization, add an organization association edge between the two nodes; When the time difference between the publication time of the manuscript index data and the publication time of the second candidate public index data is lower than a preset time difference threshold, a time correlation edge is added between the two nodes. When the manuscript index data and the second candidate public index data appear in the same historical document, add a co-occurrence relationship edge between the two nodes.

[0012] The method for comparing index consistency further includes, after calculating the comprehensive matching score using the multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data: The first candidate public indicator set is filtered based on the comprehensive matching score to generate a third candidate public indicator set; the third candidate public indicator set includes several third candidate public indicator data. Calculate the spatiotemporal perceived confidence of the third candidate public index data and the manuscript index data; The matching public index data of the manuscript to be verified is determined based on the spatiotemporal awareness confidence level. A consistency comparison report is generated based on the manuscript indicator data and the matching public indicator data. The spatiotemporal perception confidence level is determined by the semantic similarity between the document indicator name and the publicly available indicator name, the semantic similarity between the document indicator caliber and the publicly available indicator caliber, and / or the time difference between the public release time and the current time.

[0013] The spatiotemporal perception confidence level is further determined by the authoritative influence factor and / or time-related influence factor of the third candidate publicly available indicator data. The authority impact factor is determined based on the authority ranking of the third candidate public indicator data's public release institution among all public release institutions in the third candidate public indicator set; the time impact factor is determined based on the newness ranking of the public release time of the third candidate public indicator data among all public release times in the third candidate public indicator set.

[0014] The step of generating a consistency comparison report based on the manuscript indicator data and the matching public indicator data includes: The consistency between the manuscript indicator values ​​of the manuscript indicator data and the public indicator values ​​of the matching public indicator data is determined. Based on the consistency determination results, the consistency comparison report is generated.

[0015] To address the aforementioned technical problems, this application proposes an indicator consistency comparison device, which includes a memory and a processor coupled to the memory. The memory is used to store program data, and the processor is used to execute the program data to implement the indicator consistency comparison method as described above.

[0016] To address the aforementioned technical problems, this application proposes a computer storage medium for storing program data, which, when executed by a computer, is used to implement the aforementioned indicator consistency comparison method.

[0017] Compared with existing technologies, the beneficial effects of this application are as follows: This application extracts publicly available indicator data from a knowledge base and the publicly available multi-dimensional feature vectors of the publicly available indicator data; determines the manuscript indicator data of the manuscript to be verified; constructs a multi-dimensional feature vector of the manuscript based on the manuscript indicator data; filters out a first candidate set of publicly available indicators using the matching results between the manuscript indicator names in the manuscript indicator data and the publicly available indicator names of the publicly available indicator data; wherein, the first candidate set of publicly available indicators includes several first candidate sets of publicly available indicator data; calculates a comprehensive matching score using the manuscript multi-dimensional feature vectors of the manuscript indicator data and the publicly available multi-dimensional feature vectors of the first candidate sets of publicly available indicator data; determines the matching publicly available indicator data of the manuscript to be verified based on the comprehensive matching score; and generates a consistency comparison report based on the manuscript indicator data and the matching publicly available indicator data. By constructing a knowledge base database and based on a large language model, this application can achieve the matching between the indicator meanings in the reported manuscripts and the indicator meanings in the documents in the knowledge base database. By constructing relevant data verification rules, it can complete the verification of relevant data indicators in the manuscript to be published, especially the matching of data indicator names and the multi-dimensional accuracy verification of indicator data, which can better assist manual verification and modification. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart illustrating the first embodiment of the indicator consistency comparison method provided in this application; Figure 2 This is a schematic diagram of the intelligent agent index consistency comparison framework provided in this application; Figure 3 This is a flowchart illustrating the second embodiment of the indicator consistency comparison method provided in this application; Figure 4 This is a flowchart illustrating the third embodiment of the indicator consistency comparison method provided in this application; Figure 5 This is a flowchart illustrating the fourth embodiment of the indicator consistency comparison method provided in this application; Figure 6 This is a schematic diagram of an embodiment of the index consistency comparison device provided in this application; Figure 7 This is a schematic diagram of the structure of an embodiment of the computer storage medium provided in this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0020] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a particular order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] In recent years, Large Language Models (LLMs) have demonstrated powerful capabilities in natural language understanding and generation, offering new solutions to the aforementioned problems. Researchers have begun exploring the use of LLMs to extract structured information from unstructured text and perform simple semantic matching. However, existing LLM-based applications still face a series of challenges in achieving high-precision consistency comparisons: 1) Insufficient granularity and precision of semantic matching: Although the semantic similarity calculation of general LLM is better than keyword matching, its matching results are still coarse for highly specialized and numerically sensitive indicator comparisons. It may match semantically related but actually different indicators together (such as "sales revenue" and "sales volume"), while ignoring the core elements that determine the comparability of indicators—the definition of the criteria and the numerical value itself.

[0022] 2) Ignoring multimodal and unstructured information: Indicator data in publicly available literature are often presented in tables, charts, and specific formatting, forming a cohesive whole with surrounding explanatory text. Existing methods typically only process plain text, losing these important structured or visual cues, resulting in incomplete information extraction.

[0023] 3) Lack of deep-level associative reasoning ability: The authenticity of a single indicator often needs to be cross-validated within a network of related indicators. For example, a company's "net profit" growth rate has an inherent logical relationship with its "operating revenue," "gross profit margin," and other indicators. Most existing methods perform isolated indicator matching, lacking a mechanism to mine from a knowledge base and utilize such as the network of relationships between indicators for consistent reasoning.

[0024] 4) Poor system adaptability and optimizability: Different fields (such as macroeconomics, biomedicine, and finance) have vastly different indicator expression habits, data source preferences, and authoritative standards. Existing solutions often have fixed parameters or require manual adjustment, making it impossible to automatically adapt to different scenarios through feedback, and thus difficult to achieve domain-optimal performance while ensuring universality.

[0025] To address this, this application provides a method and system for intelligently comparing the consistency of cited data indicators. This method and system deeply integrates the deep semantic understanding capabilities of a large language model while effectively overcoming the aforementioned limitations, achieving high precision, automation, interpretability, and adaptability. It considers that some central and state-owned enterprises cite internal or publicly available data when reporting news. To prevent confusion for readers and damage to the author's authority, the data indicators need to be consistent in their external reporting. The solution in this application constructs a knowledge base database on an intelligent agent to store relevant official documents. After news reports are uploaded, a large language model performs segment-by-segment comparisons, determining whether relevant data indicators are consistent based on semantics. This allows for the annotation of inconsistent content and provides the reference source for the data.

[0026] Please refer to the details. Figure 1 and Figure 2 , Figure 1 This is a flowchart illustrating the first embodiment of the indicator consistency comparison method provided in this application. Figure 2 This is a schematic diagram of the consistency comparison framework for agent metrics provided in this application.

[0027] It should be noted that the indicator consistency comparison method provided in this application can be applied to an indicator consistency comparison device, which executes the following indicator consistency comparison process through fixed code, or it can be applied to an intelligent agent, such as... Figure 2 As shown, the application layer, algorithm layer, and data layer work together to determine the consistency comparison scheme of execution indicators based on the user's needs.

[0028] like Figure 1 As shown, the indicator consistency comparison method in this application embodiment specifically includes the following steps: Step S11: Extract the public indicator data of the knowledge base and the public multi-dimensional feature vector of the public indicator data.

[0029] In this application embodiment, the indicator data involved in this application are all metadata of the source data, used to describe the data attributes of the manuscript and document, such as publicly available indicator data P. i Including indicator name N i Indicator caliber D i Indicator value V i Data source S i Release time T i and the issuing agency O i wait.

[0030] The construction process of the knowledge base used in this application is as follows: Publicly available document data is multimodally encoded to obtain a unified feature vector, which is then stored in an unstructured vector database; simultaneously, structured publicly available indicator data P is extracted from the publicly available documents. i Stored in a structured database.

[0031] Specifically, the publicly available literature data used to build the knowledge base is not limited to non-textual modal data, but includes data in multiple modalities. The data sources are not limited to plain text, but include, but are not limited to: text (such as text content in PDF, Word, TXT), images (such as scanned documents, screenshots, charts), tables (such as table structures in Excel and PDF), and formatting information (such as paragraph structure and heading levels).

[0032] This application employs multimodal coding to uniformly convert publicly available documents from different formats and presentation forms (such as PDF text, tables, images, and scanned copies) into feature vectors that computers can understand and retrieve, and store them in a vector database for subsequent semantic matching and comparison with the document to be verified.

[0033] This application provides an embodiment of a specific knowledge base construction scheme, the steps of which are broken down as follows: 1. Data Input and Preprocessing Before encoding, the system first needs to parse the original file: Text-based (Word / TXT / PDF text layer): Directly extracts plain text content.

[0034] Tables (Excel / PDF tables): Extract the table structure, preserving the correspondence between rows, columns, headers, and cells.

[0035] Image-based (scanned images / charts): Use OCR (Optical Character Recognition) technology to recognize text in images, and use image recognition technology to recognize chart structures (such as bar chart trends, pie chart percentages).

[0036] 2. Multimodal Feature Extraction This is the process of converting data of different modalities into vectors respectively.

[0037] A. Text semantic encoding (processing plain text, OCR results) Tools: Use pre-trained models such as Large Language Model (LLM) or BERT (Bidirectional Encoder Representations from Transformers).

[0038] Operation: Input the extracted text paragraphs (such as indicator definitions and numerical descriptions) into the model, and output the semantic embedding vector EsemanticEsemantic of that text. This vector captures the meaning and context of the indicator.

[0039] B. Table structure encoding (processing tabular data) Challenge: The indicators in a table are often in the header, while the values ​​are in the cells. Simply converting it to text will lose the row and column correspondence.

[0040] operate: Transform a two-dimensional table structure into a structured sequence (e.g., serialize by row or by column).

[0041] Use a dedicated table encoder (such as TAPAS or a structure-aware Transformer) to encode the correspondence between "header + cell" and generate a structural feature vector to ensure that the system can understand the structural relationship such as "the first column header is 'year' and the second column header is 'GDP'".

[0042] C. Visual layout coding (handling the layout of charts and scanned documents) operate: For images containing charts, use CNN (Convolutional Neural Network) or ViT (Visual Transformer) to extract visual features and identify the text location and color blocks in the image (to distinguish between line charts and bar charts).

[0043] For PDF scans, analyzing the text layout (such as larger font size and center alignment of titles) and generating layout feature vectors helps to understand which text is the indicator name and which is the body description.

[0044] 3. Feature Alignment and Fusion By unifying the vectors from the different sources mentioned above into a single space, a final unified feature vector is formed.

[0045] Alignment: Since text vectors, table vectors, and image vectors may have different dimensions, they need to be mapped to the same dimensional space through a mapping layer (fully connected network).

[0046] Fusion: Following the approach in S31 (although S31 is for manuscript metrics, the principle is equally applicable to document processing), a final vector is generated through weighted fusion:

[0047] The weights here can be dynamically adjusted based on the document type. For example, if the document is primarily plain text, the semantic weights... The weight is higher; if it's a statistical yearbook, the table weight is higher. It's a bit higher.

[0048] 4. Stored in an unstructured vector database Storage content: The generated unified feature vector E is stored in a vector database. Additionally, metadata is stored to allow for future source tracing, such as: Original document ID, page number, and document type; The structured public indicator data P extracted from it i (i.e., indicator name N) i , caliber D i Numerical value V i Source S i These metadata fields (such as vectors) are typically stored in the corresponding fields in a relational database, but are associated with vectors.

[0049] Index structure: The following mention of "adopting a hierarchical index structure" usually refers to algorithms like HNSW (Hierarchical Navigable Small World) or IVF (Inverted Index File) to accelerate subsequent retrieval—when a new manuscript is received, it can quickly find the most similar public document vectors from a massive number of document vectors.

[0050] Step S12: Determine the manuscript metrics data for the manuscript to be verified.

[0051] In this embodiment of the application, the document to be verified is processed, and the document index data Q is extracted. j The manuscript index data Q j Including the document indicator name M j Manuscript Indicators E j Manuscript index value U j and the paragraph R in which it is located j .

[0052] It should be noted that the data extraction scheme for the manuscript indicators of the document to be verified is basically the same as the data extraction scheme for the public indicators of the publicly available documents in step S11, and will not be repeated here.

[0053] Step S13: Construct a multi-dimensional feature vector of the manuscript based on the manuscript index data.

[0054] In this embodiment of the application, for each document indicator data Q j Generates a document with the metric name M. j , caliber E j Numerical U j and context R j The multidimensional feature vector E(Q) of the feature j ).

[0055] Specifically, the multi-dimensional feature vector E(Q) j It can be generated in the following ways: Name the manuscript indicator M j Manuscript Indicators E j and the paragraph R in which it is located j The text input is a large language model, and the semantic embedding vector E is obtained. semantic .

[0056] For the manuscript index value U j Perform numerical feature encoding to obtain the numerical feature vector E. numeric The encoding includes a vectorized representation of the numerical value, unit, and numerical type.

[0057] For the paragraph R j Encode the context structure to obtain the context feature vector E contextual .

[0058] Finally, a weighted fusion method is used to generate the final multi-dimensional feature vector:

[0059] Wherein, α, β, and γ are adjustable fusion weight coefficients, and α + β + γ = 1. The adjustable fusion weight coefficients α, β, and γ are preset and configured according to different indicator types and domain characteristics, and are dynamically optimized through user feedback and reinforcement learning mechanisms.

[0060] Step S14: Filter out a first candidate set of public indicators by matching the names of the manuscript indicators in the manuscript indicator data with the names of the public indicators in the public indicator data; wherein, the first candidate set of public indicators includes several first candidate public indicator data.

[0061] In this embodiment of the application, a two-stage matching method is used to achieve accurate matching of manuscript indicator data and public indicator data. Step S14 achieves the first stage matching, while step S15 achieves the second stage matching.

[0062] Specifically, in the first stage of matching, this application matches the manuscript indicator names in the manuscript indicator data with the public indicator names in the public indicator data. The name matching method is to use cosine similarity calculation to filter out the first candidate public indicator data with similarity higher than the dynamic threshold.

[0063] It should be emphasized that the threshold used for comparing matching cases in this application is a dynamic threshold, and the calculation method for this dynamically adjusted semantic similarity threshold is as follows: First, analyze historical matching records for a specific domain and calculate the historical matching accuracy for that domain. domain The specific domain refers to the domain to which the document to be verified belongs, which can be automatically identified or categorized based on the content of the document.

[0064] Secondly, the similarity threshold is dynamically adjusted based on historical matching accuracy:

[0065] in, The base similarity threshold, λ is the adjustment magnitude coefficient, γ is the accuracy impact coefficient, and Accuracy domain ∈[0,1].

[0066] In one specific implementation, the basic similarity threshold The value range is 0.75-0.90, the adjustment amplitude coefficient λ ranges from 0.05-0.15, and the accuracy influence coefficient γ ranges from 1.0-2.0.

[0067] It is important to emphasize that the dynamic adjustment scheme adopted in this application is mainly reflected in two core aspects: first, the dynamic calculation of the matching threshold, and second, the dynamic scoring of the matching model. It is not a static, unchanging rule, but rather changes in real time based on historical data and input features.

[0068] First-stage matching (coarse screening): Using cosine similarity to quickly scan massive amounts of data. The "dynamic" aspect here lies in its reliance on the dynamic threshold calculated in the first stage. Only those indicators that exceed this dynamic threshold can proceed to the next round.

[0069] Second-stage matching (fine-tuning): After the candidate set is narrowed down, the system launches a more complex attention-based comprehensive scoring model. This is a more refined dynamic calculation process.

[0070] The dynamic adjustment here is reflected in: Input variable: Accuracy in the formula domain(Historical matching accuracy in a specific domain) is a dynamically changing value. If the system's matches in this domain are frequently corrected by users recently (accuracy decreases), this value will decrease.

[0071] Output: If the system has recently performed well in this domain (Accuracy) domain high), It will automatically increase (higher requirements, better to have fewer but better quality systems). If the system has recently performed poorly in this area (Accuracy) domain Low), This will automatically lower the threshold (it's better to recall more candidates and then let the more refined model filter them).

[0072] Step S15: Calculate the comprehensive matching score using the multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data.

[0073] In this embodiment, during the second-stage matching, the application matches the multi-dimensional feature vectors of the manuscript index data and the multi-dimensional feature vectors of the public index data. The feature vector matching can also be done using cosine similarity calculation to determine the comprehensive matching score for each candidate public index data. The higher the comprehensive matching score of the candidate public index data, the more likely it is to be used as the matching public index data for the manuscript to be verified.

[0074] Furthermore, this application also provides a specific scheme for calculating the comprehensive matching score, employing an attention-based comprehensive scoring model to rank and filter candidate indicators. Please continue reading. Figure 3 , Figure 3 This is a flowchart illustrating the second embodiment of the indicator consistency comparison method provided in this application.

[0075] like Figure 3 As shown, the indicator consistency comparison method in this application embodiment specifically includes the following steps: Step S21: Calculate the second cosine similarity between the manuscript multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data.

[0076] In this embodiment, the cosine similarity between the multi-dimensional feature vector of the manuscript index data and the multi-dimensional feature vector of the first candidate public index data is calculated. Compared with the cosine similarity of the index name in the first stage, more information is introduced for matching, which can filter out more accurate candidate public index data.

[0077] Step S22: Calculate the attention score between the manuscript index data and the first candidate public index data.

[0078] In this embodiment, a multi-head attention mechanism layer is constructed to calculate the document metric data Q. j Compared with candidate publicly available indicator data P i Attention score between them.

[0079] Specifically, the manuscript indicator data Q serves as the query, and the candidate public indicator data P... i As keys and values, the three are first projected onto different subspaces through multiple sets of learnable linear transformations, resulting in H head queries. ,key ,value Within each head, scaled dot product attention is computed. With all The similarity scores are then normalized using Softmax to obtain the attention weights. This weight represents P from the h-th viewpoint. i Q j The importance of [the factor / object]. Subsequently, weights were used to [distribute / adjust / adjust]. Weighted summation is performed to obtain the output of each head. The outputs of all heads are concatenated, and then a linear transformation is applied to fuse the multi-view information, generating the final enhanced representation or using it directly as the matching basis. Attention weights. Cross-head aggregation (such as averaging or taking the maximum value) can be used to obtain Q. j With each P i The comprehensive matching score, which integrates multiple relationships such as semantics and structure, is used for the optimized ranking of the candidate set.

[0080] Step S23: Determine the historical consistency score based on the historical matching records of the first candidate public indicator data.

[0081] In this embodiment of the application, the candidate public index data P is queried. i Calculate historical consistency score based on historical matching records. The calculation method can be as follows: Statistical data of publicly available indicators P i The number of times a match was accepted as a positive match in historical matches. Total number of matches Calculate the historical consistency score:

[0082] Step S24: Calculate the comprehensive matching score based on the historical consistency score, the second cosine similarity, and the attention score.

[0083] In this embodiment of the application, by combining the above three scores, the application can calculate a comprehensive matching score for each candidate disclosed indicator data. :

[0084] in, , , The weight parameters are learnable, and + + =1.

[0085] Step S16: Determine the public matching index data of the manuscript to be verified based on the comprehensive matching score.

[0086] In this embodiment of the application, the candidate disclosure index data with the highest comprehensive matching score is selected as the matching disclosure index data of the document to be verified.

[0087] Step S17: Generate a consistency comparison report based on the manuscript indicator data and the matching public indicator data.

[0088] In this embodiment of the application, the numerical value U of the comparative document index is... j Compared with the publicly disclosed indicator value V i A consistency determination is performed, and then a consistency comparison report is generated and displayed. The report includes a list of document indicator data, matching results, highlighted original text, and annotations of matching data. At the same time, user feedback is received, and key parameters in the matching and verification process are dynamically adjusted based on the feedback data through a reinforcement learning strategy.

[0089] Furthermore, this application can also introduce graph structure network analysis of indicator relationship diagrams to screen candidate public indicator data. Therefore, after calculating the comprehensive matching score in the above embodiments, this application also provides a specific scheme for graph structure network analysis. Please continue to refer to... Figure 4 , Figure 4 This is a flowchart illustrating the third embodiment of the indicator consistency comparison method provided in this application.

[0090] like Figure 4 As shown, the indicator consistency comparison method in this application embodiment specifically includes the following steps: Step S31: Filter the first candidate public indicator set according to the comprehensive matching score to generate a second candidate public indicator set; the second candidate public indicator set includes several second candidate public indicator data.

[0091] In this application embodiment, the comprehensive matching score in the above embodiment is used to further screen the first candidate public indicator set. A preset number of candidate public indicator data can be selected from high to low according to the comprehensive matching score to form the second candidate public indicator set, or candidate public indicator data with a comprehensive matching score higher than a preset threshold can be selected to form the second candidate public indicator set.

[0092] Step S32: Construct an indicator relationship diagram between the several second candidate public indicator data and the manuscript indicator data.

[0093] In this embodiment of the application, the application is based on the manuscript index data Q. j And all publicly available indicator data P in the secondary candidate matching set C2 i The graph represents the relationship between nodes. A graph neural network model is used to update the node representations and infer relationships, thereby optimizing the secondary candidate matching set C2.

[0094] Specifically, when constructing the indicator relationship diagram, this application uses the manuscript indicator data Q. j And all publicly available indicator data P in the secondary candidate matching set C2 i For each node, construct edges between nodes according to the following rules: (1) If the semantic similarity between two indicator names is higher than the preset threshold, add a similarity relationship edge.

[0095] (2) If two indicators belong to the same issuing organization, add an organization association edge.

[0096] (3) If the two indicators are close in time, add a time-related edge.

[0097] (4) If two indicators frequently appear together in historical documents, add a co-occurrence relation edge.

[0098] Step S33: Use a graph neural network model to update the node representations and perform relational reasoning on the indicator relationship graph, so as to update the manuscript indicator data and / or the second candidate public indicator data.

[0099] In this embodiment, in the initial stage of the indicator relationship graph, the initial representation of each node can be a multi-dimensional feature vector to describe the initial features of that node. Then, this application uses a graph neural network model to update and infer the indicator relationship graph, including at least the following steps: Message passing: In each layer of GNN (Graph Neural Network), each node collects information about its neighboring nodes.

[0100] Manuscript Indicator Q j It will ask about its neighbors (P connected by similar edges).i "What are your eigenvectors? What are their numerical values? Where is your organization located?" Meanwhile, the publicly disclosed indicator P i It will also ask its neighbors (including other Ps) i Similar issues to those related to indicators connected through institutions.

[0101] Information aggregation: Each node aggregates all the neighbor information it receives through some mathematical method (such as summation, averaging, or more complex attention mechanisms). This process is the physical implementation of "relational reasoning." Because it aggregates the information of its neighbors, the node knows its own context in this relational network.

[0102] Representation update: A node combines its old feature vector with the newly aggregated neighbor information, and through a neural network layer, generates a new, more powerful feature vector. This is called "updating the node representation".

[0103] After multiple updates by the GNN, each node (including all document metrics and candidate public metrics) has a final, context-aware representation that incorporates graph structure information.

[0104] Step S34: Calculate the updated manuscript index data and / or the second candidate public index data, recalculate the comprehensive matching score, and determine the matching public index data of the manuscript to be verified.

[0105] In this embodiment, the comprehensive matching score in the above embodiments can be recalculated based on the updated indicator relationship graph: for each manuscript indicator Q j Using its final representation, sum all public indices P in the quadratic candidate set C2. i The final representation recalculates the similarity (e.g., calculates the cosine similarity between vectors).

[0106] This new similarity score, because it incorporates complex relationships such as institution, time, and co-occurrence, is much more accurate and comprehensive than the original similarity score based solely on name semantics.

[0107] This application utilizes a fixed graph structure as a skeleton and uses a GNN to allow information to flow and merge between nodes, so that the representation of each node carries information about its contextual relationships. Finally, based on these higher-quality node representations, the matching relationship is re-evaluated, thereby optimizing the candidate set.

[0108] Furthermore, this application also provides a specific scheme for determining the final matching public indicator data for manuscript indicator data using spatiotemporal awareness confidence. Please continue reading. Figure 5 , Figure 5This is a flowchart illustrating the fourth embodiment of the indicator consistency comparison method provided in this application.

[0109] like Figure 5 As shown, the indicator consistency comparison method in this application embodiment specifically includes the following steps: Step S41: Filter the first candidate public indicator set according to the comprehensive matching score to generate a third candidate public indicator set; the third candidate public indicator set includes several third candidate public indicator data.

[0110] In this application embodiment, the comprehensive matching score calculated in the above embodiment is used for further screening of the candidate public index set.

[0111] Step S42: Calculate the spatiotemporal confidence of the third candidate public index data and the manuscript index data.

[0112] In this embodiment of the application, the application selects and optimizes each public indicator data P from the candidate public indicator set. i Computational Enhancement of Spatiotemporal Awareness Confidence The The calculation integrates name similarity, caliber similarity, authoritative factor of the issuing organization, timeliness factor of release time, and consistency deviation factor of indicator values ​​obtained based on graph relationship reasoning.

[0113] In response, this application provides enhanced spatiotemporal awareness confidence. The calculation formula is:

[0114] in: M is the name of the manuscript indicator. j With public indicator name N i Semantic similarity.

[0115] For manuscript indicators E j Compared with the publicly disclosed indicator caliber D i Semantic similarity.

[0116] ρ is the time-dependent gain coefficient, with a value ranging from 0.1 to 0.3.

[0117] λ is the time decay coefficient, with a value ranging from 0.05 to 0.2.

[0118] Δt is the time difference between the current time and the publication time Ti.

[0119] This is the scaling factor, usually set to 100.

[0120] As an authoritative impact factor, based on the publishing institution O i The authoritative ranking calculation.

[0121] The time-dependent impact factor is based on the publication time T. i The old and new sorting calculations.

[0122] Indicates taking and The average value.

[0123] η is the balance factor, with a value ranging from 0.5 to 0.8.

[0124] σ is a consistency-sensitive parameter, with a value range of 0.5-2.0.

[0125] The publicly disclosed indicator value V i The absolute difference between the numerical reference value of the correlation index obtained based on graph neural network inference and the value of the correlation index.

[0126] These are reference values ​​for related indicators.

[0127] Among them, the aforementioned authoritative influence factors The calculation methods include: All indicators in the secondary candidate matching set C2 are sorted according to the authority level of the issuing institution, with the institution with the highest authority corresponding to a ranking value of 1, and so on, increasing sequentially.

[0128] Calculate the variance μ of the numerical sequence of the index.

[0129] Calculate the authoritative impact factor:

[0130] in, For public indicator data P i The authoritative ranking value, This is a preset gradient factor, with a value range of 5-10.

[0131] Among them, the aforementioned time-related influencing factors The calculation methods include: According to the release time T i Sort all indicators in the secondary candidate matching set C2 from newest to oldest, with the latest published indicator having a sort value of 1, and so on.

[0132] Calculate the variance of the numerical series of indicators .

[0133] Calculate the time-related impact factor:

[0134] in, For public indicator data P i Time-ordered values, This is a preset gradient factor, with a value range of 5-10.

[0135] Step S43: Determine the matching public index data of the document to be verified based on the spatiotemporal awareness confidence level.

[0136] Step S44: Generate a consistency comparison report based on the manuscript indicator data and the matching public indicator data.

[0137] For a summary of the consistency comparison methods for indicators from the above embodiments, please refer to [link / reference needed]. Figure 2 This application provides a consistency comparison system for reference data based on a large language model, which is specifically used in conjunction with a data layer, an algorithm layer, and an application layer.

[0138] The data layer includes a large language model storage unit, an unstructured vector database (which supports storing vectorized representations of documents in various formats such as PDF, Word, TXT, and images, and uses a hierarchical index structure to achieve efficient similarity retrieval) and a structured relational database (which uses a distributed architecture, supports horizontal scaling, and can store and manage more than ten million levels of structured indicator data).

[0139] The algorithm layer includes: a multimodal coding module for extracting and fusing multimodal features from publicly available literature and manuscript data; a semantic matching and attention filtering module for performing two-stage semantic matching and comprehensive scoring based on an attention mechanism; a graph relationship reasoning module for constructing an index relationship graph and performing graph neural network reasoning; and an enhanced confidence calculation module for calculating enhanced spatiotemporal awareness confidence. An adaptive parameter management module is used to dynamically optimize system parameters through reinforcement learning; and a feedback learning module is used to collect user feedback and update the reinforcement learning model.

[0140] The application layer includes: an upload module for uploading publicly available documents and manuscripts to be verified; a verification module for triggering and executing the consistency comparison process; a display module for displaying the comparison report and interactive interface; and a management module for system configuration and backend management.

[0141] Furthermore, the comparison report generated by the display module in the application layer may include, but is not limited to, the following parts: (1) Manuscript indicator data sequence table, which displays the matching status and verification results of each manuscript indicator data, and supports manual modification and re-verification.

[0142] (2) Full-text view of the proposed manuscript, which supports highlighting the corresponding position in the original text when clicking on a sequence list item.

[0143] (3) Matching details display area, which displays detailed information of the public indicator data matched by each document indicator data, including indicator name, scope, value, source, release time and release organization.

[0144] The feedback mechanism implemented by the feedback learning module in the algorithm layer includes: recording all correction operations performed by the user on the matching results; analyzing the patterns and rules of the correction operations; converting the correction data into training samples for reinforcement learning; and periodically updating system parameters and model weights.

[0145] Furthermore, the system of this application may also include an API interface module, providing a RESTful API interface to support third-party system integration and invocation of the consistency comparison service. The services provided by the API interface module include: document uploading and parsing services; indicator data extraction services; consistency comparison verification services; and report generation and export services.

[0146] The system in this application adopts a microservice architecture, in which each module can be deployed and expanded independently, and communication and data flow between modules are realized through message queues and service mesh.

[0147] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0148] To implement the above-mentioned indicator consistency comparison method, this application also proposes an indicator consistency comparison device, which can be found in the following details. Figure 6 , Figure 6 This is a schematic diagram of an embodiment of the index consistency comparison device provided in this application.

[0149] The consistency comparison device 500 of this embodiment includes a processor 51, a memory 52, an input / output device 53, and a bus 54.

[0150] The processor 51, memory 52, and input / output device 53 are respectively connected to the bus 54. The memory 52 stores program data, and the processor 51 is used to execute the program data to implement the index consistency comparison method described in the above embodiment.

[0151] In this embodiment, processor 51 can also be referred to as a CPU (Central Processing Unit). Processor 51 may be an integrated circuit chip with signal processing capabilities. Processor 51 can also be a general-purpose processor, 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. The general-purpose processor can be a microprocessor, or processor 51 can be any conventional processor.

[0152] This application also provides a computer storage medium; please refer to the following: Figure 7 , Figure 7 This is a schematic diagram of a computer storage medium according to an embodiment of the present application. The computer storage medium 600 stores a computer program 61, which, when executed by a processor, is used to implement the index consistency comparison method of the above embodiment.

[0153] When the embodiments of this application are implemented as software functional units and sold or used as independent products, they 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 all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the 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.

[0154] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for comparing the consistency of indicators, characterized in that, The method for comparing the consistency of the indicators includes: Extract the publicly available indicator data from the knowledge base and the publicly available multi-dimensional feature vectors of the publicly available indicator data; Determine the manuscript metrics data for the document to be verified; Construct a multi-dimensional feature vector for the manuscript based on the manuscript index data; A first candidate set of public indicators is selected by matching the names of the manuscript indicators in the manuscript indicator data with the names of the public indicators in the public indicator data; wherein, the first candidate set of public indicators includes several first candidate public indicator data. A comprehensive matching score is calculated using the multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data. The matching public index data of the manuscript to be verified is determined based on the comprehensive matching score; A consistency comparison report is generated based on the document index data and the matching public index data.

2. The indicator consistency comparison method according to claim 1, characterized in that, The step of filtering out the first candidate set of public indicators by matching the names of the manuscript indicators in the manuscript indicator data with the names of the public indicators in the public indicator data includes: Calculate the first cosine similarity between the manuscript indicator names in the manuscript indicator data and the public indicator names in the public indicator data; The first candidate public index set is constructed based on the public index data whose first cosine similarity is higher than the dynamic similarity threshold; The dynamic similarity threshold is dynamically adjusted based on the historical matching accuracy of the domain to which the document to be verified belongs.

3. The indicator consistency comparison method according to claim 1, characterized in that, The step of calculating a comprehensive matching score using the multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data includes: Calculate the second cosine similarity between the manuscript multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data; Calculate the attention score between the manuscript index data and the first candidate public index data; The comprehensive matching score is calculated based on the second cosine similarity and the attention score.

4. The indicator consistency comparison method according to claim 3, characterized in that, The step of calculating the comprehensive matching score based on the second cosine similarity and the attention score includes: Based on the historical matching records of the first candidate publicly available indicator data, a historical consistency score is determined; The comprehensive matching score is calculated based on the historical consistency score, the second cosine similarity, and the attention score.

5. The indicator consistency comparison method according to claim 4, characterized in that, The step of determining the historical consistency score based on the historical matching records of the first candidate public indicator data includes: Based on the historical matching records of the first candidate public indicator data, the first candidate public indicator data is determined to be the number of times a match has been determined and the total number of matches. The historical consistency score is determined based on the ratio of the number of determined matches to the total number of matches.

6. The indicator consistency comparison method according to claim 1, characterized in that, After calculating the comprehensive matching score using the multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data, the index consistency comparison method further includes: The first candidate public indicator set is filtered based on the comprehensive matching score to generate a second candidate public indicator set; the second candidate public indicator set includes several second candidate public indicator data. Construct an indicator relationship diagram between the several second candidate public indicator data and the manuscript indicator data; The node representations of the indicator relationship graph are updated and the relationships are inferred using a graph neural network model, so as to update the manuscript indicator data and / or the second candidate public indicator data; The updated manuscript index data and / or the second candidate public index data are used to recalculate the comprehensive matching score, and the matching public index data of the manuscript to be verified is determined.

7. The indicator consistency comparison method according to claim 6, characterized in that, The construction of the indicator relationship diagram between the plurality of second candidate public indicator data and the manuscript indicator data includes: Nodes are generated based on the second candidate public indicator data and the manuscript indicator data; Construct edges between adjacent nodes using any one or more of the following rules: When the semantic similarity between the manuscript indicator name of the manuscript indicator data and the public indicator name of the second candidate public indicator data is higher than a preset semantic threshold, a similarity relationship edge is added between the two nodes. When the publishing organization of the manuscript index data and the publishing organization of the second candidate public index data are the same publishing organization, add an organization association edge between the two nodes; When the time difference between the publication time of the manuscript index data and the publication time of the second candidate public index data is lower than a preset time difference threshold, a time correlation edge is added between the two nodes. When the manuscript index data and the second candidate public index data appear in the same historical document, add a co-occurrence relationship edge between the two nodes.

8. The indicator consistency comparison method according to claim 1, characterized in that, After calculating the comprehensive matching score using the multi-dimensional feature vector of the manuscript index data and the public multi-dimensional feature vector of the first candidate public index data, the index consistency comparison method further includes: The first candidate public indicator set is filtered based on the comprehensive matching score to generate a third candidate public indicator set; the third candidate public indicator set includes several third candidate public indicator data. Calculate the spatiotemporal perceived confidence of the third candidate public index data and the manuscript index data; The matching public index data of the manuscript to be verified is determined based on the spatiotemporal awareness confidence level. A consistency comparison report is generated based on the manuscript indicator data and the matching public indicator data. The spatiotemporal perception confidence level is determined by the semantic similarity between the document indicator name and the publicly available indicator name, the semantic similarity between the document indicator caliber and the publicly available indicator caliber, and / or the time difference between the public release time and the current time.

9. The indicator consistency comparison method according to claim 8, characterized in that, The spatiotemporal perception confidence level is also determined by the authoritative influence factor and / or time influence factor of the third candidate publicly available indicator data; The authority impact factor is determined based on the authority ranking of the third candidate public indicator data's public release institution among all public release institutions in the third candidate public indicator set; the time impact factor is determined based on the newness ranking of the public release time of the third candidate public indicator data among all public release times in the third candidate public indicator set.

10. The indicator consistency comparison method according to claim 1, characterized in that, The process of generating a consistency comparison report based on the manuscript indicator data and the matching publicly available indicator data includes: The consistency between the manuscript indicator values ​​of the manuscript indicator data and the public indicator values ​​of the matching public indicator data is determined. Based on the consistency determination results, the consistency comparison report is generated.

11. A device for comparing index consistency, characterized in that, The indicator consistency comparison device includes a memory and a processor coupled to the memory; The memory is used to store program data, and the processor is used to execute the program data to implement the indicator consistency comparison method as described in any one of claims 1 to 10.

12. A computer storage medium, characterized in that, The computer storage medium is used to store program data, which, when executed by the computer, is used to implement the indicator consistency comparison method as described in any one of claims 1 to 10.

Images