Standard document authoring method and system enhanced based on knowledge retrieval

Abstract

The application discloses a standard document writing method and system based on knowledge retrieval enhancement, comprising the following steps: acquiring field multi-source heterogeneous knowledge by knowledge retrieval and constructing a field knowledge graph and a content increment model; performing semantic extraction, structure division, content segmentation and text granularity perception; determining first matching information and second matching information; matching standard paragraph structures and standard sentence structures to generate a standard template; performing compliance review to obtain a compliant text and recommended text; performing abnormality screening and abnormality correction to obtain a compliance value; inputting the compliant text, the recommended text and the compliance value into the content increment model to perform normalized conversion to obtain standard materials; and inputting the standard materials of each sentence into the standard template to obtain a standard document. The method can improve the efficiency and accuracy of standard document writing, has good interpretability, and can be directly applied to a standard document writing system based on knowledge retrieval enhancement.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of document writing, in particular to a standard document writing method and system based on knowledge retrieval enhancement. BACKGROUND

[0002] With the vigorous development of knowledge economy and the continuous deepening of specialization in various fields, standard documents play a crucial role in standardizing industry operations, ensuring product quality, and promoting technical exchange and cooperation. Precise and efficient standard document writing is of great importance to promoting the standardization process, improving overall operational efficiency, and protecting the interests of all parties.

[0003] Traditional standard document writing techniques have obvious shortcomings. First, manual writing is easily influenced by subjective factors, making it difficult to ensure the accuracy and consistency of document content, and differences in personnel understanding can lead to uneven document quality. At the same time, in the face of complex and multi-source heterogeneous knowledge materials, manual selection and integration is inefficient, making it difficult to quickly respond to standard updates and business change requirements. In addition, existing auxiliary writing techniques mostly focus on single text material processing, lack deep semantic mining and system integration of knowledge, and fail to fully activate multi-source heterogeneous knowledge resources, providing limited intelligent support for the writing process. The present application designs a standard document writing method and system based on knowledge retrieval enhancement, which integrates multi-source heterogeneous knowledge by building a domain knowledge graph, dynamically generates standard templates using semantic extraction, text granularity perception, and other technologies, and realizes the standardized transformation of knowledge by combining compliance review and content increment models, effectively solving the limitations of traditional methods, significantly improving the automation level, content quality, and compliance of standard document writing, and having important significance for promoting the standardization construction of various industries and improving enterprise operational efficiency. SUMMARY

[0004] The purpose of the present application is to provide a standard document writing method and system based on knowledge retrieval enhancement.

[0005] To achieve the above purpose, the present application is implemented according to the following technical solutions:

[0006] The present application includes the following steps:

[0007] Knowledge retrieval is performed to obtain multi-source heterogeneous knowledge in the field and build a domain knowledge graph. The multi-source heterogeneous knowledge in the field is divided according to general expression and standard expression, and a content increment model is constructed according to the division results;

[0008] The semantic extraction is performed on the user's writing requirements to obtain format requirements and first semantic features, and the structure division and content segmentation are performed on the materials to be written to obtain text materials and numerical materials corresponding to each sentence. The structure division includes paragraph division and sentence division;

[0009] The semantic extraction is performed on the text material to obtain second semantic features, text granularity perception is performed on the text material to obtain granularity features, first matching information is determined according to the first semantic features and the granularity features, and second matching information is determined according to the second semantic features and the granularity features;

[0010] A standard paragraph structure is determined according to the first matching information, a standard sentence structure is determined according to the second matching information, and a standard template is generated according to the format requirement, the standard paragraph structure and the standard sentence structure;

[0011] The to-be-edited material is input into the domain knowledge graph for compliance review to obtain compliant text, recommended text and specification numerical range, a domain numerical multi-sphere is constructed according to the specification numerical range for abnormal screening and abnormal correction to obtain compliant numerical values;

[0012] The compliant text, the recommended text and the compliant numerical values are input into the content increment model for standardized conversion to obtain specification materials, and the specification materials of each sentence are input into the standard template to obtain a standard document.

[0013] Further, the method for constructing the domain knowledge graph comprises:

[0014] Internal structured data, internal unstructured data, external authoritative data, external academic data and real-time data streams are collected by knowledge retrieval, the collected data are respectively subjected to format standardization, entity recognition and cleaning, relationship annotation and knowledge fusion to obtain domain multi-source heterogeneous knowledge, core class knowledge and edge relationship are respectively defined according to the domain multi-source heterogeneous knowledge, structured data is mapped into data triples according to (entity, attribute, value), a BERT-BiLSTM-CRF joint model is used to extract entity relationship to obtain text triples, a graph database Neo4j is used to store the data triples and the text triples to construct the domain knowledge graph;

[0015] The domain knowledge graph is subjected to graph optimization by using a hybrid pruning algorithm, dynamic weight rebalancing and a double-channel reinforcement mechanism; the hybrid pruning algorithm specifically comprises semantic density pruning and topological importance pruning; the double-channel reinforcement mechanism comprises a real-time channel and a batch channel;

[0016] The semantic density pruning divides semantic clusters by using adaptive Gaussian mixture clustering, and a cluster representative node is selected according to a compactness-dispersity balance condition, and the expression is as follows:

[0017] ;

[0018] ;

[0019] wherein is the i-th semantic cluster reserved representative node, the reserved condition is the maximum point satisfying the balance condition of compactness-dispersity, node set the first node, node quantity, the first semantic cluster size, semantic compactness of node and node , balance factor dynamically adjusted by KL divergence, , embedding vector of node and , cluster center selection condition, cluster index, cluster quantity, semantic compactness of node and cluster center position , regularization coefficient, prior probability of semantic cluster ;

[0020] The topological importance pruning calculates the importance score of each node by improving the PageRank algorithm and introducing relationship type weight and time decay factor, and retains the nodes in the key path whose importance score is less than the importance score threshold, the expression is:

[0021] ;

[0022] Wherein importance score of node , importance score of neighbor node , damping coefficient, inbound neighbor set of node , base weight of edge pointing from node to node , relationship type reinforcement function, adjust the weight according to the type of edge , embedding vector of relationship type , learnable projection matrix, target node eigenvectors, For nodes The set of outgoing neighbors, For nodes Pointing to node edge The basic weights, For decay rate parameters, The current system time. for timestamp, As a reward factor for recent events, For the indicator function, when the timestamp The value is 1 if the time threshold is exceeded, and 0 otherwise.

[0023] The dynamic weight rebalancing is obtained based on the graph structure update after hybrid pruning, and the expression is:

[0024] ;

[0025] ;

[0026] in side Rebalancing weights For nodes With nodes semantic similarity weights, For nodes With nodes Topological association weights, This is the node importance score vector. For the embedding matrix gradient operator, Nodes in the diagram after pruning With nodes Meta-path similarity, Nodes in the original graph With nodes Meta-path similarity, Temperature coefficient;

[0027] The real-time channel reconstructs and optimizes the graph structure through event-driven subgraph reconstruction.

[0028] The batch channel uses incremental reinforcement learning to decay and enhance node weights and adjusts the importance score threshold in topological importance pruning.

[0029] Furthermore, the method for constructing the content incremental model includes:

[0030] The domain multi-source heterogeneous knowledge is input into a pre-trained bag-of-words model to obtain different terms of the same entity to form an entity term set, the authoritative frequency and information entropy of each term in external authoritative data and external academic data are calculated to obtain a standard score, the term with the highest standard score in the entity term set is taken as a standard expression, and the remaining terms in the entity term set are general expressions;

[0031] The expression of the standard score is:

[0032] ;

[0033] Wherein is the term standard score, is the authoritative frequency weight, is the information entropy weight, is the user usage frequency weight, is the term frequency of occurrence in external authoritative data and external academic data, is the term context environment, is the term context environment probability of occurrence, is the term frequency of occurrence in internal unstructured data;

[0034] The highest frequency of data unit and data magnitude associated with the standard expression is perceived by the BERT perception model, the domain ontology is taken as a framework, the standard expression is taken as a parent node, and the general expression is taken as a child node to construct a semantic mapping relationship;

[0035] According to the division result, a content increment model is constructed, and the specific structure includes an ontology recognition library, a mapping engine, a unit normalizer and a full connection layer; the ontology recognition library is used for labeling the input text as a standard expression or a general expression, the standard expression is directly input into the unit normalizer and the full connection layer, and the general expression is input into the mapping engine; the mapping engine is embedded with a semantic mapping relationship, and the general expression is converted into a standard expression and input into the unit normalizer and the full connection layer; the unit normalizer is used for extracting the data unit and data magnitude associated with the standard expression and comparing with the input data, adjusting the data unit and optimizing the data magnitude, and outputting the optimized data; the full connection layer is used for connecting the ontology recognition library, the mapping engine and the unit normalizer to output a normalized content increment; the normalized content increment includes the standard expression and corresponding data unit and data magnitude.

[0036] Further, the method for obtaining the text material and numerical material corresponding to each sentence comprises:

[0037] When the paragraph format of the material to be compiled is obvious, the text separator is recognized and the paragraphs are divided, the end-of-sentence punctuation of each paragraph is recognized by using a regular expression, and the sentences are segmented at the punctuation positions by using the NLTK sentence segmentation function;

[0038] When the paragraph format of the material to be compiled is not obvious, the sentences are directly segmented according to the end-of-sentence punctuation, the expression content of each sentence is determined by semantic recognition, and the sentences with the same expression content are divided into the same paragraph; the expression content includes background research, market summary, standard , operation method, data analysis, development goal and significance;

[0039] The numerical mode of each sentence of the material to be compiled is recognized by using a regular expression, the numerical material is extracted from each sentence, the remaining material is taken as text material, and the numerical material is associated with the corresponding text material.

[0040] Further, the method for determining the first matching information comprises:

[0041] The explicit description and implicit intention of the user's writing requirement are recognized by using a rule-based pattern matching engine to obtain the page size and page layout, the file format of the user's writing requirement is determined by using a keyword trigger-context verification mechanism, the field label is output by using a field-adapted BERT classifier, and the format requirement is composed of the page size, the page layout, the file format and the field label;

[0042] The user's writing requirement is divided into multiple text units according to punctuation, the core theme cluster of each text unit is recognized by using a hierarchical Dirichlet process, the theme coherence of each text unit is analyzed according to the core theme cluster, adjacent text units with consistent core themes are merged, the core theme cluster of the merged text unit is taken as the paragraph theme, the paragraph order is obtained by performing time sequence dependence analysis on the merged text unit, and the first semantic feature is composed of the paragraph theme and the paragraph order;

[0043] The text material of the material to be compiled is input into a context-aware recommendation engine to obtain a text level and a text term frequency; the text level is determined according to the inclusion relationship of the text;

[0044] The text outline size is determined according to the text level of each paragraph and the number of paragraphs, the text terms are classified and the cumulative frequency of the same type of term is counted, the ratio of the cumulative frequency of the same type of term to the corresponding standard frequency is calculated to obtain the term professional degree, the mean value of the professional degrees of all types of terms in the field is taken as the field professional degree, the text terms are input into the field knowledge graph to obtain the corresponding associated clauses, the maximum value of the associated clause level of the same type of text term is taken as the term association degree, and the mean value of the association degrees of all types of terms in the field is taken as the field association degree;

[0045] The text outline size, field expertise and field correlation degree are combined to form granularity features, and the first semantic features and the granularity features are spliced to obtain first matching information.

[0046] Further, the method for determining second matching information comprises:

[0047] According to the expression content of each sentence text material, the paragraph theme is determined to determine the paragraph positioning, a ternary relationship classifier is constructed to identify the relationship between each sentence in the same paragraph, each sentence in the same paragraph is sorted according to the appearance order of the to-be-edited material, and the absolute position coding is input into the position coding-aware Transformer to obtain the sentence order, and the second semantic features are composed of the paragraph positioning, the sentence relationship and the sentence order; the relationship between each sentence includes a reference relationship, an emphasis relationship and a lead-out relationship;

[0048] The second semantic features and the granularity features are spliced to obtain second matching information.

[0049] Further, the method for performing compliance review comprises:

[0050] The text material of the to-be-edited material is input into the field knowledge graph to perform compliance review, and the specific steps comprise:

[0051] The compliance of the text terms in the text material in the field knowledge is checked, and the non-compliant text terms are removed, replaced and supplemented to obtain compliant text;

[0052] The recommended text of the text material not explicitly mentioned by the user is recommended through graph relationship reasoning; the basis of the graph relationship reasoning is that the text terms corresponding to the nodes have strong relationship edges, and the text terms corresponding to the other nodes of the strong relationship edges do not appear in the to-be-edited material; the edge weight of the most recently updated edge of the strong relationship edge is greater than 0.85; the recommended text includes related mandatory provisions and best practices;

[0053] The specification numerical range of the nodes where the compliant text and the recommended text are located is extracted, the field numerical multi-sphere is constructed according to the specification numerical range, the numerical material of the to-be-edited material is input into the field numerical multi-sphere to perform abnormal screening and abnormal correction to obtain compliant numerical values; the abnormal correction specifically means that the abnormal numerical values determined by the abnormal screening are manually verified, and the abnormal numerical values are adjusted according to the manual verification results.

[0054] Further, the method for obtaining specification material comprises:

[0055] The compliant text and the recommended text are input into a content increment model to convert general expressions in the text into specification expressions to obtain specification text, and the associated data unit and data magnitude of the specification text are extracted, the data unit is adjusted according to the compliant numerical value, and the data magnitude is optimized to obtain specification data, and the specification text and the specification data are combined to form specification material.

[0056] According to the second semantic feature, the standard template position of each sentence is determined, and the standard document is obtained by inputting the specification material of each sentence into the standard template according to the standard template position.

[0057] In a second aspect, the standard document writing system based on knowledge retrieval enhancement comprises:

[0058] The model module is used for knowledge retrieval to obtain field multi-source heterogeneous knowledge and construct a field knowledge graph, divides the field multi-source heterogeneous knowledge according to general expression and standard expression, and constructs a content increment model according to the division result.

[0059] The material analysis module is used for semantic extraction of user writing requirements to obtain format requirements and first semantic features, structure division and content segmentation of the material to be written to obtain text material and numerical material corresponding to each sentence, semantic extraction of the text material to obtain second semantic features, and text granularity perception of the text material to obtain granularity features.

[0060] The template module is used for determining first matching information according to the first semantic features and the granularity features, and determining second matching information according to the second semantic features and the granularity features, determining a standard paragraph structure according to the first matching information, determining a standard sentence structure according to the second matching information, and generating a standard template according to the format requirements, the standard paragraph structure and the standard sentence structure.

[0061] The review module is used for inputting the material to be written into the field knowledge graph to perform compliance review and obtain compliant text, recommended text and specification numerical range, constructing a field numerical multi-sphere according to the specification numerical range to perform abnormal screening and abnormal correction to obtain compliant numerical values.

[0062] The writing module is used for inputting the compliant text, the recommended text and the compliant numerical values into the content increment model to perform standardized conversion and obtain specification material, and inputting the specification material of each sentence into the standard template to obtain a standard document.

[0063] The present application has the following beneficial effects:

[0064] Compared with the prior art, the present application has the following technical effects:

[0065] The application can improve the material preprocessing ability and the intelligent level of standard document writing during standard document writing through the steps of knowledge retrieval, semantic extraction, text granularity perception, compliance review, standardized conversion and model construction, thereby improving the efficiency and precision of standard document writing, optimizing the standard document writing technology, greatly saving resources, improving work efficiency, realizing the writing of standard documents, significantly improving the automation degree, content quality and compliance of standard document writing, and having important significance for promoting the standardization construction of various industries and improving the operation efficiency of enterprises. The application can adapt to different standard document writing systems and the standard document writing needs of different users, and has certain universality. BRIEF DESCRIPTION OF DRAWINGS

[0066] Figure 1 A step flowchart of the standard document writing method based on knowledge retrieval enhancement of the application. DETAILED DESCRIPTION

[0067] The application will be further described below through specific embodiments. The illustrative embodiments of the application and the description are used to explain the application, but do not limit the application.

[0068] The standard document writing method and system based on knowledge retrieval enhancement of the application include the following steps:

[0069] As shown in Figure 1 In this embodiment, the following steps are included:

[0070] Knowledge retrieval is performed to obtain field multi-source heterogeneous knowledge and construct a field knowledge graph. The field multi-source heterogeneous knowledge is divided according to general expression and standard expression, and a content increment model is constructed according to the division result.

[0071] Semantic extraction is performed on user writing requirements to obtain format requirements and first semantic features. Structure division and content segmentation are performed on the materials to be written to obtain text materials and numerical materials corresponding to each sentence. The structure division includes paragraph division and sentence division.

[0072] Semantic extraction is performed on the text materials to obtain second semantic features. Text granularity perception is performed on the text materials to obtain granularity features. First matching information is determined according to the first semantic features and the granularity features. Second matching information is determined according to the second semantic features and the granularity features.

[0073] Standard paragraph structures are determined according to the first matching information. Standard sentence structures are determined according to the second matching information. Standard templates are generated according to the format requirements, the standard paragraph structures and the standard sentence structures.

[0074] The material to be compiled is input into the domain knowledge graph for compliance review to obtain compliant text, recommended text and standard numerical range. Based on the standard numerical range, a domain numerical multisphere is constructed to perform anomaly screening and anomaly correction to obtain compliant values.

[0075] The compliant text, the recommended text, and the compliant values ​​are input into the content incremental model for standardized transformation to obtain standardized materials. The standardized materials for each sentence are then input into the standard template to obtain a standard document.

[0076] In this embodiment, the method for constructing a domain knowledge graph includes:

[0077] Knowledge retrieval is performed by collecting internal structured data, internal unstructured data, external authoritative data, external academic data, and real-time data streams. The collected data is then subjected to format standardization, entity recognition and cleaning, relation annotation, and knowledge fusion to obtain multi-source heterogeneous knowledge in the domain. Core knowledge classes and edge relations are defined based on the multi-source heterogeneous knowledge in the domain. Structured data is mapped to data triples according to (entity, attribute, value). The BERT-BiLSTM-CRF joint model is used to extract entity relations to obtain text triples. The graph database Neo4j is used to store the data triples and text triples to construct a domain knowledge graph.

[0078] The domain knowledge graph employs a hybrid pruning algorithm, dynamic weight rebalancing, and a dual-channel reinforcement mechanism for graph optimization. The hybrid pruning algorithm specifically includes semantic density pruning and topological importance pruning. The dual-channel reinforcement mechanism includes a real-time channel and a batch channel.

[0079] The semantic density pruning uses adaptive Gaussian mixture clustering to divide semantic clusters, and selects representative nodes for each cluster based on a density-dispersion balance condition, expressed as follows:

[0080] ;

[0081] ;

[0082] in For the first semantic clusters The representative nodes that are retained are those that satisfy the maximum value of the density-dispersion balance condition. For node set The first in 1 node For the number of nodes, For the first semantic clusters scale For nodes With nodes semantic density of the node a balance factor dynamically adjusted for KL divergence, 、 an embedding vector of the node and a cluster center, a cluster center selection condition, a cluster index, a cluster number, semantic density of the node and a cluster center point position a semantic density of the node a regularization coefficient, a prior probability of the semantic cluster ;

[0083] The topological importance pruning calculates the importance score of each node by improving the PageRank algorithm and introducing relationship type weight and time decay factor, and retains the nodes in the key path whose importance score is less than the importance score threshold, and the expression is:

[0084] ;

[0085] wherein an importance score of the node , an importance score of the neighbor node , a damping coefficient, an in-edge neighbor set of the node , a base weight of the edge pointing from the node to the node , a relationship type reinforcement function, adjusting the weight according to the type of the edge , an embedding vector of the relationship type , a learnable projection matrix, a feature vector of the target node , an out-edge neighbor set of the node , a base weight of the edge pointing from the node to the node , a decay rate parameter, a current system time, a timestamp of , a recent event reward factor, is a function indicating 1 when the timestamp is greater than the time threshold, and 0 otherwise;

[0086] The dynamic weight rebalancing is obtained based on the mixed pruned graph structure update, and the expression is:

[0087]

[0088]

[0089] wherein the rebalanced weight of the edge is the semantic similarity weight of the node and the node is the semantic similarity weight of the node and the node is the topological correlation weight of the node and the node is the gradient operator with respect to the embedding matrix is the meta-path similarity of the node and the node in the pruned graph is the meta-path similarity of the node and the node in the original graph

[0090] The real-time channel optimizes the graph structure through event-driven subgraph reconstruction;

[0091] The batch channel decays and enhances the node weight through incremental reinforcement learning, and adjusts the importance score threshold in the topological importance pruning;

[0092] ​​​​​​In actual evaluation, the specific steps of knowledge retrieval include obtaining internal structured data through SQL database interface, recognizing internal unstructured data through OCR, obtaining external authoritative data through open data platform, obtaining external academic data through academic search engine API, and collecting real-time data stream through message queue Kafka. The internal structured data includes enterprise history standard library, product specification library, term library, and BOM table. The internal unstructured data includes design documents, test reports, process manuals, and meeting minutes. The external authoritative data includes international / national / industry standard library (ISO / IEC, GB, ANSI), patent database (WIPO, USPTO), and regulation library (FDA, CE). The external academic data includes academic papers (IEEE Xplore, Springer, CNKI), technical white paper, and industry research report. The real-time data stream includes industry information website, regulatory agency announcement, and competitive product dynamics.

[0093] The results of entity recognition include term entity, clause entity, test method entity, and numerical entity. Knowledge fusion includes entity alignment, attribute complement, and conflict resolution.

[0094] Core class knowledge includes standard class knowledge, clause class knowledge, term class knowledge, and test class knowledge. The attributes of standard class knowledge are [number, name, release date]. The attributes of clause class knowledge are [number, content, constraint level]. The attributes of term class knowledge are [definition, synonym]. The attributes of test class knowledge are [method, equipment, precision].

[0095] Edge relationships include inclusion relationship, reference relationship, term relationship, and version evolution relationship.

[0096] Regularization coefficient in semantic density pruning Damping coefficient in topological importance pruning Decay rate parameter The value of importance score threshold is 0.03 Temperature coefficient in dynamic weight rebalancing ;

[0097] Real-time channel reconstructs subgraph by constructing priority mapping of events, extracting event-related subgraph, and solving constraint optimization problem. The expression is:

[0098] ;

[0099] ;

[0100] Wherein is the priority mapping of event , is the authoritative weight coefficient, is the event the authority score of, the knowledge distribution change intensity, determined by calculating the KL divergence of the knowledge probability distribution before the event and the knowledge probability distribution after the event, the extracted subgraph, the node set of the extracted subgraph, the distance of the node to the event , the subgraph radius threshold value;

[0101] In the batch channel, the expression of the node weight decay-enhancement model and the importance score threshold adjustment is:

[0102] ;

[0103] ;

[0104] wherein is the updated node weight at the moment, is the decay rate, is the update time interval, is the number of feedback events to be processed, is the user feedback signal, taking when the recommendation is accepted, and taking when the recommendation is rejected, , is the feedback intensity base, is the number of times of rejection of the node , is the updated importance score threshold at the moment, is the basic threshold value (default 0.03), is the adjustment amplitude coefficient, is the recent recommendation accuracy rate (7 days), is the baseline accuracy rate, is the sensitivity control parameter.

[0105] In the embodiment, the method for constructing the content increment model comprises:

[0106] inputting the field multi-source heterogeneous knowledge into a pre-trained bag-of-words model to obtain an entity term set composed of different terms of the same entity, calculating the authority frequency and information entropy of each term in external authority data and external academic data to obtain a standard score, taking the term with the highest standard score in the entity term set as a standard expression, and taking the remaining terms in the entity term set as general expressions;

[0107] The expression of the standard score is: ​

[0108] ;

[0109] wherein is the term normative score, is the authority frequency weight, is the information entropy weight, is the user usage frequency weight, is the term frequency of occurrence in external authority data and external academic data, is the term context environment, is the term context environment of occurrence probability of occurrence, is the term frequency of occurrence in internal unstructured data;

[0110] The data unit and data magnitude with the highest frequency of use of the normative expression are associated by the BERT perception model, and the semantic mapping relationship is constructed with the domain ontology as the framework, taking the normative expression as the parent node and the general expression as the child node;

[0111] A content increment model is constructed according to the division result, and the specific structure includes an ontology recognition library, a mapping engine, a unit normalizer and a full connection layer; the ontology recognition library is used to label the input text as a normative expression or a general expression, the normative expression is directly input into the unit normalizer and the full connection layer, and the general expression is input into the mapping engine; the mapping engine is embedded with a semantic mapping relationship, and the general expression is converted into a normative expression and input into the unit normalizer and the full connection layer; the unit normalizer is used to extract the data unit and data magnitude associated with the normative expression and compare them with the input data, adjust the data unit and optimize the data magnitude, and output the optimized data; the full connection layer is used to connect the ontology recognition library, the mapping engine and the unit normalizer to output the normalized content increment; the normalized content increment includes the normative expression and the corresponding data unit and data magnitude;

[0112] In the actual evaluation, taking the general expression and the standard expression in the oil field as an example, the entity "wellbore evaluation" has multiple term expressions in external authoritative data and external academic data, and the entity term set includes: wellbore integrity evaluation, wellbore integrity assessment, wellbore health check, borehole stability analysis, the standard scores of the terms are 0.573, 0.6, 0.412, 0.592, 0.552 respectively, and the term "wellbore integrity evaluation" is taken as the standard expression, and the rest of the terms are taken as the general expression;

[0113] The "wellbore integrity evaluation" is associated with the highest frequency data unit MPa and the typical magnitude range 35-70, and the "wellbore integrity evaluation" is taken as the parent node, and the rest of the terms are taken as the child nodes, and all the child nodes inherit the attributes (data unit, data amount and detection index set) of the parent node to construct the semantic mapping relationship;

[0114] The content increment model is used for normalizing the input content, the input "wellbore health check" ontology recognition library is directly calibrated as the general expression, the input mapping engine is optimized as the standard expression "wellbore integrity evaluation", and is input to the unit normalizer to extract the specified data unit and data magnitude (MPa, typical magnitude range 35-70), and the input data is compared, and the optimized data (55MPa) conforming to the standard is output to the full connection layer.

[0115] In the embodiment, the method for obtaining the text material and the numerical material corresponding to each sentence includes:

[0116] When the paragraph format of the to-be-edited material is obvious, the text separator is recognized and the paragraphs are divided, the end-of-sentence punctuation of each paragraph is recognized by using a regular expression, and the sentences are segmented at the punctuation positions by using an NLTK sentence segmentation function.

[0117] When the paragraph format of the to-be-edited material is not obvious, the sentences are directly segmented according to the end-of-sentence punctuation of the material, the expression content of each sentence is determined by performing semantic recognition on the sentences, and the sentences with the same expression content are divided into the same paragraph; the expression content includes background research, market summary, standard , operation method, data analysis, development goal and role significance.

[0118] The numerical mode of each sentence of the to-be-edited material is recognized by using a regular expression, the numerical material is extracted from each sentence, the remaining material is taken as the text material, and the numerical material is associated with the corresponding text material;

[0119] In actual evaluation, the obvious distinguishing marks of the paragraph format of the material to be compiled are carriage return, first line indentation format, page break, section break, etc.

[0120] In the embodiment, the method for determining the first matching information comprises:

[0121] The rule-based pattern matching engine is used to identify the explicit description and implicit intention of the user-written requirements to obtain the page size and page layout, the keyword trigger-context verification mechanism is used to determine the file format of the user-written requirements, the field-adapted BERT classifier is used to output the field label, and the format requirements are composed of the page size, page layout, file format and field label.

[0122] The user-written requirements are divided into multiple text units according to punctuation marks, the core topic cluster of each text unit is identified through a hierarchical Dirichlet process, the topic coherence of each text unit is analyzed according to the core topic cluster, adjacent text units with consistent core topics are merged, the core topic cluster of the merged text unit is taken as the paragraph topic, and the paragraph order is obtained by analyzing the temporal dependence of the merged text unit, the first semantic feature is composed of the paragraph topic and the paragraph order.

[0123] The text material of the material to be compiled is input into the context-aware recommendation engine to obtain the text level and the text term frequency; the text level is determined according to the inclusion relationship of the text.

[0124] The text outline size is determined according to the text level of each paragraph and the number of paragraphs, the text terms are classified and the cumulative frequency of the same type of term is counted, the ratio of the cumulative frequency of the same type of term to the corresponding standard frequency is calculated to obtain the term professional degree, the mean value of the professional degrees of all types of terms in the field is taken as the field professional degree, the text terms are input into the field knowledge graph to obtain the corresponding associated clauses, the maximum value of the associated clause level of the same type of text term is taken as the term association degree, and the mean value of the association degrees of all types of terms in the field is taken as the field association degree.

[0125] The text outline size, field professional degree and field association degree are combined to form the granularity feature, and the first matching information is obtained by splicing the first semantic feature and the granularity feature.

[0126] In actual evaluation, taking the oil field material to be compiled as an example, when obtaining the format requirements, the rule-based pattern matching engine is used to identify the explicit description (A4 size) and implicit intention (including chart area) to obtain the page size (A4 vertical layout) and page layout (column layout), the direct trigger word library (PDF / DOCX / XML, etc.) is combined with the context semantic verification (such as "editable" corresponding to DOCX, "read-only" corresponding to PDF) to obtain the file format, and the user-written requirements are input into the field-adapted BERT classifier to obtain the oil field label (while combining the industry knowledge graph to verify the consistency of field terms);

[0127] When acquiring the first semantic feature, the to-be-written material is divided into five text units according to semicolons and periods as the division basis, and the core theme clusters of each text unit are identified as "standard drilling parameters", "core drilling parameters", "safety standards", "safety design", "emergency response", etc.); the theme coherence is analyzed (the keyword similarity and functional correlation of adjacent text units are weighted to calculate the theme coherence score, and the weight is 0.5, and when the theme coherence score is greater than the set threshold 0.7, the text unit is merged), semantic unit 1 and semantic unit 2 are merged, and semantic unit 3 and semantic unit 4 are merged to obtain three paragraph themes; the order of the paragraphs is obtained by performing time sequence dependence analysis on the merged text units according to conjunctions (first, second, etc.), verb tenses (past tense, present tense) and logical operators (therefore, because): "safety design"-"drilling parameters"-"emergency response";

[0128] In the context-aware recommendation engine, the number of text inclusion relationships of each paragraph of the to-be-written material 3 is 2, 3 and 2, and the text outline size is determined as [2, 3, 2];

[0129] The text terms of the to-be-written material are counted, wherein the cumulative frequencies of the drilling parameter terms (drilling pressure, rotation speed, etc.), the safety equipment terms (blowout preventer, gate plate, etc.) and the emergency response terms (H2S monitoring, well shut-in procedure, etc.) are 18, 20 and 11, the corresponding standard frequencies are 15, 10 and 8, and the field professional degree is 1.525;

[0130] The text terms of the to-be-written material are input into the field knowledge graph, the maximum values of the association clause levels of the various text terms are 3, 2 and 1 respectively, the field association degree is 2, and the granularity feature [2, 3, 2, 1.525, 2] is obtained.

[0131] In this embodiment, the method for determining the second matching information comprises:

[0132] According to the expression content of each sentence text material, the paragraph positioning is determined according to the paragraph theme, the relationship between the sentences in the same paragraph is identified by constructing a ternary relationship classifier, the sentences in the same paragraph are sorted according to the appearance order of the to-be-written material, and the absolute position coding is input into the position coding-aware Transformer to obtain the sentence order. The second semantic feature is composed of the paragraph positioning, the sentence relationship and the sentence order; the relationship between the sentences includes the reference relationship, the emphasis relationship and the leading relationship;

[0133] The second semantic feature and the granularity feature are spliced to obtain the second matching information;

[0134] In the actual evaluation, taking the materials to be compiled in the oil field as an example, when obtaining the second semantic feature, the text material is mainly "drilling specification", the paragraph theme is directly matched with "drilling parameter" to determine that the paragraph positioning is the second paragraph; the relationship between sentences includes reference relationship (determined by marking symbol "[1]", verb "see" / "according to", etc.), emphasis relationship (determined by adverb "especially" / "particularly", punctuation "!", etc.) and leading relationship (determined by transition word "however" / "further", question "why", etc.); In the position coding perception Transformer, multiple groups of sentence orders are generated, and then the rationality scores of the sentence orders are calculated (determined by weighting the semantic coherence score and the temporal consistency score), and the sentence order with the highest rationality score is output.

[0135] In the embodiment, the method for conducting compliance review comprises:

[0136] The text material of the material to be compiled is input into the field knowledge graph for compliance review, and the specific steps comprise:

[0137] The compliance of the text terms in the text material in the field knowledge is checked, and the text terms that are not compliant are removed, replaced and supplemented to obtain compliant text;

[0138] The recommended text is obtained by recommending the text material not explicitly mentioned by the user through graph relationship reasoning; the basis for the graph relationship reasoning is that the text terms corresponding to the nodes have strong relationship edges, and the text terms corresponding to the other nodes of the strong relationship edges do not appear in the material to be compiled; the edge weight of the most recently updated edge of the strong relationship edge is greater than 0.85; the recommended text includes related mandatory provisions and best practices;

[0139] The specification numerical range of the nodes where the compliant text and the recommended text are located is extracted, the field numerical multi-sphere is constructed according to the specification numerical range, the numerical material of the material to be compiled is input into the field numerical multi-sphere for abnormal screening and abnormal correction to obtain compliant numerical values; the abnormal correction specifically means that the abnormal numerical values determined by the abnormal screening are manually verified, and the abnormal numerical values are adjusted according to the manual verification results;

[0140] In the actual evaluation, taking the materials to be compiled in the oil field as an example, in the compliance check, the field knowledge graph finds that the text material appears "R type sealing element", which does not conform to the expression in the field knowledge, and "API 16A RX type sealing element" is used to replace the term according to the associated knowledge of the field knowledge graph;

[0141] When the recommended texts are obtained, the strong relation edge "remote control requirements" of "blowout preventer" has a weight of 0.92 and the text term corresponding to the other node of the strong relation edge does not appear in the material to be compiled, so the recommended text "accumulator capacity ≥ 80L" is added; the strong relation edge "pressure test standard" of "API16A" has a weight of 0.91 and the text term corresponding to the other node of the strong relation edge does not appear in the material to be compiled, so the recommended text "35MPa hydrostatic test required" is added; the strong relation edge "environmental protection requirements" of "hydraulic oil" has a weight of 0.87 and the text term corresponding to the other node of the strong relation edge does not appear in the material to be compiled, so the recommended text "hydraulic oil biodegradability ≥ 90%" is added.

[0142] Based on the numerical range specified in the standard, a domain numerical multisphere was constructed. The numerical data of the materials to be compiled were screened for anomalies, and "working pressure -28MPa", "closing time -50s", and "oil temperature -70℃" were identified as abnormal data. After manual verification, the corresponding errors were found to be "design drawing error", "sensor failure" and "tropical sea area special case" respectively. The first two data were corrected to "working pressure -35MPa" (corrected according to API 16A Class 700) and "closing time -28s" (corrected according to the actual test report). The third data was noted as "tropical sea area special case".

[0143] In this embodiment, the method for obtaining standardized materials includes:

[0144] The incremental model inputs compliant text and recommended text into content, transforms general expressions in the text into standardized expressions to obtain standardized text, extracts the data units and data volume associated with the standardized text, adjusts the data units and optimizes the data volume according to the compliance values ​​to obtain standardized data, and combines the standardized text and standardized data to form standardized materials;

[0145] The standard template position for each sentence is determined based on the second semantic feature. The standard template material for each sentence is then input into the standard template to obtain the standard document.

[0146] In the actual evaluation, the cosine similarity between the first matching information and the paragraph representation of the paragraph template to be matched is calculated, and the paragraph template to be matched with the highest cosine similarity (0.88) is taken as the standard paragraph structure. The cosine similarity between the second matching information and the statement representation of the statement template to be matched is calculated, and the statement template to be matched with the highest cosine similarity (0.85) is taken as the standard statement structure. A standard template is generated according to the format requirements, standard paragraph structure and standard statement structure.

[0147] The standard template position for each sentence is determined based on the second semantic feature. The standard template material for each sentence is then input into the standard template to obtain the standard document.

[0148] In a second aspect, a standard document writing system based on knowledge retrieval enhancement comprises:

[0149] a model module configured to perform knowledge retrieval to obtain domain multi-source heterogeneous knowledge and construct a domain knowledge graph, divide the domain multi-source heterogeneous knowledge according to general expression and standard expression, and construct a content increment model according to a division result;

[0150] a material analysis module configured to perform semantic extraction on a user writing requirement to obtain a format requirement and a first semantic feature, perform structural division and content segmentation on a material to be written to obtain text material and numerical material corresponding to each sentence, perform semantic extraction on the text material to obtain a second semantic feature, and perform text granularity perception on the text material to obtain a granularity feature;

[0151] a template module configured to determine first matching information according to the first semantic feature and the granularity feature, and determine second matching information according to the second semantic feature and the granularity feature, determine a standard paragraph structure according to the first matching information, determine a standard sentence structure according to the second matching information, and generate a standard template according to the format requirement, the standard paragraph structure, and the standard sentence structure;

[0152] a review module configured to input the material to be written into the domain knowledge graph to perform compliance review to obtain a compliant text, a recommended text, and a standard numerical range, construct a domain numerical multi-sphere according to the standard numerical range to perform abnormal screening and abnormal correction to obtain a compliant numerical value;

[0153] a writing module configured to input the compliant text, the recommended text, and the compliant numerical value into the content increment model to perform standardized conversion to obtain a standard material, and input the standard material of each sentence into the standard template to obtain a standard document.

[0154] The above merely describes preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A standard document writing method based on knowledge retrieval enhancement, characterized in that: Includes the following steps: S1. Perform knowledge retrieval to obtain multi-source heterogeneous knowledge in the domain and construct a domain knowledge graph. Divide the multi-source heterogeneous knowledge in the domain according to general and standard expressions, and construct a content incremental model based on the division results. S2. Semantic extraction is performed on the user's writing requirements to obtain format requirements and first semantic features. The material to be written is structurally divided and content segmented to obtain the text material and numerical material corresponding to each sentence. The structural division includes paragraph division and sentence division. S3. Semantic extraction is performed on the text material to obtain a second semantic feature, text granularity perception is performed on the text material to obtain granularity features, a first matching information is determined based on the first semantic feature and the granularity feature, and a second matching information is determined based on the second semantic feature and the granularity feature; S4. Determine the standard paragraph structure based on the first matching information, determine the standard sentence structure based on the second matching information, and generate a standard template based on the format requirements, the standard paragraph structure, and the standard sentence structure. S5. Input the material to be compiled into the domain knowledge graph for compliance review to obtain compliant text, recommended text and standard numerical range. Construct a domain numerical multisphere based on the standard numerical range to perform anomaly screening and anomaly correction to obtain compliant values. S6. Input the compliant text, the recommended text, and the compliant value into the content incremental model for standardized transformation to obtain standardized materials, and input the standardized materials of each sentence into the standard template to obtain a standard document; The method for constructing the content incremental model includes: Multi-source heterogeneous knowledge from various domains is input into a pre-trained bag-of-words model to obtain different terms for the same entity, forming an entity term set. The authority frequency and information entropy of each term in external authoritative data and external academic data are calculated to obtain a normative score. The term with the highest normative score in the entity term set is taken as the normative expression, and the remaining terms in the entity term set are general expressions. The expression for the standardized score is: ； in For the term Standardized scoring For authoritative frequency weighting, For information entropy weights, Weighting based on user usage frequency. For the term The frequency of occurrence in authoritative external data and external academic data For the term The context in which it occurs For the term Context in which it occurs The probability of occurrence For the term Frequency of occurrence in internal unstructured data; By associating the most frequently used data units and data volumes in the BERT perception model with the canonical expressions, and using the domain ontology as a framework, a semantic mapping relationship is constructed with the canonical expressions as parent nodes and the general expressions as child nodes. Based on the segmentation results, a content increment model is constructed. The specific structure includes an ontology recognition library, a mapping engine, a unit normalizer, and a fully connected layer. The ontology recognition library is used to label the input text as either a standardized or general expression. The standardized expression is directly input into the unit normalizer and the fully connected layer, while the general expression is input into the mapping engine. The mapping engine embeds semantic mapping relationships, transforming the general expression into a standardized expression, which is then input into the unit normalizer and the fully connected layer. The unit normalizer extracts the data units and data magnitude associated with the standardized expression and compares them with the input data, adjusting the data units and optimizing the data magnitude to output optimized data. The fully connected layer connects the ontology recognition library, the mapping engine, and the unit normalizer to output standardized content increments. The standardized content increments include the standardized expression and its corresponding data units and data magnitudes. The method for determining the standard paragraph structure is as follows: calculate the cosine similarity between the first matching information and the paragraph representation of the paragraph template to be matched, and take the group of paragraph templates with the highest cosine similarity as the standard paragraph structure; the paragraph representation includes the paragraph's theme, paragraph order, number of paragraphs, and text outline size; The method for determining the standard sentence structure is as follows: calculate the cosine similarity between the second matching information and the sentence representation of the sentence template to be matched, and take the group of sentence templates with the highest cosine similarity as the standard sentence structure; the sentence representation includes the paragraph location, sentence relationship, sentence order, domain specialization and domain relevance of the sentence; The domain numerical multisphere is constructed based on a standardized numerical range; the standardized numerical range is extracted from the nodes containing the compliant text and the recommended text. The anomaly correction specifically refers to manually verifying the abnormal values ​​identified through anomaly screening and adjusting the abnormal values ​​based on the results of the manual verification.

2. The standard document writing method based on knowledge retrieval enhancement according to claim 1, characterized in that, The method for constructing a domain knowledge graph includes: Knowledge retrieval is performed by collecting internal structured data, internal unstructured data, external authoritative data, external academic data, and real-time data streams. The collected data is then subjected to format standardization, entity recognition and cleaning, relation annotation, and knowledge fusion to obtain multi-source heterogeneous knowledge in the domain. Core knowledge classes and edge relations are defined based on the multi-source heterogeneous knowledge in the domain. Structured data is mapped to data triples according to (entity, attribute, value). The BERT-BiLSTM-CRF joint model is used to extract entity relations to obtain text triples. The graph database Neo4j is used to store the data triples and text triples to construct a domain knowledge graph. The domain knowledge graph employs a hybrid pruning algorithm, dynamic weight rebalancing, and a dual-channel reinforcement mechanism for graph optimization. The hybrid pruning algorithm specifically includes semantic density pruning and topological importance pruning. The dual-channel reinforcement mechanism includes a real-time channel and a batch channel. The semantic density pruning uses adaptive Gaussian mixture clustering to divide semantic clusters, and selects representative nodes for each cluster based on a density-dispersion balance condition, expressed as follows: ； ； in For the first semantic clusters The representative nodes that are retained are those that satisfy the maximum value of the density-dispersion balance condition. For node set The first in 1 node For the number of nodes, For the first semantic clusters scale For nodes With nodes semantic tightness, The equilibrium factor for dynamic adjustment of KL divergence. , For nodes and Embedded vector, Selecting conditions for cluster centers For cluster index, For the number of clusters, For nodes Location of cluster center semantic tightness, The regularization coefficient is . semantic clusters The prior probability; The topological importance pruning method improves the PageRank algorithm by introducing relation type weights and time decay factors to calculate the importance score of each node, and retains nodes in the critical path whose importance score is less than the importance score threshold. The expression is as follows: ； in For nodes Importance score For neighboring nodes Importance score The damping coefficient is... For nodes The set of incoming neighbors, node Pointing to node edge The basic weights, For relational strengthening functions, based on the edge Type Adjust the weights, For relation type Embedded vector, For a learnable projection matrix, For the target node eigenvectors, For nodes The set of outgoing neighbors, For nodes Pointing to node edge The basic weights, For decay rate parameters, The current system time. for timestamp, As a reward factor for recent events, For the indicator function, when the timestamp The value is 1 if the time threshold is exceeded, and 0 otherwise. The dynamic weight rebalancing is obtained based on the graph structure update after hybrid pruning, and the expression is: ； ； in side Rebalancing weights For nodes With nodes semantic similarity weights, For nodes With nodes Topological association weights, This is the node importance score vector. For the embedding matrix gradient operator, Nodes in the diagram after pruning With nodes Meta-path similarity, Nodes in the original graph With nodes Meta-path similarity, Temperature coefficient; The real-time channel reconstructs and optimizes the graph structure through event-driven subgraph reconstruction. The batch channel uses incremental reinforcement learning to decay and enhance node weights and adjusts the importance score threshold in topological importance pruning.

3. The standard document writing method based on knowledge retrieval enhancement according to claim 1, characterized in that, The method for obtaining the text material and numerical material corresponding to each sentence includes: When the paragraph format of the material to be compiled is obvious, the text delimiters are identified and the paragraphs are divided. Regular expressions are used to identify the punctuation marks at the end of each paragraph, and NLTK sentence segmentation function is used to segment sentences at the punctuation mark positions. When the paragraph format of the material to be compiled is not obvious, the sentences are directly broken down according to the punctuation marks at the end of the sentences. The semantics of each sentence are used to determine the content of each sentence, and sentences with the same content are divided into the same paragraph. The content includes background research, market overview, standards and norms, operating methods, data analysis, development goals and significance. Regular expressions are used to identify the numerical patterns of each statement in the material to be compiled, the numerical data are extracted from each statement, the remaining data are treated as text data, and the numerical data are associated with the corresponding text data.

4. The standard document writing method based on knowledge retrieval enhancement according to claim 1, characterized in that, The method for determining the first matching information includes: A rule-based pattern matching engine is used to identify the explicit description and implicit intent of the user's writing requirements to obtain the page size and page layout. A keyword-triggered-context verification mechanism is used to determine the file format of the user's writing requirements. A domain-adapted BERT classifier is used to output domain labels. The format requirements consist of page size, page layout, file format and domain labels. The user-written requirements are divided into multiple text units according to punctuation marks. The core topic clusters of each text unit are identified through a hierarchical Dirichlet process. Thematic coherence analysis is performed based on the core topic clusters of each text unit. Adjacent text units with the same core topic are merged. The core topic clusters of the merged text units are taken as paragraph topics. Temporal dependency analysis is performed on the merged text units to obtain the paragraph order. The paragraph topics and paragraph order constitute the first semantic feature. The text material to be written is input into a context-aware recommendation engine to obtain the text level and the frequency of text terms; the text level is determined based on the inclusion relationship of the text. The text outline size is determined based on the text level and number of paragraphs. Text terms are classified and the cumulative frequency of terms in the same category is counted. The ratio of the cumulative frequency of terms in the same category to the corresponding standard frequency is calculated to obtain the terminology professionalism. The average of the professionalism of all categories of terms in this field is taken as the domain professionalism. Text terms are input into the domain knowledge graph to obtain the corresponding related clauses. The maximum level of the related clauses of the same category of text terms is taken as the terminology relevance. The average of the relevance of all categories of terms in this field is taken as the domain relevance. The text outline size, domain specialization, and domain relevance are combined to form granular features. The first semantic feature and the granular feature are then concatenated to obtain the first matching information.

5. The standard document writing method based on knowledge retrieval enhancement according to claim 1, characterized in that, The method for determining the second matching information includes: Paragraph location is determined by matching paragraph themes with the content of each sentence in the text material. A ternary relation classifier is constructed to identify the relationships between sentences in the same paragraph. The sentences in the same paragraph are sorted according to the order of appearance of the material to be compiled. After absolute position encoding, they are input into a position encoding-aware Transformer to obtain the sentence order. The second semantic feature is composed of paragraph location, sentence relationship, and sentence order. The relationships between sentences include citation relationship, emphasis relationship, and introduction relationship. The second semantic feature and the granular feature are concatenated to obtain the second matching information.

6. The standard document writing method based on knowledge retrieval enhancement according to claim 1, characterized in that, The methods for conducting compliance reviews include: The text materials to be compiled are input into a domain knowledge graph for compliance review. The specific steps include: Check the compliance of textual terms in the textual materials with the domain knowledge, and remove, replace and supplement non-compliant textual terms to obtain compliant text; Recommended texts are obtained by inferring textual materials not explicitly mentioned by the user through graph relation reasoning; the graph relation reasoning is based on the existence of strong relation edges between the nodes corresponding to the text terms and the fact that the text terms corresponding to the other node of the strong relation edge do not appear in the material to be compiled; the edge weight of the most recently updated strong relation edge is greater than 0.85; the recommended texts include relevant mandatory clauses and best practices; The numerical data of the material to be compiled is input into the domain numerical multisphere for anomaly screening and correction to obtain compliant values.

7. The standard document writing method based on knowledge retrieval enhancement according to claim 1, characterized in that, The method for obtaining standardized materials includes: The incremental model inputs compliant text and recommended text into content, transforms general expressions in the text into standardized expressions to obtain standardized text, extracts the data units and data volume associated with the standardized text, adjusts the data units and optimizes the data volume according to the compliance values ​​to obtain standardized data, and combines the standardized text and standardized data to form standardized materials; The standard template position for each sentence is determined based on the second semantic feature. The standard template material for each sentence is then input into the standard template to obtain the standard document.

8. A standard document writing system based on knowledge retrieval enhancement, used to perform the method according to any one of claims 1-7, characterized in that, include: Model module: Used for knowledge retrieval to obtain multi-source heterogeneous knowledge in the domain and construct a domain knowledge graph. It divides the multi-source heterogeneous knowledge in the domain according to general and standardized expressions, and constructs a content incremental model based on the division results. Material Analysis Module: Used to extract the semantics of user-written requirements to obtain format requirements and primary semantic features; This is used to perform structural division and content segmentation of the material to be compiled to obtain the text material and numerical material corresponding to each sentence; the structural division includes paragraph division and sentence division; it is used to perform semantic extraction on the text material to obtain second semantic features, and to perform text granularity perception on the text material to obtain granular features; Template module: used to determine first matching information based on the first semantic feature and the granular feature, and to determine second matching information based on the second semantic feature and the granular feature; Used to determine a standard paragraph structure based on the first matching information, to determine a standard sentence structure based on the second matching information, and to generate a standard template based on the format requirements, the standard paragraph structure, and the standard sentence structure; The review module is used to input the materials to be compiled into the domain knowledge graph for compliance review to obtain compliant text, recommended text and standard numerical range. Based on the standard numerical range, a domain numerical multisphere is constructed to perform anomaly screening and anomaly correction to obtain compliant values. The writing module is used to input the compliance text, the recommended text, and the compliance value into the content incremental model for normalization and transformation to obtain standardized materials, and to input the standardized materials of each sentence into the standard template to obtain a standard document; The content increment model includes an ontology recognition library, a mapping engine, a unit normalizer, and a fully connected layer. The ontology recognition library identifies the input text as either a standardized or general expression. The standardized expression is directly input into the unit normalizer and the fully connected layer, while the general expression is input into the mapping engine. The mapping engine embeds semantic mapping relationships, transforming the general expression into a standardized expression, which is then input into the unit normalizer and the fully connected layer. The unit normalizer extracts the data units and data magnitude associated with the standardized expression and compares them with the input data, adjusting the data units and optimizing the data magnitude to output optimized data. The fully connected layer connects the ontology recognition library, the mapping engine, and the unit normalizer to output standardized content increments. The standardized content increments include the standardized expression and its corresponding data units and data magnitudes. The method for determining the standard paragraph structure is as follows: calculate the cosine similarity between the first matching information and the paragraph representation of the paragraph template to be matched, and take the group of paragraph templates with the highest cosine similarity as the standard paragraph structure; the paragraph representation includes the paragraph's theme, paragraph order, number of paragraphs, and text outline size; The method for determining the standard sentence structure is as follows: calculate the cosine similarity between the second matching information and the sentence representation of the sentence template to be matched, and take the group of sentence templates with the highest cosine similarity as the standard sentence structure; the sentence representation includes the paragraph location, sentence relationship, sentence order, domain specialization and domain relevance of the sentence; The domain numerical multisphere is constructed based on a standardized numerical range; the standardized numerical range is extracted from the nodes containing the compliant text and the recommended text. The anomaly correction specifically refers to manually verifying the abnormal values ​​identified through anomaly screening and adjusting the abnormal values ​​based on the results of the manual verification.

Images