A large model-based knowledge graph construction method

Abstract

The application relates to the technical field of data model construction, and discloses a knowledge graph construction method based on a large model, which comprises the following steps: collecting and dividing the fields of data in a database, judging the meaning of each word in each field, judging whether each word has a special meaning in each field, testing and analyzing the marked words, generating a first-level knowledge graph, and analyzing the importance of the words in the fields to generate a second-level knowledge graph. The knowledge graph construction method based on the large model realizes the systematic integration and display of knowledge in different fields, improves the organization and management efficiency of the knowledge, accurately analyzes the semantics and special meanings of the words in different fields, improves the accuracy and reliability of the knowledge graph, enhances the reliability and authority of the information in the knowledge graph, realizes the personalized display of the knowledge graph, meets the needs of users at different levels, improves the user experience, helps users quickly obtain required information, and improves work efficiency.

Description

Technical Field

[0001] This invention relates to the field of data model construction technology, specifically a method for constructing a knowledge graph based on a large model. Background Technology

[0002] With the rapid development of artificial intelligence technology, large models have demonstrated powerful capabilities in the field of natural language processing, especially in text understanding, generation, and reasoning. Knowledge graphs, as a structured form of knowledge representation, can connect discrete knowledge points to form a vast knowledge network, providing support for various AI applications. Combining the powerful capabilities of large models with the structured advantages of knowledge graphs can greatly improve the efficiency, quality, and automation of knowledge graph construction. Traditional knowledge graph construction methods mainly rely on rule matching, statistical methods, or deep learning-based sequence labeling and relation extraction techniques. These methods often require a large amount of labeled data and complex feature engineering, and perform poorly for new domains or unseen entity relationships. Knowledge graph construction methods based on large models aim to leverage the powerful generalization ability and rich world knowledge of large models to simplify the construction process and improve construction quality. By leveraging the powerful language understanding and knowledge reasoning capabilities of large models, knowledge graph construction methods based on large models significantly improve the automation, efficiency, and flexibility of knowledge graph construction.

[0003] Existing knowledge graph construction methods based on large models cannot determine whether each word has a special meaning in each domain, such as a word representing a specific professional term in a certain domain. For words with special meanings, it is impossible to determine their importance in the corresponding domain, generate different knowledge graphs based on the importance of words in the corresponding domain, or display different knowledge graphs according to the user's situation. This makes it inconvenient and ineffective for users to view the information, resulting in the inability to use some data sources that use special words as professional terms, low utilization rate, and certain limitations in its practicality. Summary of the Invention

[0004] This invention provides a knowledge graph construction method based on a large model, which can help solve the problems mentioned in the background art.

[0005] This invention provides the following technical solution: a knowledge graph construction method based on a large model, comprising:

[0006] The data in the database is collected and classified into different domains.

[0007] For each word in the database, perform word domain analysis to determine the meaning of the word in each domain;

[0008] For each word in the database, perform special meaning analysis to determine whether the word has a special meaning in each domain;

[0009] For each word, obtain its tag information;

[0010] If words contain marker information, then the marked words are examined and analyzed;

[0011] For words that still retain labeled information after verification and analysis, a first-level knowledge graph is generated, and the importance of the words in the domain is analyzed.

[0012] Generate a two-level knowledge graph based on the importance of words within the domain;

[0013] If a word does not have any tagging information, no operation will be performed on that word.

[0014] As an optional scheme of the knowledge graph construction method based on a large model described in this invention, the corpus collection and domain division are specifically as follows:

[0015] S1. Define a set of domains: ; among them, each Representing a field, S1 represents the total number of domains; S2 represents the total number of domains. The semantic scope of the domain is initially defined, denoted as . S3, by extraction The word vectors of keywords and phrases are obtained and average pooling is performed to calculate the initial feature vector of semantic scope. : ;in, For words Word vectors, semantic range The number of elements in; S4, collect domain sample data, denoted as S5. Through statistics Words with high TF and IDF frequencies are used to extract candidate sets of feature dimensions. : ;in, For the set threshold, For sample data The processed text set obtained after preprocessing; S6, for each domain Build a domain knowledge base ; Utilizing domain knowledge bases Define the domain Set of concepts: ; among them, each Representative fields A key concept or term; from Filtering out relevant fields A set of highly relevant feature dimensions: ;in, This is the set of filtered feature dimensions. For the candidate set of feature dimensions, For the field A collection of concepts Representing feature dimension With domain concept similarity, This indicates selecting the highest similarity value from a set of similarity values. S7. Calculate the weights of each feature dimension using principal component analysis or other feature selection algorithms, based on the set similarity threshold. Importance weights in: S8, Construction Domain The final feature vector: ;in, The number of feature dimensions, It is the first Each feature dimension It is its corresponding weight; S9, based on the feature vector Update the field semantic scope: ;in, S10. Set an update decision function to determine whether the feature vector and semantic range are appropriate: (The text then goes on to describe the inclusion of high-weight feature dimensions within the semantic scope, but the provided text is incomplete and contains several errors.) ;in, To represent the difference between sets, Indicates belonging to a set But it does not belong to a set. A set of elements and The set convergence threshold, Let be the number of iterations; if If the feature vector and semantic range tend to stabilize, the repetition stops, and each domain is obtained. The semantic range and feature vector are denoted as follows: and ;like If the feature vector and semantic range have not stabilized, repeat steps S5-S10; S11, for each domain Textual data was collected from multiple sources to form a corpus: ;in, Represents a text segment, The amount of corpus collected for this field; S12, Define a corpus collection diversity index: ;in, For the set of corpus source types, Corpus source type The weight, Source type In the field The amount of corpus in the data; S13, set a corpus collection judgment function to determine whether the corpus is sufficient: ;in, The threshold for target diversity in corpus collection; if If the data collection in this domain meets the diversity requirement, then it is determined that further data collection in this domain is unnecessary; if... If the data collection for that domain does not meet the diversity requirements, then it is determined that the data collection for that domain needs to continue; S14, Statistical analysis for each domain Semantic relevance metrics of the collected corpus: ;in, Represents a text fragment With field Feature vector Similarity; S15, for each domain Define a corpus quality judgment function to determine whether the quality of the corpus meets the requirements: ;in, The threshold for semantic relevance of the corpus; if If the collected corpus in this domain meets the semantic relevance requirement, then the collected corpus will not be processed again; if If the collected data does not meet the semantic relevance requirements, the collected data will be filtered to remove data that does not match semantically.

[0016] As an optional scheme of the knowledge graph construction method based on a large model described in this invention, the following is included: word domain analysis, specifically: for each word For each field Words are obtained through word vector models. Basic semantic vector ;calculate With domain feature vectors Similarity: Adjusting the semantic vectors of words based on similarity yields the semantic vectors of the words within the domain. Corrected semantic vector in: ;in, It is an adjustment factor; for both domains and Calculate words Differences in semantic vectors between these two domains: Define a semantic bias determination function to determine whether a word has a significant semantic bias: ;in, This is the semantic deviation threshold; if Then determine the word In the field and There are significant semantic discrepancies between them; if Then determine the word In the field and There is no significant semantic deviation between them; if words In the field and If there are significant semantic discrepancies, then the domain knowledge base should be utilized. Find the words The concepts that are semantically closest, among which, : ; Obtain the concept The semantic vector, denoted as Based on the semantic meaning of concepts in the domain knowledge base, the words are... The semantic vectors are adjusted to better reflect the domain. Specific semantics: ;in, It is the adjustment factor; calculate the adjusted semantic vector. domain semantic feature vector Similarity: Define an adjustment judgment function to determine whether the semantic adjustment was successful: ;in, The set similarity threshold; if If the semantic adjustment is successful, the word is confirmed. In the field The semantic adjustment results in; if If the semantic adjustment fails, the concept is removed, and the domain knowledge base is utilized. Find the word again The concept with the closest semantics Calculate its semantic vector and determine whether the semantic adjustment was successful, until the semantic adjustment is deemed successful; update the words in the domain. Correction semantic vectors in For each field Utilize authoritative domain knowledge bases to construct a domain concept system: ;in, Representation of domain One of the concepts, For the field The number of concepts in; for words in the domain the corrected semantic vector , calculate its similarity with each concept in the domain : ; where is the semantic vector of concept ; set a concept correspondence determination function to judge whether there is a corresponding concept for the word in the domain: ; where is the domain concept mapping threshold; if , it is determined that there is a corresponding concept for the word in the domain, and the word in the domain corresponds to the concept ; record the word and its corresponding concept in the domain to form a word-domain concept mapping relationship: ; if , it is determined that there is no corresponding concept for the word in the domain, and no operation is performed.

[0017] As an optional solution of the knowledge graph construction method based on the large model described in the present invention, where: the analysis of the special meaning of words is specifically: for each domain , define a domain term set: ; where represents a professional term in this domain, is the number of terms;

[0018] Set a term matching judgment function, for each word in the database, judge whether it belongs to the term set of the domain : ; if , it is determined that the word is a professional term in the domain , has a special meaning, mark the word and the domain , generate marking information: ; save the description of the special meaning; if , it is determined that the word is not a professional term in the domain , has no special meaning, and no marking information is generated;

[0019] For each domain , construct a domain text corpus ; for the word In the field Text corpus In Chinese, calculate words In the field The frequency of occurrence in the text is defined as the usage frequency: Statistical terms In the field Common collocation patterns in text; find the common collocation patterns in text. The set of frequently co-occurring words And calculate its support and confidence indicators; analyze words In the field Contextual features in text: According to the words In the field Feature calculation comprehensive score: ;in, , as well as The comprehensive weighting coefficient is used to express the judgment in a special way. To use frequency weights, To match the mode weights, For context weighting, Indicates collocation pattern Confidence level;

[0020] Set up a special expression judgment function to judge words. In the field Does it contain any special expressions?

[0021] ;in, For special expression judgment thresholds; if Then determine the word In the field There are special expressions within memory that have special meanings; these terms and fields are marked. Record tagging information: ; Save descriptions with special meanings; if Then determine the word In the field There is no special expression or meaning within it, and no tagging information is generated.

[0022] As an optional scheme of the knowledge graph construction method based on a large model described in this invention, the following is included: performing verification analysis on the marked words, specifically: for each word with a mark... and its marked areas Define a multi-source data set: ; among them, each Represents a data source. The number of data sources; based on the reliability, authority, and relevance of each data source. Set weights : ;in, For reliability weighting coefficients, This is an authoritative weighting coefficient. This refers to the relevance weighting coefficient.

[0023] Define a set of support relationships: ; among them, each This indicates a support relationship;

[0024] For each data source Define its relation to words In the field The collection of descriptions and explanations within ;

[0025] For each data source Determine its description and explanation set. Does there exist at least one element in the set of supporting relationships? Element matching in: ;

[0026] like Then determine the data source. Clearly indicate the words In the field It has a special meaning;

[0027] like Then determine the data source. Clearly indicate the words In the field The content does not have any special meaning; if the data source Clearly indicate the words In the field If it has a special meaning, then record it. Otherwise, remember ;

[0028] Calculate words In the field Comprehensive verification score for special internal meanings: ;

[0029] Set up a credibility determination function to judge the words In the field Do the special meaning markers within have high credibility? ;in, To comprehensively verify the score threshold; if Then the word is considered In the field The special meaning markers within have high credibility, so retain the markers; if Then the word is considered In the field Special meaning markers within do not have high credibility; define an authoritative data source set. Non-authoritative data source collection ;in, and They do not intersect and ;

[0030] Define domain All data sources within the database for words Special meaning support set:

[0031] ;in, Indicates data source Does it support words? In the field It has a special meaning;

[0032] Computational authoritative data source set Supporting words In the field Number of data sources with special meanings: ;

[0033] Define an evaluation decision function to judge words. Whether to retain the marker:

[0034] ;

[0035] in, Indicates that the mark is reserved. This indicates that it is determined manually. Indicates canceling the mark. This is a threshold for non-authoritative support. As an authoritative support threshold, A collection of non-authoritative data sources Supporting words In the field The number of data sources with special meanings is calculated using the following formula:

[0036] .

[0037] As an optional solution to the knowledge graph construction method based on a large model described in this invention, the generation of a first-level knowledge graph specifically includes:

[0038] Create a collection of word nodes: ; among them, each Represents a word node that is determined to have a special meaning within the domain;

[0039] Create a collection of domain nodes: ; among them, each Represents a domain node with a label;

[0040] For each word node Set its property set: ;in, For word names, Basic information about words, This describes a specific meaning within the domain.

[0041] For each domain node Set its property set: ;in, For the domain name, A brief introduction to the field;

[0042] Create a set of relationships: ; among them, each This represents the semantic relationship between word nodes and domain nodes;

[0043] Define a set of relation types: ;

[0044] Iterate through all words that have a marker. and the fields marked thereon Based on the type of the tag, determine the relation type, denoted as ;

[0045] Create a relationship and add it to the relation set. ;

[0046] Combined with word node set Domain Node Set and relation sets This forms the basic structure of a first-level knowledge graph.

[0047] As an optional scheme of the knowledge graph construction method based on a large model described in this invention, the following is included: analyzing the importance of words within the domain, specifically:

[0048] For each marked field Obtaining domain-specific text corpora ;

[0049] For each word In the field Text corpus In the text, we can count its frequency of occurrence:

[0050] ;

[0051] Calculate words In the field Keyword weight in: ;

[0052] In keyword weighting, For words In the field The word frequencies in the text are as follows: ;

[0053] In keyword weighting, For words In the field Inverse document frequency in text, specifically:

[0054] ;

[0055] Statistical terms In the field Number of times it is cited ;

[0056] Calculate words In the field Overall score reflecting the importance of the content:

[0057] ;in, To use frequency weights, For keyword weight, Citation count weighting For the field The number of times the most frequently cited words are cited in a text corpus;

[0058] Domain experts on words In the field The importance of the content was assessed, and expert scores were obtained. ;

[0059] The final importance score is calculated by combining expert ratings with the overall score.

[0060] ;in, Weighting coefficients for expert ratings;

[0061] Define an importance determination function to determine the importance of words:

[0062] ;in, The importance threshold;

[0063] like Then determine the word In the field The importance level is classified as "high";

[0064] like Then determine the word In the field The importance level within is classified as "low".

[0065] As an optional solution to the knowledge graph construction method based on a large model described in this invention, the generation of a second-level knowledge graph specifically includes:

[0066] For each word node that has a tag and its associated domain nodes Based on the importance analysis results, define the node visibility attribute: ;

[0067] Get words In the field The degree of importance within;

[0068] If words In the field If the importance of the content is classified as "high", then the words will be... In the field Applications within this context are marked as displaying knowledge graph nodes. ;

[0069] If words In the field If the importance level is classified as "low", then the words will be... In the field Application scenarios within this scope are marked as hidden knowledge graph nodes. ;

[0070] Define a user set: ;

[0071] For each user Define identity attributes: ;

[0072] like If so, the user's identity is determined to be a professional user;

[0073] like If so, the user's identity is determined to be that of a regular user;

[0074] For each user Define authentication status: ;

[0075] user Submit authentication information ;

[0076] Define the conditions for approval ;

[0077] like If the authentication is successful, it is recorded as follows: ;

[0078] like If authentication fails, it is determined that the authentication was not successful and recorded as follows: ;

[0079] If the user is a professional user and has passed authorization, access to the hidden node is permitted. ;

[0080] If the user is a regular user or has not passed authentication, the hidden node will not be visible to the user.

[0081] As an optional solution to the knowledge graph construction method based on a large model described in this invention, generating a second-level knowledge graph further includes:

[0082] Get the collection of displayed nodes: ;

[0083] Get the collection of hidden nodes: ;

[0084] The set of visible nodes and the set of hidden nodes will be integrated into the node set of the secondary knowledge graph. middle: ;

[0085] The set of relationships between nodes will be displayed. The set of relationships between hidden nodes Relationship set integrated into the second-level knowledge graph middle: ;

[0086] Based on the integrated node set and relation set Constructing a two-level knowledge graph : ;

[0087] For user query requests According to user identity and node visibility properties ,like Then to the user Display the node and its related relationships;

[0088] like If the user If the user's identity is that of a professional user and they have passed authorization authentication, then the user... Display the node and its related relationships;

[0089] like If the user If the user's identity is a regular user or they have not passed authentication, no notification will be sent to the user. Show hidden nodes and their relationships.

[0090] The present invention has the following beneficial effects:

[0091] 1. This knowledge graph construction method based on a large model analyzes each word in the database to determine its meaning in each domain, such as whether the word is a common adjective or a professional term. It also determines whether each word has a special meaning in each domain, such as whether the word represents professional terms with different meanings in different domains. By using techniques such as semantic bias correction and domain concept mapping, it accurately analyzes the semantics and special meanings of words in different domains, thereby improving the accuracy and reliability of the knowledge graph.

[0092] 2. This knowledge graph construction method based on a large model marks words with special meanings and their corresponding domains. By acquiring information on the usage of these words as professional terms in various data sources, including authoritative and non-authoritative data sources, it determines whether the special meaning of the words is valid. By integrating multi-source data and conducting credibility assessment, it enhances the reliability and authority of information in the knowledge graph.

[0093] 3. This knowledge graph construction method based on a large model assesses the importance of a word within its corresponding domain for valid words with special meanings. Based on this importance, it categorizes the content associated with that word within the domain and determines whether to display it to ordinary users. If a word is widely used within the domain, it is considered important, and ordinary users can view the content associated with that word. Conversely, if a word is not widely used, it is considered less important, and the content associated with that word is hidden from ordinary users when they view the knowledge graph. This personalized display of the knowledge graph, based on the word's importance and the user's identity, meets the needs of different user levels, improves user experience, and supports efficient semantic querying and analysis, helping users quickly obtain the information they need and improving work efficiency.

[0094] 4. This knowledge graph construction method based on a large model generates a first-level knowledge graph for words and domains according to their meanings in each domain, and generates a second-level knowledge graph for words and the content associated with words in each domain. The second-level knowledge graphs are further differentiated according to the importance of words in each domain, such as showing or hiding them for different users. By constructing a hierarchical knowledge graph, the method achieves systematic integration and display of knowledge from different domains, improves the efficiency of knowledge organization and management, helps users quickly obtain the information they need, and improves work efficiency. Attached Figure Description

[0095] Figure 1 This is a flowchart of the knowledge graph construction method based on a large model according to the present invention. Detailed Implementation

[0096] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0097] Example 1: A method for constructing a knowledge graph based on a large model (see [reference]). Figure 1 ,include:

[0098] The data in the database is collected and classified into different domains.

[0099] For each word in the database, perform word domain analysis to determine the meaning of the word in each domain;

[0100] For each word in the database, perform special meaning analysis to determine whether the word has a special meaning in each domain;

[0101] For each word, obtain its tag information;

[0102] If words contain marker information, then the marked words are examined and analyzed;

[0103] For words that still retain labeled information after verification and analysis, a first-level knowledge graph is generated, and the importance of the words in the domain is analyzed.

[0104] Generate a two-level knowledge graph based on the importance of words within the domain;

[0105] If a word does not have any tagging information, no operation will be performed on that word.

[0106] By constructing a hierarchical knowledge graph, a systematic integration and display of knowledge from different domains can be achieved, improving the efficiency of knowledge organization and management. Using techniques such as semantic bias correction and domain concept mapping, the semantics and special meanings of words in different domains can be accurately analyzed, improving the accuracy and reliability of the knowledge graph. By integrating multi-source data and conducting credibility assessments, the reliability and authority of information in the knowledge graph are enhanced. Personalized display of the knowledge graph is achieved based on the importance of words in the domain and the user's identity, meeting the needs of users at different levels, improving user experience. The constructed knowledge graph supports efficient semantic querying and analysis, helping users quickly obtain the information they need and improving work efficiency.

[0107] Example 2 is an improvement upon Example 1. This method for constructing a knowledge graph based on a large model, including corpus collection and domain segmentation, specifically includes:

[0108] S1. Define a set of domains: ; among them, each Representing a field, Total number of fields;

[0109] S2, for each domain Based on actual application scenarios and needs, the semantic scope of the domain is initially defined, denoted as . The semantic scope can be a set of core topics, key concepts, etc. in the field;

[0110] S3, through extraction The word vectors of keywords and phrases are obtained and average pooling is performed to calculate the initial feature vector of semantic scope. : ;in, For words Word vectors, semantic range The number of elements in;

[0111] S4. Collect domain sample data, denoted as... This includes textual content such as professional books, academic papers, and news reports in the field;

[0112] S5. Through statistics Words with high TF and IDF frequencies are used to extract candidate sets of feature dimensions. : ;in, The set threshold is used to determine whether the data meets the requirements. For sample data The processed text collection obtained after preprocessing, which includes operations such as word segmentation, stop word removal, and stemming.

[0113] S6, For each domain Build a domain knowledge base This includes key concepts, terms and their semantic relationships within the domain. The domain knowledge base can be constructed through domain expert annotation, professional literature extraction and other methods.

[0114] Utilizing domain knowledge bases Define the domain Set of concepts: ; among them, each Representative fields A key concept or term;

[0115] from Filtering out relevant fields A set of highly relevant feature dimensions:

[0116] ;

[0117] in, This is the set of filtered feature dimensions. For the candidate set of feature dimensions, For the field A collection of concepts Representing feature dimension With domain concept similarity, This indicates selecting the highest similarity value from a set of similarity values. The set similarity threshold is used to determine whether two pieces of data are similar, i.e., whether they meet the requirements;

[0118] S7. Calculate the weights of each feature dimension using Principal Component Analysis (PCA) or other feature selection algorithms to determine the weight of each feature dimension in the domain. Importance weights in: ;

[0119] S8, Construction Domain The final feature vector: ;in, The number of feature dimensions, It is the first Each feature dimension It is its corresponding weight;

[0120] S9. Based on the eigenvectors Update the field semantic scope: ;

[0121] in, The set weight threshold is used to determine whether the data meets the pre-defined requirements and to include high-weight feature dimensions into the semantic scope.

[0122] S10. Set an update decision function to determine whether the feature vector and semantic range are appropriate:

[0123] ;in, To represent the difference between sets, Indicates belonging to a set But it does not belong to a set. A set of elements and The set convergence threshold is used to determine whether the iteration meets the stopping condition. Let be the number of iterations; if If the feature vector and semantic range tend to stabilize, the repetition stops, and each domain is obtained. The semantic range and feature vector are denoted as follows: and ;like If the feature vector and semantic range have not stabilized, repeat steps S5-S10.

[0124] S11, For each domain Textual data was collected from multiple sources to form a corpus: ;in, Represents a text segment, The amount of corpus collected for this field;

[0125] S12. Define a corpus collection diversity index to assess the diversity of corpus sources, including sources from different countries, regions, official, non-official, and authoritative sources: ;in, For the set of corpus source types, Corpus source type The weight, Source type In the field The amount of corpus in it;

[0126] S13. Define a corpus collection judgment function to determine whether the corpus is sufficient:

[0127] ;in, A threshold for the diversity of targets in the corpus is used to determine whether the corpus is sufficient; if If the data collection in this domain meets the diversity requirement, then it is determined that further data collection in this domain is unnecessary; if... If the data is not collected, it is determined that the data collection in this field does not meet the diversity requirements and it is necessary to continue collecting data in this field.

[0128] S14, Statistics for Each Field The semantic relevance metrics of the collected corpus are used to assess the degree of matching between the corpus and the semantic scope of the domain: ;in, Represents a text fragment With field Feature vector Similarity;

[0129] S15, For each field Define a corpus quality judgment function to determine whether the quality of the corpus meets the requirements: ;in, This is the semantic relevance threshold for the corpus, used to determine whether the corpus semantically matches, i.e., whether the quality of the corpus meets the requirements; if If the collected corpus in this domain meets the semantic relevance requirement, then the collected corpus will not be processed again; if If the collected data does not meet the semantic relevance requirements, the collected data will be filtered to remove data that does not match semantically.

[0130] This embodiment also provides word domain analysis, specifically: for each word For each field Words are obtained through word vector models. Basic semantic vector ;

[0131] calculate With domain feature vectors The similarity is used to determine the extent to which the domain influences the semantics of words: ;

[0132] Adjusting the semantic vectors of words based on similarity yields the semantic vectors of the words in the relevant domain. Corrected semantic vector in:

[0133] ;in, It is an adjustment coefficient used to control the degree of influence of domain features on word semantic vectors;

[0134] For the two fields and Calculate words Differences in semantic vectors between these two domains:

[0135] ;

[0136] Define a semantic bias determination function to determine whether a word has a significant semantic bias: ;in, This is a semantic deviation threshold used to determine the semantic bias of words. In the field and Whether there is a significant semantic deviation between them. When the difference between the corrected semantic vectors of words in different domains exceeds a certain threshold, it is considered that there is a significant semantic deviation; if Then determine the word In the field and There are significant semantic discrepancies between them; if Then determine the word In the field and There is no significant semantic difference between them;

[0137] If words In the field and If there are significant semantic discrepancies, then the domain knowledge base should be utilized. Find the words The concepts that are semantically closest, among which, : ;

[0138] Acquire the concept The semantic vector, denoted as ;

[0139] Based on the semantic meaning of concepts in the domain knowledge base, words are... The semantic vectors are adjusted to better reflect the domain. Specific semantics: ;in, It is an adjustment coefficient used to control the magnitude of the adjustment;

[0140] Calculate the adjusted semantic vector domain semantic feature vector Similarity:

[0141] ;

[0142] Define an adjustment judgment function to determine whether the semantic adjustment was successful: ;in, The set similarity threshold is used to judge the adjusted semantic vector. domain semantic feature vector Whether the similarity meets the requirements, that is, whether the semantic adjustment was successful; if If the semantic adjustment is successful, the word is confirmed. In the field The semantic adjustment results in; if If the semantic adjustment fails, the concept is removed, and the domain knowledge base is utilized. Find the word again The concept with the closest semantics Calculate its semantic vector and determine whether the semantic adjustment was successful, until the semantic adjustment is determined to be successful;

[0143] Update words in the field Correction semantic vectors in To reflect the corrected semantic information;

[0144] For each field Utilize authoritative domain knowledge bases to construct a domain concept system:

[0145] ;in, Representation of domain One of the concepts, For the field The number of concepts in the middle;

[0146] For words In the field Correction semantic vectors in Calculate its relationship with the domain Various concepts Similarity: ;in, It is a concept semantic vector;

[0147] Set a concept mapping function to determine whether a word has a corresponding concept in the domain: ;in, This is a domain concept mapping threshold used to determine whether there is a mapping relationship between a word and a domain concept. When the similarity between a word and a domain concept exceeds this threshold, it is considered that there is a mapping relationship between the word and the domain concept.

[0148] like If the word exists, it is determined that there is a corresponding concept in the domain. In the field Corresponding concept ; words and its field The corresponding concept in Record these details to form a mapping relationship between word domain concepts: ;

[0149] like , it is determined that there is no corresponding concept for the word in the domain, and no operation is performed.

[0150] This embodiment also provides an analysis of the special meaning of words, specifically as follows:

[0151] For each domain , based on the keyword annotation in the domain term dictionary and professional literature, define the domain term set: ; where represents a professional term in this domain, is the number of terms;

[0152] Set a term matching judgment function. For each word in the database [[ID=We]] , determine whether it belongs to the term set of the domain : : ; If , it is determined that the word in the domain is a professional term and has a special meaning. Mark the word and the domain , and generate marking information: ; Save the description of the special meaning, which can be extracted from the domain term dictionary or professional literature; if , it is determined that the word in the domain is not a professional term and has no special meaning, and no marking information is generated;

[0153] For each domain , construct a domain text corpus , which contains a large amount of text data related to this domain, such as professional books, academic papers, industry reports, news reports, etc.;

[0154] For the word in the domain text corpus , calculate the frequency of the word in the domain text, and define it as the usage frequency: ;

[0155] Statistically analyze the common collocation patterns of the word B [[ID=To]] text, such as which words it frequently co-occurs with;

[0156] Use the association rule mining method to find the set of words that frequently co-occur with and calculate indicators such as its support and confidence; ;

[0157] Analyze the word ​In the field Contextual features in text, calculated by word The average of the word vectors within the surrounding window yields the context vector. and compare it with the domain feature vector Perform similarity calculations to measure their fit within the domain context:

[0158] ;

[0159] According to words In the field Feature calculation comprehensive score: ;in, , as well as The comprehensive weighting coefficient is used to express the judgment in a special way. To use frequency weights, To match the mode weights, For context weighting, Indicates collocation pattern Confidence level;

[0160] Set up a special expression judgment function to judge words. In the field Does it contain any special expressions?

[0161] ;in, A special expression threshold is used to determine words. In the field Does it contain any special expressions, that is, does it have any special meaning? If so... Then determine the word In the field There are special expressions within memory that have special meanings; these terms and fields are marked. Record tagging information: Descriptions with special meanings can be extracted and summarized based on their typical usage and context in the domain text; if Then determine the word In the field There is no special expression or meaning within it, and no tagging information is generated.

[0162] This embodiment also provides a method for verifying and analyzing marked words, specifically: for each marked word... and its marked areas Define a multi-source data set: ;

[0163] Each of them Represents a data source. The number of data sources, including different countries, regions, official, unofficial, and authoritative sources;

[0164] Based on the reliability, authority, and relevance of the data source, for each data source... Set weights For example, authoritative data sources such as data released by professional academic institutions and official documents have higher weighting, while non-authoritative data sources such as content from certain online forums that have not undergone rigorous review have lower weighting. ;in, This is a reliability weighting coefficient, evaluated based on metrics such as the historical accuracy and error rate of the data source. Reliable data sources typically have a low error rate and high data accuracy. The authority weighting factor is determined based on factors such as the publishing institution, citation frequency, and expert recognition of the data source. Authoritative data sources include data published by professional academic institutions and official government documents. This is a relevance weighting coefficient, which is evaluated based on the fit between the data source and the target domain or application scenario. Data sources with high relevance provide information that is closely related to the target domain.

[0165] Define a set of support relationships: ; among them, each It indicates a supporting relationship, such as "defined as a technical term" or "provides a professional explanation";

[0166] For each data source Define its relation to words In the field The collection of descriptions and explanations within For each data source Determine its description and explanation set. Does there exist at least one element in the set of supporting relationships? Element matching in: ;like Then determine the data source. Clearly indicate the words In the field It has a special meaning; if Then determine the data source. Clearly indicate the words In the field The content does not have any special meaning; if the data source Clearly indicate the words In the field If it has a special meaning, then record it. Otherwise, remember ;

[0167] Calculate words In the field The comprehensive verification score of the special meanings within words reflects the impact of multi-source data on the meanings of words. In the field Support level for special meaning: ;

[0168] Set up a credibility determination function to judge the words In the field Do the special meaning markers within have high credibility? ;in, To comprehensively verify the score threshold, it is used to judge words. In the field Do the special meaning markers within have high credibility? Then the word is considered In the field The special meaning markers within have high credibility, so retain the markers; if Then the word is considered In the field Special meaning markers within do not have high credibility; define an authoritative data source set. Non-authoritative data source collection ;in, and They do not intersect and ;

[0169] Define domain All data sources within the database for words Special meaning support set:

[0170] ;in, Indicates data source Does it support words? In the field It has a special meaning;

[0171] Computational authoritative data source set Supporting words In the field Number of data sources with special meanings: ;

[0172] Define an evaluation decision function to judge words. Whether to retain the marker, i.e., for words In the field Special meaning markers should be re-evaluated, especially when they are found to exist only in some non-authoritative or geographically specific data sources. Care should be taken to determine whether they represent a generally accepted special meaning within the field, and the marker should be removed if necessary. ;in, Indicates that the mark is reserved. This indicates that it is determined manually. Indicates canceling the mark. This is a non-authoritative support threshold used to determine the set of non-authoritative data sources. Supporting words In the field In cases where the data source contains special meaning, An authoritative support threshold is used to determine the set of authoritative data sources. Supporting words In the field In cases where the data source contains special meaning, A collection of non-authoritative data sources Supporting words In the field The number of data sources with special meanings is calculated using the following formula: .

[0173] Example 3 is an improvement on Example 2. In this example, a first-level knowledge graph is generated, specifically by creating a set of word nodes: ; among them, each Represents a word node that is determined to have a special meaning within the domain;

[0174] Create a collection of domain nodes: ; among them, each Represents a domain node with a label;

[0175] For each word node Set its property set: ;in, For word names, This includes basic information about words, such as part of speech and common general meanings. This describes a specific meaning within the domain.

[0176] For each domain node Set its property set: ;in, For the domain name, A brief introduction to the field;

[0177] Create a set of relationships: ; among them, each This represents the semantic relationship between word nodes and domain nodes;

[0178] Define a set of relation types: ;

[0179] Iterate through all words that have a marker. and the fields marked thereon Based on the type of marker, such as technical terms or special expressions, determine the relationship type and denote it as... ; Create a relationship and add it to the relation set. ;

[0180] Combined with word node set Domain Node Set and relation sets This forms the basic structure of a first-level knowledge graph.

[0181] This embodiment also provides the ability to analyze the importance of words within a domain, specifically: for each tag's domain... Obtaining domain-specific text corpora It contains a large amount of text data related to this field, such as professional books, academic papers, news reports, industry reports, etc.

[0182] For each word In the field Text corpus In the text, we can count its frequency of occurrence:

[0183] ;

[0184] Calculate words In the field Keyword weight in: In keyword weighting, For words In the field The word frequencies in the text are as follows: In keyword weighting, For words In the field Inverse document frequency in text, specifically:

[0185] ;

[0186] Statistical terms In the field Number of times it is cited This refers to the number of times a word is cited in relevant literature, reports, etc., within the field; it is used to calculate the number of words. In the field Overall score reflecting the importance of the content: ;in, To use frequency weights, For keyword weight, Citation count weighting For the field The citation count of the most frequently cited words in the text corpus is used to normalize the citation count;

[0187] Domain experts on words In the field The importance of the content was assessed, and expert scores were obtained. Expert rating The range is 0-1; the final importance score is calculated by combining the expert score and the overall score. ;in, The weighting coefficients for expert ratings are set according to the degree of importance attached to the expert opinions;

[0188] Define an importance determination function to determine the importance of words: ;in, This is an importance threshold used to determine the importance of words. In the field The importance of the interior; if Then determine the word In the field The importance level of the content is classified as "high"; if Then determine the word In the field The importance level within is classified as "low".

[0189] This embodiment also provides the generation of a secondary knowledge graph, specifically:

[0190] For each word node that has a tag and its associated domain nodes Based on the importance analysis results, define the node visibility attribute: ;

[0191] Get words In the field The degree of importance within; if words In the field If the importance of the content is classified as "high", then the words will be... In the field Applications within this context are marked as displaying knowledge graph nodes. If words In the field If the importance level is classified as "low", then the words will be... In the field Application scenarios within this scope are marked as hidden knowledge graph nodes. ;

[0192] Define a user set: ;

[0193] For each user Define identity attributes: ;

[0194] like If so, the user's identity is determined to be a professional user;

[0195] like If so, the user's identity is determined to be that of a regular user;

[0196] For each user Define authentication status: ;

[0197] user Submit authentication information This includes, but is not limited to, professional qualification certificates, proof of employment, academic background, etc.

[0198] Define the conditions for approval For example, the authentication information must be complete and the qualifications valid. The authentication information submitted by the user can be verified by the system or manually. Review and approve;

[0199] like If the authentication is successful, it is recorded as follows: ;

[0200] like If authentication fails, it is determined that the authentication was not successful and recorded as follows: ;

[0201] If the user is a professional user and has passed authorization, access to the hidden node is permitted. ;

[0202] If the user is a regular user or has not passed authentication, the hidden node will not be visible to the user.

[0203] The generation of a second-level knowledge graph also includes:

[0204] Get the collection of displayed nodes: ;

[0205] Get the collection of hidden nodes: ;

[0206] The set of visible nodes and the set of hidden nodes will be integrated into the node set of the secondary knowledge graph. middle: ;

[0207] The set of relationships between nodes will be displayed. The set of relationships between hidden nodes Relationship set integrated into the second-level knowledge graph middle: ;

[0208] Based on the integrated node set and relation set Constructing a two-level knowledge graph : ;

[0209] For user query requests According to user identity and node visibility properties ,like Then to the user Display the node and its related relationships;

[0210] like If the user If the user's identity is that of a professional user and they have passed authorization authentication, then the user... Display the node and its related relationships;

[0211] like If the user If the user's identity is a regular user or they have not passed authentication, no notification will be sent to the user. Show hidden nodes and their relationships.

Claims

1. A method for constructing a knowledge graph based on a large model, characterized in that: Comprise: Corpus collection and field division of data in database; For each word in the database, perform word field analysis to determine the meaning of the word in each field; For each word in the database, perform word special meaning analysis to determine whether the word has special meaning in each field; For each word, get its mark information; If the word has mark information, perform test analysis on the marked word; For the word that still has mark information after test analysis, generate a first-level knowledge graph and analyze the importance of the word in the field; According to the importance of the word in the field, generate a second-level knowledge graph; If the word does not have mark information, do not perform any operation on the word; Word special meaning analysis, specifically: For each domain define a set of domain terms: ； wherein denotes a term of the art, is the number of terms; A term matching judgment function is set up to judge whether each word in the database belongs to the domain's term set. :​​ ； If , determine the term In the field , there is a special meaning of the term, mark the term and field , generate the marking information: ； Save the description of the special meaning; If , then determine the word In the field It is not a professional term, there is no special meaning, no marker information is generated; For each domain , a domain text corpus is constructed ; For a word In a corpus of text in a domain , the frequency of occurrence of the word in the text of the domain is defined as usage frequency: ； statistical terms in the art common collocation patterns in text; Finding out the frequent co-occurrence word set of the frequency of the co-occurrence of the word set and calculating its support and confidence indicators; analyzing the words in the field contextual features in the text: ； According to the words In the field of feature computing integrated score: ； wherein, , and is a comprehensive weight coefficient for special expression judgment, is a frequency weight, is a collocation pattern weight, is a context weight, represents a confidence degree of the collocation pattern . A special expression judging function is set to judge the word In the field Whether there is a special expression: ； wherein is a special expression decision threshold value; If , then determine the word In the field There is a special expression, with special meaning, mark the word and field , record the mark information: ； Save the description of the special meaning; If , then the term is determined to be not special, not to have a special meaning, and no flag information is generated. Test analysis on the marked word, specifically: For each word that is tagged and its tagged domain define a multi-source data set: ； wherein each represents a data source, is the number of data sources; According to the reliability, authority and relevance of the data sources, weights are set for each data source Setting weights : ； wherein, is a reliability weight coefficient, is an authority weight coefficient, is a relevance weight coefficient; Define a support relationship set: ； wherein each represents a support relationship; For each data source , a set of descriptions and explanations of its relation to terms in the domain ; For each data source , determine if at least one element of its description and explanation set matches an element of the support relationship set : ； If , then determine the data source explicitly indicate the word has a special meaning within the field ; If , then determine the data source explicitly indicate the word in the field has no special meaning; If the data source explicitly indicates that a word has a special meaning in the field , then record ; Otherwise, record ; Computing terms In the field Synthetic validation score for special meaning within the field: ； A credibility decision function is set to determine whether a word In the field of special meaning markers has higher credibility: ； wherein, is a comprehensive verification score threshold value; If , the term is considered to have a high degree of confidence in the special meaning in the field , and the tag is retained. If , the term is considered to have a special meaning in the field , the set of definition-authoritative data sources , the set of non-authoritative data sources ; wherein and are disjoint and ; Definition field In all data sources for the term Support set for the special meaning of the term: ； wherein, represents a data source whether the term is supported in the field has a special meaning; Set of computing authority data sources In-domain support terms In-domain Number of data sources with special meaning within domain: ； A function to evaluate the decision function is defined to determine whether the word is marked for retention: ； wherein, denotes a reserved marker, denotes a manual judgment, denotes a cancel marker, is a non-authoritative support threshold, is an authoritative support threshold, is a set of non-authoritative data sources supports the term in the domain has a special meaning, and the number of data sources is given by the formula: ； Generate a first-level knowledge graph, specifically: Create a word node set: ； wherein each represents a word node determined to have a special meaning within the domain; Create a field node set: ； wherein each represents a tagged domain node; For each word node , set its property set: ； wherein, is a word name, is basic information of the word, is a description of a special meaning within the field; For each domain node , a set of its properties is set: ； wherein is a field name, is a field brief; Create a relationship set: ； wherein each represents a semantic association relationship between a term node and a domain node; Define a relationship type set: ； traversing all the marked terms and their marked fields , depending on the type of marking, determining the type of relationship, denoted by ; Create relationship and add it to the relationship collection ; Combined with word node set Domain Node Set and relation sets This forms the basic structure of a first-level knowledge graph; Generate a second-level knowledge graph, specifically: For each word node that is tagged as present and its associated domain node define the node visibility property according to the importance analysis result: ； acquiring terms importance within a field in the field If the importance classification of the word in the field is "high", the word is marked as a knowledge graph node In the field In the field In the field In the field ; If the importance classification of the word in the field is "low", the word is hidden in the knowledge graph In the field In the field In the field The application scenario of the word in the field is marked as a hidden knowledge graph node ; Define a user set: ； For each user Define identity attributes: ； If then the identity of the user is determined to be a professional user; If then the identity of the user is determined to be a normal user; For each user Define authentication status: ； User Submit authentication information ; Defining the condition for audit pass ; If , then it is determined that the permission authentication is passed, denoted as ; If , then determine that the permission authentication fails, and record as ; If the user's identity is a professional user and passes the permission authentication, access to the hidden node is allowed ; Generating a second-level knowledge graph also includes: Get the display node set: ； Get the hidden node set: ； Integrating the set of displayed nodes and the set of hidden nodes into the set of nodes of the secondary knowledge graph In: ； a set of relationships between the display nodes a set of relationships between the hidden nodes a set of relationships integrated into the secondary knowledge graph in: ； According to the integrated node set and the relationship set , a secondary knowledge graph is constructed : ； For a user query request , according to the user identity and the node visibility attribute , if , the node and its related relationships are displayed to the user ; If , if the identity of the user is a professional user and is authenticated by the permission, the node and its related relationship are shown to the user ; If , if the identity of the user is an ordinary user or has not passed the permission authentication, the hidden nodes and their relationships are not shown to the user .

2. The large model-based knowledge graph construction method of claim 1, wherein: Corpus collection and field division, specifically: S1, define a field set: ； wherein each represents one field, is the total number of fields; S2, for each domain , preliminarily define the semantic range of the domain, denoted by ; S3, by extracting the word vectors of the keywords and phrases in the semantic scope and performing an average pooling operation to calculate an initial feature vector of the semantic scope : ； wherein, is a word vector for the word , is a number of elements in a semantic range . S4, collect domain sample data, denoted as ; S5、By statistics TF and inverse document frequency (IDF) of the higher frequency of words, extract feature dimension candidate set : ； wherein, is a set threshold value, is a set of processed texts obtained after pre-processing the sample data is pre-processed. S6. For each domain , build a domain knowledge base ; Utilizing a domain knowledge base , defining a concept set of a domain : ； wherein each represent a key concept or term in the field ; From the set of feature dimensions highly relevant to the field of art is selected: ； wherein, is a set of selected feature dimensions, is a candidate set of feature dimensions, is a domain of concepts, denotes a similarity of a feature dimension to a domain concept, denotes a similarity of a feature dimension to a domain concept, is a set similarity threshold value; S7、Through principal component analysis or other feature selection algorithm, the weight of feature dimension is calculated to determine the importance weight of each feature dimension in the field ​ ； S8. Constructing the field Final feature vector: ； in, The number of feature dimensions, It is the first Each feature dimension It is its corresponding weight; S9. Updating the semantic range of the feature vector , updating the field of semantics: ； wherein, is a set weight threshold value, and the feature dimension with high weight is included in the semantic range. S10, set an update determination function to determine whether the feature vector and semantic range are appropriate: ； wherein, denotes the set difference, denotes the set of elements belonging to the set but not to the set , and is a set convergence threshold, is the number of iterations; If , then determine that the eigenvector and semantic range tend to be stable, stop repeating, and get the semantic range and eigenvector of each field , respectively, denoted as and ; If then it is determined that the feature vector and semantic range are not converging and steps S5-S10 are repeated. S11. For each domain Collecting text corpus from multiple channels to form a corpus collection: ； wherein, represents a text segment, the amount of corpus collected for the field; S12, define a corpus collection diversity index: ； wherein, is a set of corpus source types, is a corpus source type is a weight, is a source type is a number of corpora in a domain . S13, set a corpus collection determination function to determine whether the corpus is sufficient: ； wherein, a target diversity threshold for the corpus collection; If then it is determined that the corpus collection for this domain satisfies the diversity requirement and there is no need to continue collecting the corpus for this domain; If then it is determined that the corpus collection for the field does not satisfy the diversity requirement and the collection of the corpus for the field needs to continue; S14, counting each field Semantic relevance indicators of the collected corpus: ； wherein, representing a text segment with the field characteristic vector similarity; S15, for each field , define a corpus quality judgment function to judge whether the quality of the corpus meets the requirements: ； wherein, is a corpus semantic relevance threshold; If , it is determined that the corpus collection of the field meets the semantic correlation requirement, and no further processing is performed on the collected corpus; If then it is determined that the corpus collection in this field does not meet the semantic relevance requirement, and the collected corpus is filtered to remove the corpus that does not match the semantics. 3.The large model-based knowledge graph construction method of claim 1, wherein: Word field analysis, specifically: For each word For each field Words are obtained through word vector models. Basic semantic vector ; Computing the similarity of the field feature vector to the field feature vector ； According to the similarity, the semantic vector of the word is adjusted to obtain a corrected semantic vector of the word in the domain : ； wherein is an adjustment factor; For both domains and the semantic vector difference of the words in these two domains is computed: ； Define a semantic deviation determination function to determine whether the word has significant semantic deviation: ； wherein, is a semantic bias threshold; If , then the term is determined to be semantically biased between the domain and ; If , then the term is determined to be semantically neutral between the domain and ; If the word In the field And There is a significant semantic deviation, then use the domain knowledge base , find the concept closest to the semantics of the word , where, : ； acquire the semantic vector of the concept , denoted as ; According to the concept semantics in the domain knowledge base, the semantic vector of the word is adjusted to be closer to the specific semantics of the domain : ； wherein is an adjustment factor; computing an adjusted semantic vector similarity to a domain semantic feature vector ​ ； Set an adjustment determination function to determine whether the semantic adjustment is successful: ； wherein, is a set similarity threshold value; If , then determine that the semantic adjustment is successful, and confirm the word in the domain ; semantic adjustment result; like If the semantic adjustment fails, the concept is removed, and the domain knowledge base is utilized. Find the words again The concept with the closest semantics Calculate its semantic vector and determine whether the semantic adjustment was successful, until the semantic adjustment is determined to be successful; Updating a correction semantic vector for a term in a domain ;​ For each domain , build a domain concept hierarchy with an authoritative domain knowledge base: ； wherein, represents a concept in the field , is the number of concepts in the field . For a word In a domain Correct semantic vector , compute its similarity to each concept in the domain ​ ； wherein, is a semantic vector of the concept ; Set a concept corresponding determination function to determine whether the word has a corresponding concept in the field: ； wherein, is a domain concept mapping threshold; If , then determine that the word has a corresponding concept in the domain, the word has a corresponding concept in the domain ; and ; The word and its corresponding concept in the field are recorded, forming a word-field concept mapping relationship: ； If then the term is determined not to have a corresponding concept in the domain and no operation is performed. 4.The large model-based knowledge graph construction method of claim 1, wherein: Analyze the importance of the word in the field, specifically: for each tagged domain , obtaining a domain text corpus ; For each word in the text corpus in the field count its frequency of occurrence: ； Computing terms In the field of key words weighting: ； In the keyword weight, For the word In the field The word frequency in the text, specifically: ； In the keyword weight, For the word In the field Inverse document frequency in the text, specifically: ； statistical terms in the art number of times cited ; Computing term Importance within field Overall importance score: ； wherein, is a frequency weight, is a keyword weight, is a citation count weight, is a field citation count of the most cited word in the text corpus; Domain expert assesses the importance of the word In the domain and gets an expert score ; Combine expert scores and comprehensive scores to calculate the final importance score: ； wherein, is the weight coefficient for the expert score; Define an important determination function to determine the importance of the word: ； wherein, is an importance threshold value; like Then determine the word In the field The importance level is classified as "high"; If , then the term is classified as "low" in importance within the field . ​

Images