Intelligent knowledge graph construction method based on AI

Abstract

The invention belongs to the technical field of knowledge graph construction, and particularly relates to an AI-based intelligent knowledge graph construction method, which comprises the following steps: acquiring a target subject domain, and collecting knowledge data from a data source according to the target subject domain; performing data cleaning on the collected knowledge data to obtain cleaned knowledge data; performing data extraction on the cleaned knowledge data to obtain knowledge extraction data; extracting polysemous words in the knowledge extraction data, and adopting a multi-scoring mechanism for the polysemous words to obtain polysemous word meanings; performing knowledge representation and storage according to knowledge extraction data and polysemous word meanings; integrity evaluation is carried out on the constructed knowledge graph to obtain an overall integrity index, and the overall integrity index is used for measuring the integrity of the knowledge graph; and checking and optimizing the constructed knowledge graph according to the overall integrity index. According to the method, the situation that ambiguity exists in the paragraph due to the fact that the meaning of the polysemy word is uncertain is effectively prevented, and the accuracy of the knowledge graph is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of knowledge graph construction, and in particular to an AI-based intelligent knowledge graph construction method. Background Art

[0002] Knowledge graph construction refers to the process of integrating scattered and heterogeneous data sources, automatically identifying and extracting entities, relationships and attributes through natural language processing and machine learning technologies to form a structured semantic knowledge base. This knowledge base is represented in the form of a graph, consisting of nodes (representing entities) and edges (representing relationships between entities), which can provide rich background knowledge for machine learning models and enhance their understanding and reasoning capabilities. Knowledge graphs are widely used in search engines, recommendation systems, intelligent question and answer, education industry and other fields. Today, knowledge graphs have become a key technology for organizing and understanding massive amounts of information, and how to derive more accurate knowledge graphs is now widely concerned.

[0003] In the prior art, CN114328951B discloses a knowledge graph construction method that integrates information acquisition and triple extraction, including the following steps: S1: using crawler technology to regularly crawl ocean-related text content including news from specified web pages; S2: using natural language processing tools to perform entity extraction and relationship extraction on the text content to obtain triples of news, and then storing the triples of news in a database; S3: constructing a knowledge graph based on the triples in the database, and realizing the visualization of the knowledge graph in a data browser; S4: obtaining knowledge associations based on the visualized knowledge graph.

[0004] CN115017335A discloses a knowledge graph construction method and system, including the following steps: setting a new word discovery algorithm, and organizing special nouns and an open knowledge graph as a database for word segmentation recognition; obtaining an input text, and extracting triples including a subject, a predicate, and an object in the text as knowledge extraction results based on a database and a word segmentation extractor; querying the open knowledge graph for triples related to the nodes corresponding to the extracted triples, and forming a new knowledge graph with all triples.

[0005] However, there are still some problems with the prior art and the above-mentioned existing patents, mainly in that many words and phrases have multiple meanings. The ambiguity caused by multiple meanings can lead to inaccurate entity recognition and relationship extraction, which in turn affects the quality and usability of the knowledge graph, and affects users' trust in the data and the performance of applications based on the knowledge graph. Summary of the invention

[0006] In view of the above deficiencies in the prior art, the purpose of the present invention is to provide an AI-based intelligent knowledge graph construction method, which effectively prevents the ambiguity of paragraphs caused by the uncertain meaning of polysemous words and enhances the accuracy of the knowledge graph.

[0007] In order to achieve the above purpose, the technical solution adopted by the present invention is a method for constructing an intelligent knowledge graph based on AI, comprising the following steps: Obtain the target subject domain and collect knowledge data from the data source according to the target subject domain; Perform data cleaning on the collected knowledge data to obtain cleaned knowledge data; Perform data extraction on the cleaned knowledge data to obtain knowledge extraction data; Extract polysemous words from knowledge extraction data, use a multiple scoring mechanism to determine the meaning of polysemous words, and derive the meaning of polysemous words; Extract data and polysemous meanings based on knowledge, and perform knowledge representation and storage; Conduct integrity assessment on the constructed knowledge graph and obtain an overall integrity index, which is used to measure the integrity of the knowledge graph. The constructed knowledge graph is verified and optimized according to the overall completeness index.

[0008] Preferably, the step of collecting knowledge data from a data source according to a target subject domain comprises the following steps: Obtain the target subject domain covered by the knowledge graph; Connect to the data source and identify the data source type; Determine the data collection tool according to the data source type, and use data mining algorithms and text analysis techniques to collect and extract knowledge data based on the target subject domain.

[0009] Preferably, the data cleaning of the collected knowledge data to obtain the cleaned knowledge data, and data extraction of the cleaned knowledge data to obtain the knowledge extraction data include the following steps: Remove duplicate records through unique identifiers and unify data in different formats into a standard format; By sorting out different data sources, merging relevant data, and removing redundant data, we can obtain cleaned knowledge data; Based on the cleaned knowledge data, entity extraction, relationship extraction and attribute extraction are performed through data extraction rules to obtain knowledge extraction data.

[0010] Preferably, the method of extracting polysemous words from the knowledge extraction data adopts a multiple scoring mechanism for the polysemous words, determines the meaning of the polysemous words through the multiple scoring mechanism, and obtains the meaning of the polysemous words, including the following steps: Based on the knowledge extraction of polysemous words in the data, the meaning 1, meaning 2, ..., meaning N of the polysemous words are extracted, where N is the meaning number; The meaning 1, meaning 2, ..., meaning N of the polysemous word are scored through a multiple scoring mechanism, and the meaning of the polysemous word is obtained based on the scoring results.

[0011] Preferably, the multiple scoring mechanism comprises the following steps: Get the meaning 1, meaning 2, ..., meaning N of a polysemous word; Substitute meaning 1, meaning 2, ..., meaning N into the original paragraph where the polysemous word is located one by one, and obtain paragraph 1, paragraph 2, ..., paragraph G, where G is the paragraph number; perform context consistency scoring on the original paragraphs and paragraph 1, paragraph 2, ..., paragraph G in turn, and obtain a context consistency score sequence, which includes context consistency score 1, context consistency score 2, ..., context consistency score F, where F is the context consistency score number; Count the number of identical highest scores in a sequence of context consistency scores: If the highest scores in the context consistency score sequence are the same, the meaning of the polysemous word corresponding to the highest score is the adapted meaning of the polysemous word in its original paragraph; If the number of the same highest scores in the context consistency score sequence is not unique, then the commonness score is performed on the paragraphs with the highest context consistency score to obtain a commonness score sequence. The commonness score includes commonness score 1, commonness score 2, ..., commonness score M, where M is the number of the same highest scores in the context consistency score sequence; If there are only a few words with the same highest score in the commonness score sequence, the meaning of the polysemous word corresponding to the highest score is the adaptive meaning of the polysemous word in its original paragraph; If the number of paragraphs with the same highest score in the commonness score sequence is not unique, then the paragraphs with the highest commonness score are scored for sentiment consistency to obtain a sentiment consistency score sequence, which includes sentiment consistency score 1, sentiment consistency score 2, ..., sentiment consistency score P, where P is the number of paragraphs with the same highest score in the commonness score sequence; If the number of highest scores in the sentiment consistency score sequence is the same, the meaning of the polysemous word corresponding to the highest score is the adapted meaning of the polysemous word in its original paragraph. Otherwise, the management personnel are notified to manually adapt the meaning of the polysemous word.

[0012] Preferably, the construction of a knowledge graph based on knowledge extraction data and polysemous word meanings comprises the following steps: Through string matching and semantic similarity calculation, similar entities are identified and aligned, and the attribute values ​​of knowledge extraction data are filled using the weighted average method; The knowledge extraction data and the meaning of polysemous words are structured, and the structured knowledge extraction data and the meaning of polysemous words are converted into triples to obtain nodes and edges of the knowledge graph; The nodes and edges of the obtained knowledge graph are stored in the graph database to construct the knowledge graph.

[0013] Preferably, the completeness evaluation of the constructed knowledge graph to obtain the overall completeness index comprises the following steps: Obtain integrity evaluation indicators, perform quantitative analysis on nodes, edges, and attributes in the knowledge graph based on the obtained integrity evaluation indicators, and calculate the actual value of each indicator; Based on the actual value of the integrity assessment indicator, the overall integrity index of the knowledge graph is obtained through the overall integrity index calculation formula.

[0014] Preferably, the checking and optimizing of the constructed knowledge graph according to the overall integrity index comprises the following steps: Based on the obtained overall completeness index of the knowledge graph, the overall completeness index of the knowledge graph is compared with the minimum expected threshold of completeness; If the overall integrity index of the knowledge graph is not less than the minimum expected threshold of completeness, it is determined that the completeness of the knowledge graph has reached the expected level; If the overall integrity index of the knowledge graph is less than the minimum expected threshold for completeness, it is determined that the integrity of the knowledge graph does not reach the expected level, the knowledge graph is output as not meeting the integrity standard, and the administrator is notified to verify and optimize the knowledge graph.

[0015] Preferably, the steps for obtaining the overall integrity index are: The completeness evaluation indicators include the difference in the number of entities, the difference in the number of relationships, the difference in the number of attributes, the total number of retrievals, and the number of unsuccessful comparisons; Get a dictionary of word meanings; Through the deep learning model that is different from the data extraction rules, entity extraction, relationship extraction and attribute extraction are performed again from the knowledge extraction data; Combine the word meaning dictionary to construct a comparison vocabulary of the newly extracted entities, relations or attributes; Count and compare the number of entities, the number of relationships, and the number of attributes in the vocabulary and knowledge graph respectively, and compare them with each other to obtain the difference in the number of entities, the difference in the number of relationships, and the difference in the number of attributes; Randomly extract an entity, relationship or attribute from the comparison vocabulary, put it into the knowledge graph for comparison and search, and count the total number of searches and the number of unsuccessful comparisons; The overall integrity index is calculated using the overall integrity index calculation formula by combining the difference in the number of entities, the difference in the number of relationships, the difference in the number of attributes, the total number of searches and the number of unsuccessful comparisons.

[0016] Preferably, the overall integrity index calculation formula is: ; ; Where β is the overall integrity index, is the difference in the number of entities, is the difference in the number of relations, is the difference in the number of attributes, Y is the total number of searches, X is the number of unsuccessful comparisons, e is a natural constant, α1 is the weight of the difference in the number of entities for the overall completeness index, α2 is the weight of the difference in the number of relationships for the overall completeness index, α3 is the weight of the difference in the number of attributes for the overall completeness index, α4 is the weight of the ratio of the number of unsuccessful comparisons to the total number of searches for the overall completeness index, and f is the transition function.

[0017] The present invention has the following advantages: The present invention substitutes several meanings contained in polysemous words into the original paragraph respectively, and then scores the contextual consistency of the new paragraph with the substituted meanings and the original paragraph, and effectively reduces the ambiguity caused by polysemous words by evaluating the coherence and logical consistency of the paragraph. By substituting the meaning of the polysemous words into the original paragraph for commonality scoring, the meanings with the same contextual consistency score can be scored again, which effectively improves the efficiency of eliminating ambiguity. By substituting the meaning of the polysemous words into the original paragraph for sentiment consistency scoring, the meanings with the same commonality score can be scored three times, which improves the accuracy of polysemous word meaning adaptation and enhances the accuracy of the knowledge graph.

[0018] The present invention structures the knowledge extraction data and the meaning of polysemous words and converts them into triples to obtain the nodes and edges of the knowledge graph, thereby making the knowledge graph easy to store and query, facilitating subsequent knowledge reasoning, and enabling the knowledge graph to support complex queries and intelligent applications, thereby enhancing the queryability and reasonability of the knowledge graph and improving the practicality and flexibility of the knowledge graph.

[0019] The present invention performs integrity assessment on the constructed knowledge graph to obtain an overall integrity index, and verifies and optimizes the knowledge graph based on the overall integrity index, thereby ensuring that the knowledge graph can continuously reflect the latest domain knowledge. Through automated evaluation and optimization mechanisms, the maintenance efficiency and update speed of the knowledge graph are improved, thereby ensuring the timeliness and reliability of the knowledge graph.

[0020] The knowledge graph constructed by the present invention can be applied in the large-scale model dedicated to education and teaching. By combining big data, machine learning, knowledge graph and other technologies, an intelligent science and education environment can be built to create a teaching and scientific research service platform with professionalism, personalization, systematicness and one-stop characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 Flowchart of the AI-based intelligent knowledge graph construction method provided by the present invention. DETAILED DESCRIPTION

[0022] The present invention will be further described below in conjunction with the accompanying drawings.

[0023] The general idea in the present invention is as follows: By substituting the several meanings contained in the polysemous words into the original paragraph respectively, and then scoring the contextual consistency of the new paragraph with the substituted meanings and the original paragraph, by evaluating the coherence and logical consistency of the paragraph, and by substituting the meanings of the polysemous words into the original paragraph for commonality scoring, the meanings with the same contextual consistency score can be scored again, and by substituting the meanings of the polysemous words into the original paragraph for sentiment consistency scoring, the meanings with the same commonality score can be scored three times, thereby achieving the effect of quickly reducing the ambiguity caused by polysemous words and improving the accuracy of the meaning adaptation of polysemous words.

[0024] like Figure 1 As shown in the figure, the AI-based intelligent knowledge graph construction method includes the following steps: Obtain the target subject domain and collect knowledge data from the data source according to the target subject domain; Perform data cleaning on the collected knowledge data to obtain cleaned knowledge data; Perform data extraction on the cleaned knowledge data to obtain knowledge extraction data; Extract polysemous words from knowledge extraction data, use a multiple scoring mechanism to determine the meaning of polysemous words, and derive the meaning of polysemous words; Extract data and polysemous meanings based on knowledge, and perform knowledge representation and storage; Conduct integrity assessment on the constructed knowledge graph and obtain an overall integrity index, which is used to measure the integrity of the knowledge graph. The constructed knowledge graph is verified and optimized according to the overall completeness index.

[0025] In this embodiment, during the data collection phase, a large amount of information is obtained from various data sources, including structured basic information such as colleges, majors, courses, and textbooks from school databases, academic affairs systems, etc., and structured or semi-structured information such as directory structure, basic theories, main viewpoints, and unstructured information such as innovation trends and social practices are extracted from professional textbooks, course outlines, academic papers, research reports, scientific research data, internship training cases and other text resources through a large model (natural language processing technology).

[0026] The IQR (interquartile range) technology is used to remove outliers from the collected data, and the principal component analysis (PCA) technology is used to remove noise from the collected data. Standardized naming of keywords such as disciplines, departments, and majors can ensure data consistency, improve data accuracy, and facilitate subsequent knowledge extraction. Machine learning algorithms, such as conditional random fields (CRF) and long short-term memory networks (LSTM), are used to automatically label and classify knowledge extraction information, extract required information, and extract knowledge elements such as entities, relationships, and attributes, and convert the extracted knowledge elements into triples to initially form a knowledge graph.

[0027] By converting unstructured data into structured knowledge representation, the readability of the data is enhanced. When encountering polysemous words, a multiple scoring mechanism is used to determine their meaning in a specific context, which effectively solves the ambiguity problem in natural language processing and improves the accuracy and usability of the knowledge graph. After the extraction is completed, the extracted knowledge elements are stored in a structured form to facilitate subsequent query and reasoning, thereby improving the application value and practicality of knowledge.

[0028] Integrity assessment helps identify missing parts in the knowledge graph by measuring the completeness of the knowledge graph, while verification optimization further improves the accuracy and reliability of the knowledge graph. It not only improves the quality of the knowledge graph, but also ensures that the knowledge graph can continue to provide support for various intelligent applications, enhancing the practicality and effectiveness of the knowledge graph.

[0029] Collecting knowledge data from data sources according to the target subject domain includes the following steps: obtaining the target subject domain covered by the knowledge graph; connecting the data source and identifying the data source type; determining the data collection tool according to the data source type, and collecting and extracting knowledge data based on the target subject domain using data mining algorithms and text analysis techniques.

[0030] In this embodiment, combined with the expected application scenario, the target subject domain is clearly defined to collect knowledge data from the data source to ensure that the constructed knowledge graph can meet actual needs. The data source is connected through NLP technology, and the type of the connected data source is identified through a deep learning model. The structured data in the data source is extracted by using ETL tools, the semi-structured data in the data source is extracted by using JSON parsing tools, and the unstructured data in the data source is extracted by API calls. Choosing appropriate tools can improve the efficiency and accuracy of data collection. Finally, based on the target subject domain, data mining algorithms and text analysis techniques are used to collect and extract knowledge data, thereby ensuring that high-quality knowledge data related to the target subject domain is extracted from the data source.

[0031] The collected knowledge data is cleaned to obtain the cleaned knowledge data, and the cleaned knowledge data is extracted to obtain the knowledge extraction data, including the following steps: removing duplicate records through unique identifiers and unifying data in different formats into a standard format; obtaining the cleaned knowledge data by collating different data sources, merging related data, and removing redundant data; based on the cleaned knowledge data, entity extraction, relationship extraction, and attribute extraction are performed through data extraction rules to obtain knowledge extraction data.

[0032] In this embodiment, the data extraction rules include rule-based extraction and statistics-based extraction. By removing duplicate records through unique identification and unifying data in different formats into a standard format, duplicate data can be avoided in the data, the quality and accuracy of the data can be improved, and keywords such as disciplines, departments, and majors can be standardized and named, and data from different sources can be integrated into a unified data set to ensure data consistency and improve the quality of data collection.

[0033] Among them, the statistical-based extraction method: uses machine learning algorithms, such as conditional random fields (CRF), long short-term memory networks (LSTM), etc., to extract the required information from the cleaned knowledge data. The extraction process includes entity extraction, relationship extraction and attribute extraction. The entity extraction process is to automatically discover entities such as disciplines, departments, majors, authors, publishers, etc. from the cleaned knowledge data through named entity recognition (NER) technology. By training the model through machine learning methods, it can learn how to identify and classify the relationship between entities from a large amount of annotated data, so as to extract the relationship between entities; for knowledge points, through in-depth semantic understanding and analysis, specific attributes such as directory structure, basic theory, and main viewpoints are extracted from the cleaned knowledge data to complete attribute extraction and obtain knowledge extraction data.

[0034] Extract polysemous words in knowledge extraction data, adopt a multiple scoring mechanism for the polysemous words, determine the meaning of the polysemous words through the multiple scoring mechanism, and obtain the meaning of the polysemous words, including the following steps: based on the polysemous words in the knowledge extraction data, extract the meaning 1, meaning 2, ..., meaning N of the polysemous words, where N is the meaning number; score the meaning 1, meaning 2, ..., meaning N of the polysemous words through the multiple scoring mechanism, and obtain the meaning of the polysemous words based on the scoring results.

[0035] The multiple scoring mechanism includes the following steps: obtaining the meaning 1, meaning 2, ..., meaning N of the polysemous word; substituting the meaning 1, meaning 2, ..., meaning N into the original paragraph where the polysemous word is located one by one to obtain paragraph 1, paragraph 2, ..., paragraph G, where G is the paragraph number; and scoring the original paragraph with paragraph 1, paragraph 2, ..., paragraph G in turn for contextual consistency. The contextual consistency scoring method is: evaluating the consistency of the paragraph structure through dependency syntactic analysis, and then calculating the correctness, completeness and logic of the dependency relationship to score, and obtaining a contextual consistency scoring sequence. The contextual consistency scoring sequence includes contextual consistency score 1, contextual consistency score 2, ..., contextual consistency score F, where F is the contextual consistency scoring number; counting the number of the same highest scores in the contextual consistency scoring sequence: if the number of the same highest scores in the contextual consistency scoring sequence is unique, then the polysemous word meaning corresponding to the highest score is the adaptive meaning of the polysemous word in its original paragraph, which has the effect of selecting the adaptive meaning of the polysemous word through one scoring, saving working time.

[0036] If the number of the same highest scores in the context consistency score sequence is not unique, then the paragraph with the highest context consistency score is scored for commonality. The commonality scoring method is as follows: based on the statistical data of the N-gram model, the commonality of a meaning is determined. The score is scored by calculating the frequency of occurrence of a specific meaning in the corpus. The higher the frequency, the higher the commonality score. A commonality score sequence is obtained. The commonality score includes commonality score 1, commonality score 2, ..., commonality score M, where M is the number of the same highest scores in the context consistency score sequence. If the number of the same highest scores in the commonality score sequence is unique, the meaning of the polysemous word corresponding to the highest score is the adapted meaning of the polysemous word in its original paragraph. By scoring the paragraph twice, the situation where a single context consistency score may not be sufficient to distinguish polysemous words with similar meanings is prevented, thereby enhancing the robustness of the system and improving the user experience.

[0037] If the number of paragraphs with the same highest score in the commonality score sequence is not unique, then the paragraph with the highest commonality score is scored for sentiment consistency. The sentiment consistency scoring method is: by using sentiment analysis tools, such as the Snownlp library, to perform sentiment polarity analysis on the paragraph, output the probability of positive or negative sentiment, and then score by calculating the sentiment polarity probability output by the Snownlp library. The more consistent the sentiment tendency is with the original paragraph, the higher the sentiment consistency score is, and a sentiment consistency score sequence is obtained. The sentiment consistency score sequence includes sentiment consistency score 1, sentiment consistency score 2, ..., sentiment consistency score P, where P is the number of paragraphs with the same highest score in the commonality score sequence; if the number of paragraphs with the same highest score in the sentiment consistency score sequence is unique, the meaning of the polysemous word corresponding to the highest score is the adapted meaning of the polysemous word in its original paragraph, otherwise, the management personnel are notified to manually adapt the meaning of the polysemous word. The multiple scoring mechanism helps to provide more accurate results and reduce users' doubts about the results and the need for manual corrections.

[0038] Based on the knowledge extraction data and the meaning of polysemous words, a knowledge graph is constructed, which includes the following steps: similar entities are identified through string matching and semantic similarity calculation. Machine learning algorithms such as cosine similarity or Jaccard similarity can be used to quantify the similarity between entities to ensure the accuracy of entity recognition. Entity alignment is then performed, and the attribute values ​​of the knowledge extraction data are filled using the weighted average method.

[0039] The knowledge extraction data and polysemous meanings are structured through natural language processing (NLP) tools, and the structured knowledge extraction data and polysemous meanings are converted into triples to obtain the nodes and edges of the knowledge graph, which can clearly represent the relationship and attributes between entities; the nodes and edges of the obtained knowledge graph are stored in a graph database to construct a knowledge graph.

[0040] Conducting integrity assessment on the constructed knowledge graph and obtaining an overall integrity index includes the following steps: obtaining integrity assessment indicators, conducting quantitative analysis on the nodes, edges, and attributes in the knowledge graph based on the obtained integrity assessment indicators, and calculating the actual value of each indicator; obtaining the overall integrity index of the knowledge graph through the overall integrity index calculation formula based on the actual value of the integrity assessment indicator, which helps to improve the reliability and availability of the knowledge graph and ensure that the information in the knowledge graph is accurate and reliable.

[0041] The constructed knowledge graph is checked and optimized according to the overall integrity index, including the following steps: based on the overall integrity index of the obtained knowledge graph, the overall integrity index of the knowledge graph is compared with the minimum expected threshold of completeness; if the overall integrity index of the knowledge graph is not less than the minimum expected threshold of completeness, it is determined that the completeness of the knowledge graph has reached the expected level; If the overall integrity index of the knowledge graph is less than the minimum expected threshold for completeness, it is determined that the integrity of the knowledge graph has not reached the expected level, and the output of the knowledge graph does not meet the integrity standard. The administrator is notified to verify and optimize the knowledge graph. When new knowledge is added, it can be ensured that the newly added knowledge is seamlessly integrated with the existing knowledge graph to maintain overall consistency and integrity.

[0042] The steps for obtaining the overall completeness index are as follows: the completeness evaluation indicators include the difference in the number of entities, the difference in the number of relationships, the difference in the number of attributes, the total number of searches and the number of unsuccessful comparisons; obtain a dictionary of word meanings; use a deep learning model that is different from the data extraction rules to extract entities, relationships and attributes from the knowledge extraction data again; combine the word meaning dictionary to construct a comparison vocabulary with the newly extracted entities, relationships or attributes; count the number of entities, the number of relationships and the number of attributes in the comparison vocabulary and knowledge graph respectively, and compare them with each other to obtain the difference in the number of entities, the difference in the number of relationships and the difference in the number of attributes; randomly select an entity, relationship or attribute from the comparison vocabulary, put it in the knowledge graph for comparison and search, and count the total number of searches and the number of unsuccessful comparisons; combine the difference in the number of entities, the difference in the number of relationships, the difference in the number of attributes, the total number of searches and the number of unsuccessful comparisons, and calculate the overall completeness index through the overall completeness index calculation formula.

[0043] In this embodiment, the deep learning model specifically refers to: pre-training the BERT pre-trained language representation model on a large amount of unlabeled text data and unsupervised pre-training tasks, so that the BERT pre-trained language representation model has richer language representation capabilities and better context understanding capabilities, and through the BERT pre-trained language representation model, entity extraction, relationship extraction and attribute extraction are performed from the knowledge extraction data, and the newly extracted entities, relationships or attributes are constructed into a comparison vocabulary in combination with the word meaning dictionary, which can more accurately identify the key information in the text and improve the extraction coverage.

[0044] The overall integrity index calculation formula is: ; ; Where β is the overall integrity index, is the difference in the number of entities, is the difference in the number of relations, is the difference in the number of attributes, Y is the total number of searches, X is the number of unsuccessful comparisons, e is a natural constant, α1 is the weight of the difference in the number of entities for the overall completeness index, α2 is the weight of the difference in the number of relationships for the overall completeness index, α3 is the weight of the difference in the number of attributes for the overall completeness index, α4 is the weight of the ratio of the number of unsuccessful comparisons to the total number of searches for the overall completeness index, and f is the transition function.

[0045] An example calculation is when =1, α1=0.2, =1, α2=0.2, =1, α3=0.2, X=0, Y=28, α4=0.4, we can use the above formula to calculate β=f=0.8019. The closer the overall integrity is to 1, the better the integrity. The administrator sets the minimum expected threshold for integrity. When the overall integrity is lower than the minimum expected threshold, the output of the knowledge graph does not meet the integrity standard, and the administrator will conduct the next step of verification and optimization.

[0046] In summary, the present invention substitutes the meanings of polysemous words into the original paragraphs and scores them through a multiple scoring mechanism, and derives the adaptive meaning of the polysemous words according to the scoring results, thereby effectively preventing the ambiguity of paragraphs caused by the uncertainty of the meaning of polysemous words and enhancing the accuracy of the knowledge graph.

Claims

1. The AI-based intelligent knowledge graph construction method is characterized by The following steps are involved: Obtain the target subject domain and collect knowledge data from the data source according to the target subject domain; Perform data cleaning on the collected knowledge data to obtain cleaned knowledge data; Perform data extraction on the cleaned knowledge data to obtain knowledge extraction data; Extract polysemous words from knowledge extraction data, use a multiple scoring mechanism to determine the meaning of polysemous words, and derive the meaning of polysemous words; Extract data and polysemous meanings based on knowledge, and perform knowledge representation and storage; Conduct integrity assessment on the constructed knowledge graph and obtain an overall integrity index, which is used to measure the integrity of the knowledge graph. The constructed knowledge graph is verified and optimized according to the overall completeness index.

2. The AI-based intelligent knowledge graph construction method according to claim 1, characterized in that: The collecting of knowledge data from the data source according to the target subject domain includes the following steps: Obtain the target subject domain covered by the knowledge graph; Connect to the data source and identify the data source type; Determine the data collection tool according to the data source type, and use data mining algorithms and text analysis techniques to collect and extract knowledge data based on the target subject domain.

3. The AI-based intelligent knowledge graph construction method according to claim 1, characterized in that: The data cleaning of the collected knowledge data to obtain the cleaned knowledge data, and data extraction of the cleaned knowledge data to obtain the knowledge extraction data include the following steps: Remove duplicate records through unique identifiers and unify data in different formats into a standard format; By sorting out different data sources, merging relevant data, and removing redundant data, we can obtain cleaned knowledge data; Based on the cleaned knowledge data, entity extraction, relationship extraction and attribute extraction are performed through data extraction rules to obtain knowledge extraction data.

4. The AI-based intelligent knowledge graph construction method according to claim 1, characterized in that: The method of extracting polysemous words from the knowledge extraction data adopts a multiple scoring mechanism for the polysemous words, determines the meaning of the polysemous words through the multiple scoring mechanism, and obtains the meaning of the polysemous words, including the following steps: Based on the knowledge extraction of polysemous words in the data, the meaning 1, meaning 2, ..., meaning N of the polysemous words are extracted, where N is the meaning number; The meaning 1, meaning 2, ..., meaning N of the polysemous word are scored through a multiple scoring mechanism, and the meaning of the polysemous word is obtained based on the scoring results.

5. The AI-based intelligent knowledge graph construction method according to claim 4, characterized in that: The multiple scoring mechanism comprises the following steps: Get the meaning 1, meaning 2, ..., meaning N of a polysemous word; Substitute meaning 1, meaning 2, ..., meaning N into the original paragraph where the polysemous word is located one by one, and obtain paragraph 1, paragraph 2, ..., paragraph G, where G is the paragraph number; perform context consistency scoring on the original paragraphs and paragraph 1, paragraph 2, ..., paragraph G in turn, and obtain a context consistency score sequence, which includes context consistency score 1, context consistency score 2, ..., context consistency score F, where F is the context consistency score number; Count the number of identical highest scores in a sequence of context consistency scores: If the highest scores in the context consistency score sequence are the same, the meaning of the polysemous word corresponding to the highest score is the adapted meaning of the polysemous word in its original paragraph; If the number of the same highest scores in the context consistency score sequence is not unique, then the commonness score is performed on the paragraphs with the highest context consistency score to obtain a commonness score sequence. The commonness score includes commonness score 1, commonness score 2, ..., commonness score M, where M is the number of the same highest scores in the context consistency score sequence; If the number of highest scores in the commonness score sequence is the same, then the meaning of the polysemous word corresponding to the highest score is the adaptive meaning of the polysemous word in its original paragraph; If the number of paragraphs with the same highest score in the commonness score sequence is not unique, then the paragraphs with the highest commonness score are scored for sentiment consistency to obtain a sentiment consistency score sequence, which includes sentiment consistency score 1, sentiment consistency score 2, ..., sentiment consistency score P, where P is the number of paragraphs with the same highest score in the commonness score sequence; If the number of highest scores in the sentiment consistency score sequence is the same, the meaning of the polysemous word corresponding to the highest score is the adapted meaning of the polysemous word in its original paragraph. Otherwise, the management personnel are notified to manually adapt the meaning of the polysemous word.

6. The AI-based intelligent knowledge graph construction method according to claim 1, characterized in that: The construction of a knowledge graph based on knowledge extraction data and polysemous word meanings includes the following steps: Through string matching and semantic similarity calculation, similar entities are identified and aligned, and the attribute values ​​of knowledge extraction data are filled using the weighted average method; The knowledge extraction data and the meaning of polysemous words are structured, and the structured knowledge extraction data and the meaning of polysemous words are converted into triples to obtain nodes and edges of the knowledge graph; The nodes and edges of the obtained knowledge graph are stored in the graph database to construct the knowledge graph.

7. The AI-based intelligent knowledge graph construction method according to claim 1, characterized in that: The integrity evaluation of the constructed knowledge graph to obtain the overall integrity index includes the following steps: Obtain integrity evaluation indicators, perform quantitative analysis on nodes, edges, and attributes in the knowledge graph based on the obtained integrity evaluation indicators, and calculate the actual value of each indicator; Based on the actual value of the integrity assessment indicator, the overall integrity index of the knowledge graph is obtained through the overall integrity index calculation formula.

8. The AI-based intelligent knowledge graph construction method according to claim 7, characterized in that: The verification and optimization of the constructed knowledge graph according to the overall integrity index includes the following steps: Based on the obtained overall completeness index of the knowledge graph, the overall completeness index of the knowledge graph is compared with the minimum expected threshold of completeness; If the overall integrity index of the knowledge graph is not less than the minimum expected threshold of completeness, it is determined that the completeness of the knowledge graph has reached the expected level; If the overall integrity index of the knowledge graph is less than the minimum expected threshold for completeness, it is determined that the integrity of the knowledge graph does not reach the expected level, the knowledge graph is output as not meeting the integrity standard, and the administrator is notified to verify and optimize the knowledge graph.

9. The AI-based intelligent knowledge graph construction method according to claim 7, characterized in that: The steps for obtaining the overall integrity index are: The completeness evaluation indicators include the difference in the number of entities, the difference in the number of relationships, the difference in the number of attributes, the total number of retrievals, and the number of unsuccessful comparisons; Get a dictionary of word meanings; Through the deep learning model that is different from the data extraction rules, entity extraction, relationship extraction and attribute extraction are performed again from the knowledge extraction data; Combine the word meaning dictionary to construct a comparison vocabulary of the newly extracted entities, relations or attributes; Count and compare the number of entities, the number of relationships, and the number of attributes in the vocabulary and knowledge graph respectively, and compare them with each other to obtain the difference in the number of entities, the difference in the number of relationships, and the difference in the number of attributes; Randomly extract an entity, relationship or attribute from the comparison vocabulary, put it into the knowledge graph for comparison and search, and count the total number of searches and the number of unsuccessful comparisons; The overall integrity index is calculated using the overall integrity index calculation formula by combining the difference in the number of entities, the difference in the number of relationships, the difference in the number of attributes, the total number of searches and the number of unsuccessful comparisons.

10. The AI-based intelligent knowledge graph construction method according to claim 1, characterized in that: The overall integrity index calculation formula is: ； ； Where β is the overall integrity index, is the difference in the number of entities, is the difference in the number of relations, is the difference in the number of attributes, Y is the total number of searches, X is the number of unsuccessful comparisons, e is a natural constant, α1 is the weight of the difference in the number of entities for the overall completeness index, α2 is the weight of the difference in the number of relationships for the overall completeness index, α3 is the weight of the difference in the number of attributes for the overall completeness index, α4 is the weight of the ratio of the number of unsuccessful comparisons to the total number of searches for the overall completeness index, and f is the transition function.

Images