An automatic entity relation extraction method and system in knowledge graph construction

Abstract

This invention relates to the field of knowledge graph technology, and discloses an automated entity relation extraction method and system for knowledge graph construction. The method includes: deep cleaning of heterogeneous data to obtain target text data; contextual structure parsing of the target text data to identify core entities and inter-entity event descriptions; latent constraint analysis of the core entities to obtain a set of deep relations; logical conflict determination between the explicit relation statement set and the deep relation set of the target text data to obtain a set of non-contradictory relations; semantic connection completion of the set of non-contradictory relations to construct a fused relation network; and association reinforcement reconstruction of the fused relation network to confirm the final entity relation set of the target text data. This invention can improve the efficiency of automated entity relation extraction in knowledge graph construction.

Description

Technical Field

[0001] This invention relates to the field of knowledge graph technology, and in particular to an automated entity relation extraction method and system for knowledge graph construction. Background Technology

[0002] In the process of knowledge graph construction, automated entity relation extraction faces the challenge of complex heterogeneous data sources and diverse formats. Existing technologies struggle to completely remove non-textual content and redundant information during data cleaning, and insufficient character encoding uniformity and text filtering accuracy result in inconsistent target text data quality, creating potential problems for subsequent entity recognition and relation extraction. Furthermore, the contextual analysis of target text lacks systematicity, the relevance and accuracy of core entity selection are insufficient, and the structured integration of event descriptions between entities is inadequate, failing to fully preserve the key semantic information of entity relationships.

[0003] Existing technologies have limitations in entity relationship mining. They struggle to effectively infer potential relationships between core entities based on the inherent connections between semantic elements, lack a scientific mechanism for resolving logical conflicts between explicit and deep relationships, and their quantitative analysis of relationship priority is not precise enough. Furthermore, the completeness of semantic connection completion is insufficient, and the mining of potential relationships between non-directly connected entity pairs is inadequate. This leads to omissions and insufficient association strength in the constructed relationship network, ultimately affecting the completeness and reliability of the knowledge graph. Therefore, improving the accuracy and efficiency of automated entity relationship extraction in knowledge graph construction has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides an automated entity relation extraction method and system for knowledge graph construction to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, this invention provides an automated entity relation extraction method for knowledge graph construction, comprising: S1. Perform deep cleaning on the heterogeneous data to obtain the target text data of the heterogeneous data; S2. Perform contextual structured parsing on the target text data and identify the core entities and inter-entity event descriptions in the target text data; S3. Based on the event description between entities, perform potential constraint analysis on the core entities to obtain the deep relationship set between the core entities; S4. Perform a logical conflict determination on the explicit relation statement set and the deep relation set of the target text data to obtain the non-contradictory relation set of the target text data; S5. Perform semantic connection completion on the set of contradictory relations to construct the fusion relation network of the target text data; S6. Based on the deep relation set and the fused relation network, perform association strengthening and reconstruction on the fused relation network to confirm the final entity relation set of the target text data.

[0006] In a preferred embodiment, the deep cleaning of heterogeneous data to obtain target text data of the heterogeneous data includes: It combines web page data sources, database storage records, and structured documents into a raw document; The storage format of the original document is parsed, and document structure tags, header and footer information and non-text content in the original document are removed to obtain the plain text content of the original document; The plain text content is encoded uniformly, and the target text content in the standard text is filtered out based on the language category of the unified standard text. The target text content is processed into a continuous text stream to obtain the target text data of the heterogeneous data.

[0007] In a preferred embodiment, the step of performing contextualized structuring parsing on the target text data and identifying the core entities and inter-entity event representations in the target text data includes: Based on the topic's development and coherence in the target text data, the target text data is divided into continuous text intervals. Perform syntactic and semantic analysis on the continuous text intervals to obtain the semantic dependencies of the continuous text intervals; Based on the semantic dependency relationship, argument role recognition is performed on the continuous text interval to obtain candidate entities for the continuous text interval; Based on the centrality of the candidate entities and their semantic role relationship with the main predicates in the continuous text interval, the core entities among the candidate entities are selected. Based on the core entity and the main predicate, extract the relevant adverbial, attributive, and complement components from the continuous text interval; The core entities, main predicates, adverbs, attributives, and complements are structurally integrated to obtain the inter-entity event representation of the target text data.

[0008] In a preferred embodiment, the step of performing potential constraint analysis on the core entities based on the inter-entity event representation to obtain a deep set of relationships between the core entities includes: The syntactic components of the inter-entity event representation are decomposed to obtain the semantic elements of the inter-entity event representation; Based on the inherent relationships of the semantic elements, the inherent relationships of the core entities are inferred to obtain the potential causal relationships, temporal relationships and purpose relationships between the core entities; The potential causal relationships, the temporal relationships, and the purpose relationships are aggregated into an initial set of relationship hypotheses for the inter-entity event representation; Based on the contextual completeness and evidence strength of the inter-entity event descriptions, a logical strength quantification analysis is performed on the hypothetical relationships in the initial relation hypothesis set to obtain a semantic reasonableness score for the initial relation hypothesis set. The hypothetical relationships with semantic reasonableness scores higher than a preset threshold in the initial relation hypothesis set are integrated into a deep relation set between the core entities.

[0009] In a preferred embodiment, the step of performing logical conflict determination on the explicit relation statement set and the deep relation set of the target text data to obtain the conflict-free relation set of the target text data includes: Extract the set of explicit relational statements from the target text data; The explicit relation statement set and the deep relation set are unified in entity identification to identify the same entity pairs in the explicit relation statement set and the deep relation set; Calculate the priority value of the relation statement for the same entity pair to determine the relation type of the same entity pair; Based on the relationship type, triples are aggregated between the explicit relationship statement set and the non-conflicting relationship statements in the deep relationship set to obtain the non-conflicting relationship set of the target text data.

[0010] In a preferred embodiment, the formula for calculating the priority value of the relation statement is as follows: ; In the formula, Indicates the candidate relation type Priority of relational statements This represents the pre-defined explicit relationship credibility weight. This represents the preset depth relationship credibility weight. Indicates the candidate relation type The confidence level in the set of explicit relation statements. Indicates the candidate relation type The confidence level in the deep relation set. This represents the preset context consistency compensation coefficient. Indicates support for candidate relation types The set of semantic elements, Representing semantic elements For candidate relation types Support strength, This indicates a summation operation.

[0011] In a preferred embodiment, the step of performing semantic connection completion on the set of contradictory relations to construct a fusion relation network for the target text data includes: Traverse the set of relation triples without contradictions and extract the unique entity from the relation triples; A coupling analysis is performed on the number of existing relations and the semantic strength of the relation types of the unique entity to obtain the direct association degree between the unique entities; Based on the direct association degree, shared context parsing is performed on non-directly connected entity pairs in the set of no-conflict relationships to obtain the semantic proximity degree between the non-directly connected entity pairs. Based on the semantic proximity, implicit connection inference is performed on the non-directly connected entity pairs to obtain the potential relation supplement set of the non-directly connected entity pairs; Using the unique entity as a node and the relationships between entities in the contradictory relationship set and the potential relationship supplement set as connecting edges, a fusion relationship network for the target text data is constructed.

[0012] In a preferred embodiment, the step of performing shared context resolution on non-directly connected entity pairs in the conflict-free relation set based on the direct association degree to obtain the semantic proximity between the non-directly connected entity pairs includes: Locate the text interval in which the non-directly connected entity pairs co-occur in the target text data; The inter-entity event representations associated with the non-directly connected entity pair in the text interval are used as the shared semantic context of the non-directly connected entity pair. The interaction relationship of the non-directly connected entities in the shared semantic context is analyzed to obtain the interaction pattern between the semantic roles. Analyze the co-occurrence frequency and overlapping of descriptive attributes of the non-directly connected entity pairs in the shared semantic context; The semantic proximity between the non-directly connected entity pairs is obtained by weighted fusion of the intensity of the interaction pattern between the characters, the co-occurrence frequency, and the overlap of the descriptive attributes.

[0013] In a preferred embodiment, the step of performing association enhancement and reconstruction on the fused relationship network based on the deep relationship set and the fused relationship network to confirm the final entity relationship set of the target text data includes: Identify connections that are explicitly supported in the deep relation set but missing in the fused relation network as candidate edges to be enhanced in the fused relation network; A necessity tracing analysis is performed on the candidate edges to be enhanced to obtain the enhancement confidence of the candidate edges to be enhanced; Based on the enhanced confidence, the candidate edges to be enhanced that meet the preset enhancement conditions are integrated into the fused relation network to obtain the enhanced relation network of the fused relation network; The connection edges of the enhanced relation network are transformed into final relation triples, and the final relation triples are output as the final entity relation set of the target text data.

[0014] To address the aforementioned problems, this invention also provides an automated entity relation extraction system for knowledge graph construction, the system comprising: The heterogeneous data cleaning module is used to perform deep cleaning on heterogeneous data to obtain the target text data of the heterogeneous data. The contextual structure parsing module is used to perform contextual structure parsing on the target text data and identify the core entities and inter-entity event descriptions in the target text data. The potential constraint analysis module is used to perform potential constraint analysis on the core entities based on the event representation between the entities, and obtain the deep relationship set between the core entities; The logical conflict determination module is used to make a logical conflict determination between the explicit relation statement set and the deep relation set of the target text data, so as to obtain the non-contradictory relation set of the target text data. A semantic connection completion module is used to perform semantic connection completion on the set of contradictory relations in order to construct a fusion relation network of the target text data; The association strengthening and reconstruction module is used to perform association strengthening and reconstruction on the fused relationship network based on the deep relationship set and the fused relationship network, and to confirm the final entity relationship set of the target text data.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention achieves effective removal of redundant information, standardization and unification of character encoding, and accurate screening of target text through a multi-dimensional heterogeneous data deep cleaning process. It significantly improves the quality of text data at the data source level, providing a standardized and high-quality foundation for subsequent entity relationship extraction. It completes text interval division based on topic coherence, and achieves accurate screening of core entities by combining semantic dependency and semantic role analysis. By structurally integrating multiple types of grammatical components to form a complete inter-entity event description, the accuracy of core entity identification is greatly improved, and key semantic information related to entity relationships is fully preserved, laying a precise semantic foundation for subsequent deep relationship mining.

[0016] 2. This invention accurately infers the potential relationships between core entities through the inherent associations of semantic elements, and selects a high-reliability deep relationship set through quantitative analysis. By employing a scientific method for calculating the priority value of relationship statements, it achieves a seamless fusion of explicit and deep relationships, effectively avoiding relationship conflicts. Semantic proximity analysis fully explores the potential relationships between non-directly connected entities, completing semantic connection completion. Necessity tracing analysis then strengthens and reconstructs the fusion relationship network, comprehensively addressing any omissions in the relationship network. This series of operations comprehensively improves the completeness and accuracy of entity relationship extraction, effectively enhancing the overall efficiency of automated entity relationship extraction in knowledge graph construction. It also outputs a structurally sound and accurately accurate final entity relationship set, providing high-quality and highly reliable core data support for knowledge graph construction, ensuring the completeness and practicality of the knowledge graph. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an automated entity relation extraction method for knowledge graph construction according to an embodiment of the present invention. Figure 2 This is a functional module diagram of an automated entity relation extraction system for knowledge graph construction provided in an embodiment of the present invention; The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0018] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] This application provides an automated entity relation extraction method for knowledge graph construction. The execution subject of this automated entity relation extraction method for knowledge graph construction includes, but is not limited to, at least one of the following electronic devices that can be configured to execute the method provided in this application: a server, a terminal, etc. In other words, the automated entity relation extraction method for knowledge graph construction can be executed by software or hardware installed on a terminal device or a server device. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster. The server can be an independent server or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms.

[0020] Reference Figure 1The diagram shown is a flowchart illustrating an automated entity relation extraction method for knowledge graph construction according to an embodiment of the present invention. In this embodiment, the automated entity relation extraction method for knowledge graph construction includes: S1. Perform deep cleaning on the heterogeneous data to obtain the target text data of the heterogeneous data; In this embodiment of the invention, the step of performing deep cleaning on heterogeneous data to obtain target text data of the heterogeneous data includes: It combines web page data sources, database storage records, and structured documents into a raw document; The storage format of the original document is parsed, and document structure tags, header and footer information and non-text content in the original document are removed to obtain the plain text content of the original document; The plain text content is encoded uniformly, and the target text content in the standard text is filtered out based on the language category of the unified standard text. The target text content is processed into a continuous text stream to obtain the target text data of the heterogeneous data.

[0021] The system retrieves all web page content from a specified domain using web crawlers, extracts all records from the target data table using database queries, and collects structured documents with the extensions .docx, .pdf, and .xlsx from a specified storage path using file reading tools. All the retrieved web page content, database records, and complete information from the structured documents are then integrated and stored in a single file, which is the original document.

[0022] The storage format of the original document is identified based on its file extension. For HTML documents, its tag hierarchy is parsed; for docx documents, its underlying XML data structure is parsed; for PDF documents, its text flow and page layout relationship structure is parsed; and for SQL documents, its data fields and statement structure are parsed. For each document format, document structure tags are identified and removed one by one. Specifically, all tags containing "<" and ">" are removed from HTML documents; paragraph style tags and XML control tags are removed from docx documents; page layout control tags are removed from PDF documents; and structural symbols other than statement keywords are removed from SQL documents. At the same time, fixed content segments that appear repeatedly at the top and bottom of each page are identified as header and footer information and removed. Content in the document that does not belong to Chinese or English text, numbers, or common punctuation marks, including images, audio, video, formulas, and special symbols, is filtered out and removed. The remaining content containing only Chinese or English text, numbers, and common punctuation marks is the plain text content of the original document.

[0023] A character encoding detection tool is used to scan the encoding identifier of each character in the plain text content to determine the corresponding encoding format. All characters with different encoding formats such as UTF-8, GBK, GB2312, and ISO-8859-1 are uniformly converted to UTF-8 encoding format. During the conversion process, a mapping table between each encoding format and UTF-8 encoding is used to ensure accurate semantic correspondence for each character. The converted text is the standard text. The language category is determined by detecting the Unicode encoding range of the characters in the standard text: Chinese corresponds to the Unicode encoding range 4E00-9FFF, English corresponds to 0041-005A and 0061-007A, and numbers correspond to 0030-0039. (The text then abruptly shifts to a different topic: preset target...) The target language category is Chinese. Text segments in the standard text whose Unicode encoding of all characters falls within the range of 4E00-9FFF are selected. These text segments are then concatenated in their original order to form the target text content. After unifying the character encoding, Chinese and English can be preset as target language categories according to actual needs. Text segments in the standard text that conform to the encoding range of Chinese Unicode 4E00-9FFF and English Unicode 0041-005A and 0061-007A are selected respectively. The two types of text segments are concatenated in their original order to form the Chinese and English target text content, and each is processed as a continuous text stream to obtain the target text data corresponding to the Chinese and English heterogeneous corpora, thus achieving the classification standardization of multilingual corpora.

[0024] The target text content is split into multiple independent paragraphs by paragraph delimiters. The beginning and end of each paragraph are scanned one by one, and blank characters such as spaces, tabs, and newlines are identified and removed until the beginning and end of the paragraph are all valid characters. Then, according to the original order of each paragraph in the target text content, all processed paragraphs are connected in sequence to form a continuous text string without breaks, repetitions, or extra blank characters. This continuous text string is the target text data of the heterogeneous data.

[0025] The beneficial effects are that through a complete process of multi-source data aggregation, format parsing and redundant content removal, encoding standardization and language filtering, and continuous text stream processing, heterogeneous data is systematically cleaned. The final target text data contains only valid text information of the target language category, without redundant information such as document structure tags, headers and footers, and non-text content. The character encoding is unified and the text is continuous, providing a high-quality and standardized data source for the subsequent contextual structure parsing of the target text data, ensuring the accuracy and effectiveness of subsequent core entity recognition and inter-entity event description extraction.

[0026] S2. Perform contextual structured parsing on the target text data and identify the core entities and inter-entity event descriptions in the target text data; In this embodiment of the invention, the step of performing contextualized structuring parsing on the target text data and identifying the core entities and inter-entity event representations in the target text data includes: Based on the topic's development and coherence in the target text data, the target text data is divided into continuous text intervals. Perform syntactic and semantic analysis on the continuous text intervals to obtain the semantic dependencies of the continuous text intervals; Based on the semantic dependency relationship, argument role recognition is performed on the continuous text interval to obtain candidate entities for the continuous text interval; Based on the centrality of the candidate entities and their semantic role relationship with the main predicates in the continuous text interval, the core entities among the candidate entities are selected. Based on the core entity and the main predicate, extract the relevant adverbial, attributive, and complement components from the continuous text interval; The core entities, main predicates, adverbs, attributives, and complements are structurally integrated to obtain the inter-entity event representation of the target text data.

[0027] For target text data in different language categories, such as Chinese and English, appropriate grammatical and semantic analysis models are invoked respectively. Chinese text is matched with Modern Chinese grammar rules, and English text is matched with English grammar rules to conduct semantic dependency analysis. At the same time, the corresponding language argument role recognition system is adopted. Chinese is based on the Modern Chinese predicate-argument system, and English is based on the PennTreebank argument role system to complete the candidate entity screening. The core entity screening and the structured integration of event descriptions between entities all follow the grammatical logic and semantic expression habits of the corresponding language, so as to achieve accurate parsing and recognition in multilingual contexts.

[0028] Based on the connection features and referential logic of topics in the target text data, the topic initiation markers, transition markers, and summary markers in the text are scanned. At the same time, the referents of pronouns between sentences are checked. If the entity referred to by the pronoun exceeds the current paragraph and there is no clear semantic connection, or if the above-mentioned topic marker words appear, the text is divided into multiple continuous text intervals based on this point. The topic remains consistent within each interval, and the referential relationship is unbroken.

[0029] The analysis is carried out sentence by sentence in a continuous text interval. First, the part of speech and grammatical components of words in the sentence are identified. Then, the semantic relationship between words is analyzed, including the relationship between agent and patient, modifier and modified, restrictor and restricted, etc. These grammatical and semantic relationships are organized into structured information, clarifying the semantic orientation and dependent object of each word, and finally forming the semantic dependency relationship of the continuous text interval.

[0030] Based on the established semantic dependencies, the main verbs within the interval are focused on, and the nominal components associated with these verbs are identified. The argument roles they play are determined, such as the initiator, receiver, tool used, and location of the action. All nominal components that play argument roles are screened out and aggregated to form candidate entities for the continuous text interval.

[0031] The number of times each candidate entity appears in a continuous text interval and the number of semantic associations with other candidate entities are counted. The sum of the two is the centrality of the candidate entity, with a preset centrality threshold of 3. At the same time, the verb that appears most frequently in the interval and is used as the core action expression is identified as the main predicate. The semantic role relationship between the candidate entity and the main predicate is analyzed, and candidate entities that assume core semantic roles such as agent and patient and whose centrality reaches the preset threshold are selected. These entities are the core entities.

[0032] Using core entities and main predicates as the core retrieval objects, the system extracts attributive components that modify core entities within continuous text intervals, including adjectives, noun phrases, and prepositional phrases that directly modify them; it also extracts adverbial components that restrict or supplement the main predicate, including adverbs and prepositional phrases that indicate time, place, manner, or scope; and it extracts complement components that supplement the result or degree of the main predicate's action, including supplementary words and phrases after verbs, ensuring that all extracted components are directly related to the core entity or main predicate.

[0033] Following the semantic logical order of "attributive + core entity + adverbial + main predicate + complement", the components are arranged in a structured manner. If there are multiple core entities, the semantic relationship between each core entity and the main predicate is clarified. All components are organically integrated into complete semantic units. Each semantic unit clearly presents the core entity, core action and related modifiers and supplementary information. These semantic units together constitute the inter-entity event representation of the target text data.

[0034] The beneficial effects are: accurately dividing continuous text intervals through topic and referential features to ensure semantic coherence within the intervals; clarifying semantic dependencies between words through syntactic and semantic analysis, providing a solid basis for argument role identification; screening candidate entities based on argument roles, and then combining centrality and core semantic roles to screen core entities, ensuring the accuracy of core entities; comprehensively extracting relevant modifiers and supplementary components and integrating them in a structured manner to form a complete event description between entities, providing a comprehensive and accurate semantic foundation for subsequent analysis of potential relationships between core entities.

[0035] S3. Based on the event description between entities, perform potential constraint analysis on the core entities to obtain the deep relationship set between the core entities; In this embodiment of the invention, the step of performing potential constraint analysis on the core entities based on the inter-entity event representation to obtain a deep relationship set between the core entities includes: The syntactic components of the inter-entity event representation are decomposed to obtain the semantic elements of the inter-entity event representation; Based on the inherent relationships of the semantic elements, the inherent relationships of the core entities are inferred to obtain the potential causal relationships, temporal relationships and purpose relationships between the core entities; The potential causal relationships, the temporal relationships, and the purpose relationships are aggregated into an initial set of relationship hypotheses for the inter-entity event representation; Based on the contextual completeness and evidence strength of the inter-entity event descriptions, a logical strength quantification analysis is performed on the hypothetical relationships in the initial relation hypothesis set to obtain a semantic reasonableness score for the initial relation hypothesis set. The hypothetical relationships with semantic reasonableness scores higher than a preset threshold in the initial relation hypothesis set are integrated into a deep relation set between the core entities.

[0036] The sentence boundaries for describing events between entities are defined by punctuation marks. The grammatical components are decomposed sentence by sentence. First, the core entities in each sentence are identified. Then, the main predicates that can reflect the core actions are located. Next, the adverbial components that modify the main predicates, such as time, place, manner, and scope, are selected. The attributive components that limit the attributes of the core entities are extracted. Finally, the complement components that supplement and explain the result and degree of the main predicate's actions are found. Each core entity, main predicate, adverbial, attributive, and complement is treated as an independent semantic element. Each sentence is checked to ensure that no key grammatical components involved in the formation of the event are omitted.

[0037] By systematically examining the inherent connections between all semantic elements, if the action corresponding to core entity A directly acts on core entity B through the main predicate, causing a clearly identifiable change in the attribute characteristics and behavioral patterns of core entity B, or causing core entity B to produce a specific and perceptible result, then a potential causal relationship exists between core entity A and core entity B. If the actions corresponding to each core entity contain time markers such as "first," "later," or "following" in the text, or if the order of action execution can be clearly distinguished based on the logic of event development, then a potential temporal relationship exists between the core entity corresponding to the action that occurred first and the core entity corresponding to the action that occurred later. If the action of core entity A, as embodied by the main predicate, directly targets the achievement of a pre-set goal of core entity B, or can satisfy the needs explicitly stated by core entity B, then a potential purpose relationship exists between core entity A and core entity B.

[0038] All potential causal relationships, potential temporal relationships, and potential purpose relationships inferred through semantic element associations are comprehensively collected. They are first classified and organized into three categories: causal relationships, temporal relationships, and purpose relationships. A separate subset is established for each type of relationship. Then, the three subsets are integrated into a single set, which is the initial set of relational hypotheses for the description of events between entities.

[0039] The context of the event descriptions between entities corresponding to each hypothetical relationship in the initial relation hypothesis set is examined to determine whether it contains at least two pieces of background information, supplementary explanations, or logical groundwork supporting the establishment of the relationship. This serves as the evaluation standard for contextual integrity; if it meets the standard, the context is considered complete, and if not, it is considered incomplete. The number of supporting semantic elements corresponding to each hypothetical relationship is counted; the more elements, the stronger the evidence. 1-2 supporting semantic elements indicate low evidence strength, 3-4 supporting semantic elements indicate medium evidence strength, and 5 or more supporting semantic elements indicate high evidence strength. A comprehensive judgment is made by combining contextual integrity and evidence strength. A complete context receives 4 points, an incomplete context receives 0-3 points, and a high evidence strength receives 6 points, a medium context receives 3-5 points, and a low context receives 0-2 points. The sum of the two scores is the semantic reasonableness score for each hypothetical relationship, ranging from 0 to 10 points. A higher score indicates a stronger logical strength for the relationship.

[0040] The preset semantic reasonableness score threshold is 6 points. The semantic reasonableness score of each hypothetical relationship in the initial relation hypothesis set is compared with the preset threshold one by one. All hypothetical relationships with scores greater than 6 points are selected and sorted according to the original classification method of potential causal relationship, potential temporal relationship, and potential purpose relationship to form a unified set with a clear structure. This set is the deep relation set between core entities.

[0041] The beneficial effects are as follows: by accurately dissecting grammatical components sentence by sentence, the semantic elements are ensured to be complete and independent, providing a solid foundation for relation inference; by clarifying the criteria for judging three types of potential relations, the relation inference logic is made clear and reproducible; by quantifying the completeness of the context and the strength of evidence to score the semantic reasonableness, the hypothetical relations are scientifically screened, effectively eliminating hypothetical relations with insufficient logical support. The resulting deep relation set can truly and accurately reflect the intrinsic connections between core entities, providing accurate and reliable relation data support for subsequent judgments on logical conflicts between explicit relations and deep relations, and ensuring the effectiveness and accuracy of subsequent processing steps from the source.

[0042] S4. Perform a logical conflict determination on the explicit relation statement set and the deep relation set of the target text data to obtain the non-contradictory relation set of the target text data; In this embodiment of the invention, the step of performing logical conflict determination on the explicit relation statement set and the deep relation set of the target text data to obtain a conflict-free relation set of the target text data includes: Extract the set of explicit relational statements from the target text data; The explicit relation statement set and the deep relation set are unified in entity identification to identify the same entity pairs in the explicit relation statement set and the deep relation set; Calculate the priority value of the relation statement for the same entity pair to determine the relation type of the same entity pair; Based on the relationship type, triples are aggregated between the explicit relationship statement set and the non-conflicting relationship statements in the deep relationship set to obtain the non-conflicting relationship set of the target text data.

[0043] The formula for calculating the priority value of the relation statement is as follows: ; In the formula, Indicates the candidate relation type Priority of relational statements This represents the pre-defined explicit relationship credibility weight. This represents the preset depth relationship credibility weight. Indicates the candidate relation type The confidence level in the set of explicit relation statements. Indicates the candidate relation type The confidence level in the deep relation set. This represents the preset context consistency compensation coefficient. Indicates support for candidate relation types The set of semantic elements, Representing semantic elements For candidate relation types Support strength, This indicates a summation operation.

[0044] The target text data is scanned sentence by sentence to identify sentences containing explicit relational markers such as "is", "belongs to", "related to", "leads to", and "prior to". Entity pairs and direct relational expressions between entity pairs are extracted from these sentences. Each entity pair and its corresponding direct relational expression are combined into an independent relational statement unit. The set of all relational statement units is the explicit relational statement set of the target text data.

[0045] Extract all entities from the explicit relation statement set and the deep relation set one by one, and establish an entity synonym mapping table. Associate and match the full name, abbreviation, and standard name of the entity with common expressions. For example, mark "People's Republic of China" and "China" as the same entity. According to the entity synonym mapping table, replace all entities in the two sets with the preset standard name as the entity identifier. After completing the entity identifier unification, traverse the relation statement units in the two sets and filter out the entity combinations with the same entity identifier. These entity combinations are the same entity pairs in the explicit relation statement set and the deep relation set.

[0046] The confidence weight of explicit relations is preset to 0.4, the confidence weight of deep relations to 0.3, and the context consistency compensation coefficient to 0.3. For each candidate relation type of each identical entity pair, the number of occurrences of the candidate relation type in the explicit relation statement set is counted. The number of occurrences is divided by the total number of relation statements of the identical entity pair in the explicit relation statement set to obtain the confidence of the candidate relation type in the explicit relation statement set. The semantic reasonableness score of the candidate relation type in the deep relation set is extracted, and the score is divided by 10 to obtain the confidence of the candidate relation type in the deep relation set. The core entities, main predicates, adverbs, and other semantic elements directly related to the candidate relation type are selected to form a supporting semantic element set. Each semantic element is assigned a score of 1 to 3 according to the degree of relevance. The support strength is calculated as follows: the most direct association receives 3 points, a relatively direct association receives 2 points, and an indirect association receives 1 point. The sum of the support strengths of all semantic elements is calculated. The confidence weight of explicit relations is multiplied by the confidence weight in the explicit relation set, the confidence weight of deep relations is multiplied by the confidence weight in the deep relation set, and the context consistency compensation coefficient is multiplied by the sum of the support strengths of semantic elements. The three products are added together to obtain the relation statement priority value of the candidate relation type. The relation statement priority values ​​of all candidate relation types for the same entity pair are compared, and the candidate relation type with the highest priority value is selected as the final relation type for the same entity pair. If there are multiple candidate relation types with the same and highest priority values, then all of these candidate relation types are determined as the relation type for the same entity pair.

[0047] The explicit relation statement set and the relation statement units in the deep relation set are classified by entity pairs. The relation types corresponding to different relation statement units of the same entity pair are compared. If the relation types are semantically contradictory, such as one being a causal relationship and the other being an unrelated relationship, they are determined to be conflicting relation statements. If the relation types are semantically compatible or consistent, they are determined to be non-conflicting relation statements. All non-conflicting relation statements are organized into standard relation triples according to the structure of "entity A - relation type - entity B". Only one triple is retained for the same relation type of the same entity pair. Different non-conflicting relation types form independent triples. All standard relation triples are integrated and gathered to form the set of non-conflicting relations of the target text data.

[0048] The beneficial effects are as follows: Clear explicit relation extraction rules ensure the completeness and accuracy of the explicit relation statement set; entity synonym mapping table unifies entity identification, guaranteeing the accuracy of identifying identical entity pairs; quantitative calculation of relation statement priority values ​​determines relation types, making the determination process logically clear and reproducible, avoiding errors caused by subjective judgment; clear conflict determination criteria screen non-conflicting relation statements and aggregate them into triples, effectively eliminating contradictory relations. The resulting set of conflict-free relations is structurally standardized and semantically consistent, providing a high-quality, conflict-free relation data foundation for subsequent semantic connection completion and the construction of a fused relation network, ensuring the smoothness of the subsequent construction process and the reliability of the fused relation network.

[0049] S5. Perform semantic connection completion on the set of contradictory relations to construct the fusion relation network of the target text data; In this embodiment of the invention, the step of performing semantic connection completion on the set of contradictory relations to construct the fusion relation network of the target text data includes: Traverse the set of relation triples without contradictions and extract the unique entity from the relation triples; A coupling analysis is performed on the number of existing relations and the semantic strength of the relation types of the unique entity to obtain the direct association degree between the unique entities; Based on the direct association degree, shared context parsing is performed on non-directly connected entity pairs in the set of no-conflict relationships to obtain the semantic proximity degree between the non-directly connected entity pairs. Based on the semantic proximity, implicit connection inference is performed on the non-directly connected entity pairs to obtain the potential relation supplement set of the non-directly connected entity pairs; Using the unique entity as a node and the relationships between entities in the contradictory relationship set and the potential relationship supplement set as connecting edges, a fusion relationship network for the target text data is constructed.

[0050] Based on the direct association degree, the process of performing shared context resolution on non-directly connected entity pairs in the conflict-free relation set to obtain the semantic proximity between the non-directly connected entity pairs includes: Locate the text interval in which the non-directly connected entity pairs co-occur in the target text data; The inter-entity event representations associated with the non-directly connected entity pair in the text interval are used as the shared semantic context of the non-directly connected entity pair. The interaction relationship of the non-directly connected entities in the shared semantic context is analyzed to obtain the interaction pattern between the semantic roles. Analyze the co-occurrence frequency and overlapping of descriptive attributes of the non-directly connected entity pairs in the shared semantic context; The semantic proximity between the non-directly connected entity pairs is obtained by weighted fusion of the intensity of the interaction pattern between the characters, the co-occurrence frequency, and the overlap of the descriptive attributes.

[0051] Iterate through each relation triple in the set of contradictory relations one by one, extract the subject entity and object entity from each triple, store all extracted entities into a temporary set, and remove duplicate entities by comparing the canonical names of the entities in the temporary set. The final set of entities containing only unique canonical names is the unique entity.

[0052] The number of existing relations for each unique entity in the set of no-contradiction relations is counted. At the same time, a fixed semantic strength score is assigned to different relation types: causal relations are scored 3 points, temporal relations are scored 2 points, purpose relations are scored 2 points, and other related relations are scored 1 point. The number of existing relations for each unique entity is multiplied by the semantic strength score of the corresponding relation type, and the sum is then divided by the product of the maximum number of relations that the unique entity may have and the highest semantic strength score of 3 points. The maximum number of relations is calculated by subtracting 1 from the total number of unique entities. The result is the direct association degree between the unique entity and every other unique entity. The association degree value ranges from 0 to 1.

[0053] Entity pairs that do not form direct relation triples in the set of no contradictory relations are selected as non-directly connected entity pairs. Each non-directly connected entity pair is scanned one by one to find sentences or consecutive paragraphs in the target text data that contain the canonical names of both entities in the entity pair. These sentences or paragraphs are identified as the text intervals in which the non-directly connected entity pair appears.

[0054] Within the identified co-occurring text intervals, identify all inter-entity event statements that contain any one of the entities in the non-directly connected entity pair and describe the related scenarios of the two entities. Collect and organize all these event statements as the shared semantic context of the non-directly connected entity pair.

[0055] The event representations between each entity in the shared semantic context are analyzed to determine the semantic role played by each entity in the event representation in the non-directly connected entity pair, such as agent, patient, location, and tool. Based on the pairing relationship of these semantic roles, the interaction relationship pattern between the two is judged, including three modes: collaborative, influential, and related. The collaborative mode refers to the two participating in the same action together, the influential mode refers to the action of one party having an indirect effect on the other party, and the related mode refers to the two being related through common attributes.

[0056] The number of times non-directly connected entity pairs co-occur in a shared semantic context is called the co-occurrence frequency. The descriptive attributes of each entity are extracted, including category, features, functions, etc. The average number of completely identical descriptive attributes between the two entities is calculated as the descriptive attribute overlap, which ranges from 0 to 1.

[0057] The interaction mode intensity weight is preset to 0.4, the co-occurrence frequency weight is preset to 0.3, and the description attribute overlap weight is preset to 0.3. Fixed intensity scores are assigned to the three interaction modes: 10 for collaboration, 7 for influence, and 4 for association. The interaction mode intensity score is multiplied by the corresponding weight, the co-occurrence frequency is divided by 10 and then multiplied by the corresponding weight, and the description attribute overlap is multiplied by 10 and then multiplied by the corresponding weight. The three calculation results are added together, and the resulting value from 0 to 10 is the semantic proximity between non-directly connected entity pairs.

[0058] The preset semantic proximity threshold is 6 points, and the semantic proximity score ranges from 0 to 10 points. The higher the score, the closer the semantic association between the non-directly connected entity pairs. The semantic proximity of the non-directly connected entity pairs is compared with this threshold. If the proximity is greater than 6 points, the potential relationship type between the two is inferred based on the shared semantic context and interaction pattern, including potential association relationship, potential indirect causal relationship, and potential collaborative relationship. All non-directly connected entity pairs that meet the conditions and their corresponding potential relationship types are organized into a set, which is the potential relationship supplement set of non-directly connected entity pairs.

[0059] Using unique entities as nodes in the fusion relationship network, the relationships between entities in the non-contradictory relationship set and the potential relationships in the potential relationship supplement set are used as the network's connecting edges. The relationships in the non-contradictory relationship set are the definite connecting edges, and the relationships in the potential relationship supplement set are the supplementary connecting edges. According to the semantic association logic of the relationships between entities, the nodes and connecting edges are organically combined to construct a fusion relationship network of target text data with a complete structure and comprehensive relationships.

[0060] The beneficial effects are as follows: by systematically traversing and extracting unique entities, the uniqueness and accuracy of network nodes are ensured; by quantitatively calculating direct correlation and semantic proximity, precise basis is provided for the analysis of relationships between entities; by multi-dimensionally analyzing shared context, judging interaction patterns, and analyzing co-occurrence frequency and attribute overlap, scientific inference of potential relationships between non-directly connected entities is achieved, effectively supplementing implicit connections between entities; the finally constructed fusion relationship network integrates definite and potential relationships, with clear and comprehensive logic of node and edge associations, providing a structurally complete and relationally rich network foundation for subsequent association strengthening and reconstruction, and ensuring the integrity and reliability of the final entity relationship set.

[0061] S6. Based on the deep relation set and the fused relation network, perform association strengthening and reconstruction on the fused relation network to confirm the final entity relation set of the target text data.

[0062] In this embodiment of the invention, the step of performing association enhancement and reconstruction on the fused relationship network based on the deep relationship set and the fused relationship network to confirm the final entity relationship set of the target text data includes: Identify connections that are explicitly supported in the deep relation set but missing in the fused relation network as candidate edges to be enhanced in the fused relation network; A necessity tracing analysis is performed on the candidate edges to be enhanced to obtain the enhancement confidence of the candidate edges to be enhanced; Based on the enhanced confidence, the candidate edges to be enhanced that meet the preset enhancement conditions are integrated into the fused relation network to obtain the enhanced relation network of the fused relation network; The connection edges of the enhanced relation network are transformed into final relation triples, and the final relation triples are output as the final entity relation set of the target text data.

[0063] Traverse each connection in the deep relation set, which contains entity pairs and their corresponding relation types. Compare each connection edge in the fused relation network one by one, and check if there are any connection edges with the same entity pairs and relation types. If no matching connection edge is found in the fused relation network, then the connection in the deep relation set is identified as a candidate edge to be enhanced in the fused relation network.

[0064] For each candidate edge to be enhanced, a necessity tracing analysis is performed to extract the semantic reasonableness score corresponding to the candidate edge in the deep relation set. This score accounts for 60% of the weight in the enhancement confidence calculation. At the same time, the correlation of the entity pair corresponding to the candidate edge in the fused relation network is analyzed, including the direct correlation, semantic proximity, and the association logic between the entity pair and other entities in the network. If the direct correlation is ≥0.7, the semantic proximity is ≥6, and there are at least 2 other entities that have a direct connection relationship with the entity pair and support the validity of the candidate edge, then the correlation support score is full, accounting for 40% of the weight in the enhancement confidence calculation. The semantic reasonableness score is converted proportionally and added to the correlation support score to obtain the enhancement confidence of the candidate edge to be enhanced, with a confidence range of 0-10.

[0065] Based on the confidence score range of 0-10, the preset enhancement condition is an enhancement confidence score ≥ 7. The enhancement confidence score of each candidate edge to be enhanced is compared with the preset condition. All candidate edges with a confidence score of 7 or higher are selected, standardized according to the connection edge format of the fusion relationship network, and then integrated into the fusion relationship network one by one. It is ensured that the newly added connection edges have no logical conflict with the original connection edges. The network formed after integration is the enhanced relationship network of the fusion relationship network.

[0066] Traverse all the connection edges in the enhanced relation network. Each connection edge corresponds to a set of entity pairs and a clear relation type. According to the standard structure of "entity A - relation type - entity B", each connection edge is transformed into an independent final relation triple. Collect all the transformed final relation triples and sort them according to the first letter of the entity pair to form a set with a standardized structure and clear relations. This set is the final entity relation set of the target text data. Output the final entity relation set in a readable text format through a data storage tool to ensure that the output content completely retains all triple information.

[0067] The beneficial effects are as follows: by accurately identifying the missing effective connections in the deep relation set as candidate edges to be strengthened, the targeted nature of relation supplementation is ensured; by using multi-dimensional source analysis to quantify and enhance confidence, the selection of candidate edges is made scientifically based, avoiding the redundant addition of invalid relations; by integrating high-confidence candidate edges to strengthen the relation network, the relation omissions in the original network are made up for, and the integrity and correlation strength of the network are improved; finally, the final entity relation set output has a standard structure, accurate and comprehensive relations, and fully presents all kinds of effective relations between entities in the target text data, providing high-quality and highly reliable core data support for the construction of knowledge graphs, ensuring the accuracy and practicality of knowledge graphs.

[0068] like Figure 2The diagram shown is a functional block diagram of an automated entity relationship extraction system for knowledge graph construction provided in an embodiment of the present invention.

[0069] The automated entity relation extraction system 100 for knowledge graph construction described in this invention can be installed in an electronic device. Depending on the functions implemented, the automated entity relation extraction system 100 for knowledge graph construction may include a heterogeneous data cleaning module 101, a contextual structure parsing module 102, a latent constraint analysis module 103, a logical conflict judgment module 104, a semantic connection completion module 105, and an association enhancement and reconstruction module 106. The modules described in this invention can also be referred to as units, which are a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, stored in the memory of the electronic device.

[0070] In this embodiment, the functions of each module / unit are as follows: The heterogeneous data cleaning module 101 is used to perform deep cleaning on heterogeneous data to obtain the target text data of the heterogeneous data. The contextual structure parsing module 102 is used to perform contextual structure parsing on the target text data and identify the core entities and inter-entity event descriptions in the target text data. The potential constraint analysis module 103 is used to perform potential constraint analysis on the core entities based on the inter-entity event representation to obtain a deep relationship set between the core entities; The logical conflict decision module 104 is used to make a logical conflict decision between the explicit relation statement set and the deep relation set of the target text data to obtain a non-contradictory relation set of the target text data. The semantic connection completion module 105 is used to perform semantic connection completion on the set of contradictory relations in order to construct the fusion relation network of the target text data; The association strengthening and reconstruction module 106 is used to perform association strengthening and reconstruction on the fused relationship network based on the deep relationship set and the fused relationship network, and to confirm the final entity relationship set of the target text data.

[0071] In the several embodiments provided by this invention, it should be understood that the disclosed methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and other division methods may be used in actual implementation.

[0072] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0073] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0074] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0075] This application embodiment can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results.

[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An automated entity relation extraction method for knowledge graph construction, characterized in that, The method includes: S1. Perform deep cleaning on the heterogeneous data to obtain the target text data of the heterogeneous data; S2. Perform contextual structured parsing on the target text data and identify the core entities and inter-entity event descriptions in the target text data; S3. Based on the event description between entities, perform potential constraint analysis on the core entities to obtain the deep relationship set between the core entities; S4. Perform a logical conflict determination on the explicit relation statement set and the deep relation set of the target text data to obtain the non-contradictory relation set of the target text data; S5. Perform semantic connection completion on the set of contradictory relations to construct the fusion relation network of the target text data; S6. Based on the deep relation set and the fused relation network, perform association strengthening and reconstruction on the fused relation network to confirm the final entity relation set of the target text data.

2. The automated entity relation extraction method in knowledge graph construction as described in claim 1, characterized in that, The deep cleaning of heterogeneous data to obtain the target text data of the heterogeneous data includes: It combines web page data sources, database storage records, and structured documents into a raw document; The storage format of the original document is parsed, and document structure tags, header and footer information and non-text content in the original document are removed to obtain the plain text content of the original document; The plain text content is encoded uniformly, and the target text content in the standard text is filtered out based on the language category of the unified standard text. The target text content is processed into a continuous text stream to obtain the target text data of the heterogeneous data.

3. The automated entity relation extraction method in knowledge graph construction as described in claim 1, characterized in that, The step of performing contextualized structuring parsing on the target text data and identifying the core entities and inter-entity event representations in the target text data includes: Based on the topic's development and coherence in the target text data, the target text data is divided into continuous text intervals. Perform syntactic and semantic analysis on the continuous text intervals to obtain the semantic dependencies of the continuous text intervals; Based on the semantic dependency relationship, argument role recognition is performed on the continuous text interval to obtain candidate entities for the continuous text interval; Based on the centrality of the candidate entities and their semantic role relationship with the main predicates in the continuous text interval, the core entities among the candidate entities are selected. Based on the core entity and the main predicate, extract the relevant adverbial, attributive, and complement components from the continuous text interval; The core entities, main predicates, adverbs, attributives, and complements are structurally integrated to obtain the inter-entity event representation of the target text data.

4. The automated entity relation extraction method in knowledge graph construction as described in claim 1, characterized in that, The process of performing potential constraint analysis on the core entities based on the inter-entity event representation to obtain a deep set of relationships between the core entities includes: The syntactic components of the inter-entity event representation are decomposed to obtain the semantic elements of the inter-entity event representation; Based on the inherent relationships of the semantic elements, the inherent relationships of the core entities are inferred to obtain the potential causal relationships, temporal relationships and purpose relationships between the core entities; The potential causal relationships, the temporal relationships, and the purpose relationships are aggregated into an initial set of relationship hypotheses for the inter-entity event representation; Based on the contextual completeness and evidence strength of the inter-entity event descriptions, a logical strength quantification analysis is performed on the hypothetical relationships in the initial relation hypothesis set to obtain a semantic reasonableness score for the initial relation hypothesis set. The hypothetical relationships with semantic reasonableness scores higher than a preset threshold in the initial relation hypothesis set are integrated into a deep relation set between the core entities.

5. The automated entity relation extraction method in knowledge graph construction as described in claim 1, characterized in that, The step of performing logical conflict determination between the explicit relation statement set and the deep relation set of the target text data to obtain the conflict-free relation set of the target text data includes: Extract the set of explicit relational statements from the target text data; The explicit relation statement set and the deep relation set are unified in entity identification to identify the same entity pairs in the explicit relation statement set and the deep relation set; Calculate the priority value of the relation statement for the same entity pair to determine the relation type of the same entity pair; Based on the relationship type, triples are aggregated between the explicit relationship statement set and the non-conflicting relationship statements in the deep relationship set to obtain the non-conflicting relationship set of the target text data.

6. The automated entity relation extraction method in knowledge graph construction as described in claim 5, characterized in that, The formula for calculating the priority value of the relation statement is as follows: ； In the formula, Indicates the candidate relation type Priority of relational statements This represents the pre-defined explicit relationship credibility weight. This represents the preset depth relationship credibility weight. Indicates the candidate relation type The confidence level in the set of explicit relation statements. Indicates the candidate relation type The confidence level in the deep relation set. This represents the preset context consistency compensation coefficient. Indicates support for candidate relation types The set of semantic elements, Representing semantic elements For candidate relation types Support strength, This indicates a summation operation.

7. The automated entity relation extraction method in knowledge graph construction as described in claim 1, characterized in that, The step of performing semantic connection completion on the set of contradictory relations to construct the fusion relation network of the target text data includes: Traverse the set of relation triples without contradictions and extract the unique entity from the relation triples; A coupling analysis is performed on the number of existing relations and the semantic strength of the relation types of the unique entity to obtain the direct association degree between the unique entities; Based on the direct association degree, shared context parsing is performed on non-directly connected entity pairs in the set of no-conflict relationships to obtain the semantic proximity degree between the non-directly connected entity pairs. Based on the semantic proximity, implicit connection inference is performed on the non-directly connected entity pairs to obtain the potential relation supplement set of the non-directly connected entity pairs; Using the unique entity as a node and the relationships between entities in the contradictory relationship set and the potential relationship supplement set as connecting edges, a fusion relationship network for the target text data is constructed.

8. The automated entity relation extraction method in knowledge graph construction as described in claim 7, characterized in that, Based on the direct association degree, the process of performing shared context resolution on non-directly connected entity pairs in the conflict-free relation set to obtain the semantic proximity between the non-directly connected entity pairs includes: Locate the text interval in which the non-directly connected entity pairs co-occur in the target text data; The inter-entity event representations associated with the non-directly connected entity pair in the text interval are used as the shared semantic context of the non-directly connected entity pair. The interaction relationship of the non-directly connected entities in the shared semantic context is analyzed to obtain the interaction pattern between the semantic roles. Analyze the co-occurrence frequency and overlapping of descriptive attributes of the non-directly connected entity pairs in the shared semantic context; The semantic proximity between the non-directly connected entity pairs is obtained by weighted fusion of the intensity of the interaction pattern between the characters, the co-occurrence frequency, and the overlap of the descriptive attributes.

9. The automated entity relation extraction method in knowledge graph construction as described in claim 1, characterized in that, The step of performing association enhancement and reconstruction on the fused relationship network based on the deep relationship set and the fused relationship network to confirm the final entity relationship set of the target text data includes: Identify connections that are explicitly supported in the deep relation set but missing in the fused relation network as candidate edges to be enhanced in the fused relation network; A necessity tracing analysis is performed on the candidate edges to be enhanced to obtain the enhancement confidence of the candidate edges to be enhanced; Based on the enhanced confidence, the candidate edges to be enhanced that meet the preset enhancement conditions are integrated into the fused relation network to obtain the enhanced relation network of the fused relation network; The connection edges of the enhanced relation network are transformed into final relation triples, and the final relation triples are output as the final entity relation set of the target text data.

10. An automated entity relation extraction system for knowledge graph construction, characterized in that, The system is used to implement the automated entity relation extraction method in knowledge graph construction as described in claim 1, the system comprising: The heterogeneous data cleaning module is used to perform deep cleaning on heterogeneous data to obtain the target text data of the heterogeneous data. The contextual structure parsing module is used to perform contextual structure parsing on the target text data and identify the core entities and inter-entity event descriptions in the target text data. The potential constraint analysis module is used to perform potential constraint analysis on the core entities based on the event representation between the entities, and obtain the deep relationship set between the core entities; The logical conflict determination module is used to make a logical conflict determination between the explicit relation statement set and the deep relation set of the target text data, so as to obtain the non-contradictory relation set of the target text data. A semantic connection completion module is used to perform semantic connection completion on the set of contradictory relations in order to construct a fusion relation network of the target text data; The association strengthening and reconstruction module is used to perform association strengthening and reconstruction on the fused relationship network based on the deep relationship set and the fused relationship network, and to confirm the final entity relationship set of the target text data.

Images