Method and device for shaping personnel relationship based on intelligent entity relationship extraction

Abstract

The embodiment of the application provides a personnel relationship shaping method and device based on intelligent entity relationship extraction, and the network structure and the personnel relationship shaping method based on the intelligent entity relationship extraction can be used to analyze and shape the demand of the trust relationship in specific scenes in different industry fields such as education, law and catering, realize entity identification of massive multi-source heterogeneous target network emotion data such as Internet data, passwords, public software tools, special software tools and technical knowledge bases, analyze and extract the correlation of trust entities such as authentication, communication, social contact and business, construct a trust relationship knowledge graph based on a relationship graph model, realize comprehensive shaping of network structure relationship and personnel relationship, provide reliable and effective target trust relationship data for users, and improve the utilization rate of network emotion data.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a method and apparatus for shaping human relationships based on intelligent entity relationship extraction. Background Technology

[0002] In today's era of deep integration of digitalization and networking, online public opinion analysis, social network mining, and enterprise organizational management all rely on the in-depth mining and understanding of entities and their relationships within massive amounts of multi-source heterogeneous data. These data sources are wide-ranging, including publicly available internet information, internal enterprise documents, communication records, social media content, and business system logs. Their diverse forms, varied structures, and frequent dynamic evolution constitute a highly complex information environment.

[0003] Traditional methods for analyzing network structures and interpersonal relationships primarily rely on manual rule extraction, basic statistical models, or shallow machine learning techniques, which have significant limitations. On the one hand, these methods have low levels of automation, heavily depending on manual annotation and analysis by domain experts, making it difficult to handle large-scale data and real-time requirements. On the other hand, due to insufficient understanding of textual semantics, especially when dealing with unstructured, semantically ambiguous, or implicitly relational data, their extraction accuracy and recall are both unsatisfactory, making it difficult to support the construction of fine-grained, multi-dimensional relationship networks.

[0004] While existing technologies have attempted to incorporate natural language processing and machine learning methods for entity recognition and relation extraction, they generally face the following core challenges: First, weak ability to deeply integrate and uniformly represent multi-source heterogeneous data; second, insufficient ability to jointly extract and accurately model complex trust relationships such as authentication, communication, social interaction, and business relationships; and third, a lack of end-to-end intelligent analysis frameworks that can adapt to different industry scenarios and possess semantic enhancement and noise suppression capabilities. Therefore, there is an urgent need for an intelligent method that can automatically, accurately, and across scenarios extract entities from diverse data and shape deep relational networks to improve data value density and decision support efficiency. Summary of the Invention

[0005] To address the problems in the existing technology, this application provides a method and apparatus for shaping personnel relationships based on intelligent entity relationship extraction, which can realize the comprehensive shaping of network structure relationships and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network intelligence data.

[0006] To solve at least one of the above problems, this application provides the following technical solution: Firstly, this application provides a method for shaping personnel relationships based on intelligent entity relationship extraction, including: Acquire the core resources upon which trust relationships are built; categorize these core resources and determine the information extraction rules for each type; the types of core resources include authentication relationships, communication relationships, social relationships, and business relationships; Based on the information extraction rules, raw text data is collected from core resources of each type, and named entity recognition is performed on the raw text data based on a deep learning model to obtain the basic corpus of raw text data of each type. Entity relation extraction methods, either individually or in combination, are used to extract target triples from various types of basic corpora. Specifically, authentication relations are extracted using named entity relation extraction based on joint learning, as well as entity relation extraction combining deep learning and remote supervision; connectivity relations are extracted using entity relation extraction based on deep learning, as well as entity relation extraction combining deep learning and remote supervision; social relations are extracted using joint entity relation extraction based on contrastive learning and text semantic enhancement; and business relations are extracted using named entity relation extraction based on joint learning, as well as joint entity relation extraction based on contrastive learning and text semantic enhancement. Construct a trust relationship knowledge graph based on triples of various types.

[0007] Furthermore, core resources for authentication relationships include personnel attendance records, organizational rosters, and contact lists; Core resources for communication relationships include address book, call logs, SMS sending and receiving records, and email sending and receiving records; Core resources for social relationships include household registration information or contact information, and social groups; Core resources for business relationships include internal organizational rosters, address books, business documents, API call logs, and code files.

[0008] Furthermore, the rules for extracting information about authentication relationships include: Grouping data from different organizations using organizational tag data from the social worker relationship resource database; Further grouping based on the data tags of each organization to obtain data at the departmental level; Based on information related to personnel's work and business, determine the authentication relationships of personnel at the resource access level; At the network level, the registered user list, access permissions, and historical records are obtained through the website application user database, permission database, resource access logs, and network domain / workgroup records. The rules for extracting core resource information about communication relationships include: Extract communication relationships between people from communication records; The communication intimacy level is extracted by communication frequency information, which includes call frequency, SMS sending and receiving frequency, and email sending and receiving frequency. The rules for extracting core resource information from social relationships include: Extract social relationship information from notes; Use social group data to identify and classify different types of social relationships; The rules for extracting core resource information related to business relationships include: Business relationships are obtained through business association information; the business association information includes communication and interaction records, business transaction records, and program releases. Analyze entity relationships in business processes based on business documents and API call logs.

[0009] Furthermore, the step of performing named entity recognition on the original text data based on a deep learning model to obtain the basic corpus of various types of original text data includes: The original text data is vectorized to obtain vectorized text by dynamically training word vectors using a pre-trained language model. The vectorized text is input into the feature extractor to further extract text features and obtain a high-level feature vector. The high-level feature vectors are input into the decoding network to obtain the entity label sequence; Based on entity label sequences, content containing corresponding entities is filtered from the original text data and assembled into a target sentence set, which serves as the basic corpus for relation extraction.

[0010] Furthermore, methods for extracting authentication relations using named entity relation extraction based on joint learning, and entity relation extraction combining deep learning and remote supervision, include: Weakly labeled data in the authentication corpus are correlated and matched with social engineering relationship resource database to generate weakly labeled corpus. Segmented convolutional neural network is used to filter noise in the weakly labeled corpus to obtain candidate triples. Candidate triples are associated with verifiable authentication-type basic corpora to obtain hierarchical information text; the hierarchical information text is then input into a joint model of bidirectional neural network and tree-structured neural network for joint classification to obtain corrected triples; the hierarchical information includes organizations, departments, and personnel; Integrate candidate triples and modified triples to obtain the target triple; Methods for extracting connectivity relationships using entity relation extraction based on deep learning, and entity relation extraction combining deep learning and remote supervision, include: The basic corpus of communication categories containing communication relationships is input into the neural network as a high-quality labeled corpus for feature extraction to obtain a basic relation library; Weakly labeled corpora are generated by matching basic corpora without connectivity relationships with basic relation databases, and noise in the weakly labeled corpora is filtered out using a segmented convolutional neural network to obtain enhanced triples. Integrate the basic relational database with enhanced triples to obtain the target triples; Methods for social relation extraction that employ text semantic enhancement based on contrastive learning for joint entity relation extraction include: Contrast samples are generated from basic social corpora and input into a pre-trained encoder. The encoder parameters are optimized through a contrastive learning module to obtain enhanced semantic representations. Based on enhanced semantic representation, candidate entities and candidate relation tuples are generated by the Span method. The candidate entities and candidate relation tuples are then input into an end-to-end joint neural network model for joint classification, and the target triplet is output. Methods for extracting business relationships include joint learning-based named entity relation extraction and text semantic enhancement-based entity relation joint extraction based on contrastive learning. The basic business corpus is input into an end-to-end joint neural network model for joint classification to obtain a basic business relationship database. Extract texts containing ambiguous terms from the basic business corpus, and construct contrast samples with entities in the basic business relation database as the core. Obtain enhanced semantic representations through contrastive learning. Input the enhanced semantic representations into the joint extraction sub-model to correct errors in the basic business relation database and supplement new relations. Integrate the basic business relationship database with the corrected relationship data to output the target triplet.

[0011] Furthermore, the deep learning-based entity relation extraction method includes: The entity relations of the text are annotated to obtain entity relation annotated corpus; the methods for annotating entity relations include manual annotation and generating relation tags by matching with a remote knowledge base; The entity relation annotation corpus is processed using a neural network model to construct multi-dimensional feature vectors that represent the semantic information of words; the feature vectors include word vectors, position vectors, and grammatical relation vectors. The feature vectors are input into the neural network for feature extraction; weights are calculated on the extracted features, and the possible relationship types between entities are determined based on the semantic features of the weights, and entity relationship pairs are output. The named entity relation extraction method based on joint learning includes: An end-to-end joint neural network model is constructed, which integrates bidirectional neural network and tree structure neural network modules. The bidirectional neural network is used for entity recognition tasks, and the tree structure neural network is used for relation extraction tasks. The text is segmented into continuous text fragments as candidate entities; the candidate entities are then combined in pairs to form candidate relation tuples. The candidate entity and candidate relation tuples are input into the end-to-end joint neural network model. The candidate entity and candidate relation tuples are jointly classified by the bidirectional neural network and the tree structure neural network, and the relation type is output. The entity relation extraction method combining deep learning and remote supervision includes: The text is matched with an existing knowledge base, and entity pairs in the text that are consistent with the existing knowledge base are labeled to generate weakly labeled corpus. For weakly labeled corpora, feature embeddings that fuse word vectors and position vectors are constructed, and the two types of vectors are combined to form the model input features; A segmented convolutional neural network is used to divide the model input features of entity pairs into three segments according to the entity position, and pooling operations are performed on each segment to obtain sentence feature vectors. The feature vectors of sentences containing the same pair of entities are packaged together, and a multi-instance learning method is used to select the sentence with the highest semantic relation probability within the package as the core indicator sentence of the entity pair, thereby obtaining the relation type of the entity pair. The text semantic enhancement entity relation joint extraction method based on contrastive learning includes: The same text and entity context are repeatedly input into the pre-trained language model. Text embeddings with similar semantics but different expressions are generated through standard random masks to obtain positive sample pairs. The entity contexts of other texts in the same training batch are selected and combined with the entity context of the current text to form differential sample pairs to obtain negative sample pairs. By inputting positive and negative sample pairs into the contrast learning module, the semantic representations of positive sample pairs become more similar, while the semantic representations of negative sample pairs become more distant, thus obtaining an enhanced semantic representation. The enhanced semantic representation is segmented into continuous text fragments as candidate entities; candidate entities are combined in pairs to form candidate relation tuples; candidate entities and candidate relation tuples are input into an end-to-end joint neural network model for joint classification, and the relation type is output.

[0012] Furthermore, the step of constructing a trust relationship knowledge graph based on triples of various types includes: The target triples of various types are aggregated into a unified database, and knowledge fusion is performed through entity alignment and relation deduplication; the fused triples are converted into graph structure data and stored in the graph database.

[0013] Secondly, this application provides a personnel relationship shaping device based on intelligent entity relationship extraction, comprising: The resource classification module is used to acquire the core resources upon which trust relationships are built; it classifies the core resources and determines the information extraction rules for each type of core resource; the types of core resources include authentication relationships, communication relationships, social relationships, and business relationships; The entity recognition module is used to collect raw text data from core resources of various types according to information extraction rules, and to perform named entity recognition on the raw text data based on a deep learning model to obtain the basic corpus of raw text data of various types. The relation extraction module is used to extract target triples from various types of basic corpora using individual or combined entity relation extraction methods. Specifically, authentication relations are extracted using named entity relation extraction based on joint learning, as well as entity relation extraction combining deep learning and remote supervision; connectivity relations are extracted using entity relation extraction based on deep learning, as well as entity relation extraction combining deep learning and remote supervision; social relations are extracted using joint entity relation extraction based on contrastive learning and text semantic enhancement; and business relations are extracted using named entity relation extraction based on joint learning, as well as joint entity relation extraction based on contrastive learning and text semantic enhancement. The results output module is used to construct a trust relationship knowledge graph based on triples of various types of target triples.

[0014] Thirdly, this application provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the personnel relationship shaping method based on intelligent entity relationship extraction.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the personnel relationship shaping method based on intelligent entity relationship extraction.

[0016] Fifthly, this application provides a computer program product, including a computer program / instructions, which, when executed by a processor, implement the steps of the personnel relationship shaping method based on intelligent entity relationship extraction.

[0017] As can be seen from the above technical solution, this application provides a method and apparatus for shaping personnel relationships based on intelligent entity relationship extraction. The network structure and personnel relationship shaping method based on intelligent entity relationship extraction can start from the trust relationship analysis and shaping needs of specific scenarios in different industry fields such as education, law, and catering. It can realize entity recognition of massive multi-source heterogeneous target network situation data such as Internet data, passwords, public software tools, proprietary software tools, and technical knowledge bases, analyze and extract the association relationship of trust entities such as authentication, communication, social, and business, and construct a trust relationship knowledge graph based on the relationship graph model to realize the comprehensive shaping of network structure relationships and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network situation data. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is one of the flowcharts illustrating the personnel relationship shaping method based on intelligent entity relationship extraction in the embodiments of this application; Figure 2 This is the second flowchart illustrating the personnel relationship shaping method based on intelligent entity relationship extraction in the embodiments of this application; Figure 3 This is the third flowchart illustrating the personnel relationship shaping method based on intelligent entity relationship extraction in the embodiments of this application; Figure 4 This is a structural diagram of the personnel relationship shaping device based on intelligent entity relationship extraction in the embodiments of this application; Figure 5 This is a schematic diagram of the structure of the electronic device in the embodiments of this application.

[0020] Figure label: Electronic device 9600, central processing unit 9100, memory 9140, communication module 9110, input unit 9120, audio processor 9130, display 9160, power supply 9170, buffer memory 9141, application / function storage unit 9142, data storage unit 9143, driver storage unit 9144, antenna 9111, speaker 9131, microphone 9132. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The acquisition, storage, use, and processing of data in this application all comply with the relevant provisions of national laws and regulations.

[0023] Current entity relationship data in various scenarios such as online public opinion, social networks, and enterprise organizational structures are characterized by wide sources, strong heterogeneity, high degree of unstructuredness, and frequent dynamic changes. This patent addresses the problems of low automation and low accuracy in current network structure and personnel relationship shaping by proposing a method for shaping network structure and personnel relationships based on intelligent entity relationship extraction. It utilizes key string analysis and machine learning algorithms, primarily based on natural language processing, to analyze and calculate full-dimensional target data, extracting entity and relationship elements between nodes. It shapes and analyzes trust relationships, including authentication relationships, communication relationships, social relationships, and business relationships, and constructs a trust relationship graph using knowledge graph techniques, providing users with a target trust relationship network including network structure and personnel relationships.

[0024] Specifically, considering the problems existing in the prior art, this application provides a method and apparatus for shaping personnel relationships based on intelligent entity relationship extraction. The network structure and personnel relationship shaping method based on intelligent entity relationship extraction can start from the trust relationship analysis and shaping needs of specific scenarios in different industry fields such as education, law, and catering. It can realize entity recognition of massive multi-source heterogeneous target network situation data such as Internet data, passwords, public software tools, proprietary software tools, and technical knowledge bases, analyze and extract the association relationship of trust entities such as authentication, communication, social, and business, and construct a trust relationship knowledge graph based on the relationship graph model to achieve comprehensive shaping of network structure relationships and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network situation data.

[0025] To achieve comprehensive shaping of network structure and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network intelligence data, this application provides an embodiment of a personnel relationship shaping method based on intelligent entity relationship extraction. See [link to embodiment]. Figure 1 The method for shaping personnel relationships based on intelligent entity relationship extraction specifically includes the following: Step S101: Obtain the core resources upon which the trust relationship is built; classify the core resources and determine the information extraction rules for each type of core resource; the types of core resources include authentication relationships, communication relationships, social relationships, and business relationships.

[0026] Optionally, in this embodiment, specifying the core resource types and information extraction rules upon which the trust relationship is built includes: 1) Extraction of core resource information for authentication relationships: Core resource types: personnel attendance records, organizational rosters, address books, etc.

[0027] Information extraction rules: Grouping data from different organizations using organizational tag data from the social worker relationship resource database; Further grouping based on the data tags of each organization to obtain data at the departmental level; Based on information related to personnel's work and business, determine the authentication relationships of personnel at the resource access level; At the network level, the system obtains the list of registered users, access permissions, and historical records through the website application user database, permission database, resource access logs, and network domain / workgroup records.

[0028] 2) Extraction of core resource information related to communication relationships: Core resource types: Contacts, call logs, SMS messages, emails, etc.

[0029] Information extraction rules: Extract communication relationships between people from communication records; The degree of intimacy in communication can be extracted by information such as call frequency, SMS sending and receiving frequency, and email sending and receiving frequency.

[0030] 3) Extraction of core resource information from social relationships Core resource types: household registration information or contact information, social groups, etc.

[0031] Information extraction rules: Extract social relationship information from notes; Use social group data to identify and classify different types of social relationships.

[0032] 4) Extraction of core resource information related to business relationships Core resource types: internal organizational roster, address book, business documents, API call logs, code files, etc.

[0033] Information extraction rules: Business relationships can be obtained through communication and interaction records, business transaction records, and program releases. Analyze entity relationships in business processes based on business documents and API call logs.

[0034] Step S102: Collect raw text data from each type of core resource according to the information extraction rules, and perform named entity recognition on the raw text data based on the deep learning model to obtain the entity label sequence of each type of raw text data.

[0035] Optionally, in this embodiment, performing Named Entity Recognition (NER) on multi-source network sentiment data based on a deep learning model includes: 1) Define the deep learning model used in the NER task. Pre-trained Language Models (PLMs) are deep language models that are pre-trained on large-scale unlabeled text corpora, enabling them to learn rich semantic, syntactic, and contextual information. This patent employs a PLM-based method for dynamically training word vectors, which allows for better vectorized representation of text. This enables the use of a feature extractor to extract text features, which are then used in a decoder to obtain predicted sequence labels.

[0036] 2) Basic process of NER task The input text is vectorized based on static word vectors or dynamic PLM, specifically based on character or word, or by a hybrid approach that combines information from both methods. The IR stage requires the effective fusion of word and character information and can also assist in feature engineering used by statistical machine learning methods; The Context Encoder (CE) layer or sequence modeling layer uses a feature extractor to further extract text features from the vectorized text output by the IR stage; The tag decoder (TD) layer takes the vector output from the CE layer and inputs it into the decoding network to obtain the optimal sequence name tag.

[0037] Based on the identified entity label sequence, content containing the corresponding entity is selected from the original data corpus and assembled into a target sentence set, which serves as the basic corpus for relation extraction.

[0038] Step S103: Extract target triples from the basic corpora of each type using entity relation extraction methods, either individually or in combination. Specifically, authentication relations are extracted using named entity relation extraction based on joint learning, and entity relation extraction combining deep learning and remote supervision; connectivity relations are extracted using entity relation extraction based on deep learning, and entity relation extraction combining deep learning and remote supervision; social relations are extracted using joint entity relation extraction based on contrastive learning and text semantic enhancement; and business relations are extracted using named entity relation extraction based on joint learning, and entity relation extraction based on contrastive learning and text semantic enhancement.

[0039] Optionally, in this embodiment, different entity relationship extraction methods are set for each type of core resource and / or entity list, specifically: Scenario 1: Extraction of core resource information for authentication relationships Named entity relation extraction based on joint learning + entity relation extraction combining deep learning and remote supervision 1. Explanation of Adapting to Joint Learning: This scenario requires handling three types of strongly related entities and relationships simultaneously: "organizational structure-department," "personnel-department," and "personnel-access permissions," and also presents issues such as "fuzzy departmental boundaries" and "tight binding of permissions to personnel." The Span method of joint learning can accurately locate the scope of entities (such as organizational structure, department, personnel, and permission type). The end-to-end model can avoid the cascading error of "identifying personnel first and then determining permissions." At the same time, the combination of bidirectional LSTM and tree-structured LSTM can capture the hierarchical dependency relationship of "organization-department-personnel," which fits the rule of "grouping data acquisition" in the scenario.

[0040] 2. Explanation of Remote Supervision Adaptation: The scenario involves massive amounts of weakly labeled data such as website application user databases and permission databases (e.g., access logs without manually labeled permission relationships). Remote supervision does not require extensive manual labeling; noisy data in the logs (e.g., invalid access records) can be processed using a PCNN model. The multi-instance attention mechanism can assign high weights to "valid access logs," accurately extracting the "person-permission" authentication relationship, matching the scenario's requirement to "obtain access permissions and historical records."

[0041] Using logic: First, a remotely supervised PCNN model is used to preprocess massive access logs and permission databases to extract candidate authentication relationships in batches. Then, the joint learning Span method is used to refine the extraction of high-quality data such as organizational rosters and attendance records, correcting the noise error of remote supervision, and finally outputting structured triples such as <personnel, authentication permission, resource> and <department, affiliation, organization>.

[0042] Scenario 2: Extraction of core resource information of communication relationships Entity Relation Extraction Based on Deep Learning + Entity Relation Extraction Combining Deep Learning and Distant Supervision 1. Explanation of Adaptation to Basic Deep Learning Methods: The core scenario involves extracting "person-to-person" communication relationships and the intimacy derived from "communication frequency." The relationship type is singular (only communication relationships), and the data, such as address books and call logs, is highly structured (e.g., call logs contain explicit caller / callee numbers and times). Basic deep learning methods using word vectors and location vectors can quickly capture the semantic relationships of entities (persons), making them suitable for handling these basic relationships with low annotation costs and clear semantics.

[0043] 2. Explanation of Adaptation to Remote Supervision Methods: Communication records (especially SMS and email records) are extremely large in volume, and contain a large number of unlabeled redundant records (such as spam SMS). The multi-instance method of remote supervision can treat "all communication records of the same person pair" as a package, and filter out the effective records with the highest semantic probability; the attention mechanism can assign weights according to the communication frequency, with higher frequency records having higher weights, directly supporting the calculation of "communication intimacy", which fits the core requirement of "extracting intimacy through frequency" in the scenario rules.

[0044] Using logic: Basic deep learning methods are used to process labeled address books and core call records to quickly obtain basic communication relationships; a remotely supervised PCNN model is used to process massive amounts of unlabeled SMS and email records, high-value records are filtered through weight allocation, and intimacy is calculated by combining frequency statistics, finally outputting a structured result of <person A, communication (intimacy: high / medium / low), person B>.

[0045] Scenario 3: Extraction of core resource information from social relationships Joint entity relation extraction based on contrastive learning and text semantic enhancement 1. Explanation of Adapting to Contrastive Learning: The core pain point of this scenario is "semantic ambiguity"—social relationships rely on annotation information (such as "Brother Wang" or "Sister Li") and the context of social groups. Different annotations may correspond to the same relationship (such as "husband" or "Mr." both indicating a spousal relationship), while the same annotation may correspond to different relationships (such as "General Manager Zhang" potentially indicating a superior or a client). Contrastive learning enhances semantic understanding through positive sample pairs (different expressions of the same annotation) and negative sample pairs (the context of different annotations), addressing the problem of insufficient attention to ambiguous semantics in pre-trained models and accurately distinguishing social relationship types.

[0046] 2. The auxiliary value of joint extraction: The method itself integrates the logic of joint extraction, which can simultaneously complete "personal entity recognition" and "social relationship classification", avoiding errors caused by multiple nicknames and unclear entity boundaries in the group, and adapting to the need of "identifying and classifying social relationships from social groups" in the scenario.

[0047] Using logic: Input household registration information and social group chat records into the contrastive learning model. Generate positive samples (such as different expressions of "Brother Wang - Little Wang") and negative samples (such as the context of "Brother Wang - Sister Li") through random masking. After enhancing the semantic representation, automatically identify personnel entities and classify social relationships (such as relatives, colleagues, and friends), and output triples of <person A, relatives / colleagues / friends, person B>.

[0048] Scenario 4: Extraction of core resource information related to business relationships Joint entity relation extraction based on joint learning + text semantic enhancement based on contrastive learning 1. Explanation of Adaptation to Federation Learning: The scenario involves multiple types of entities (organizations, personnel, business interfaces, code modules) and complex relationships (business transactions, business process dependencies, interface call relationships), and entities and relationships are highly bound (e.g., in an "interface call" relationship, both the caller and the callee entities must be clearly defined). The Span method of federated learning can accurately locate entity boundaries in business documents and code files, and the end-to-end model can efficiently handle "entity relationships in business processes," alleviate cascading errors, and adapt to complex scenarios with "multiple entities and multiple relationships."

[0049] 2. Supplementary explanation of the contrastive learning method: Business documents (such as contracts and requirements documents) and interface logs contain a large amount of ambiguous text (such as the same business term having different meanings in different documents, or similar interface names with different functions). Contrastive learning can enhance the breadth of semantics of text, distinguish the differences between similar business entities by comparing positive and negative samples, avoid misjudgments of relationships caused by ambiguous terms, and meet the accuracy requirements of "analyzing the relationships between business process entities" in the scenario.

[0050] Using logic: First, a joint learning model is used to process structured data such as organizational rosters and API call logs, extracting basic business relationships such as "business transactions" and "API calls". Then, a contrastive learning model is used to process unstructured data such as business documents and code files, enhancing semantic representation to resolve terminology ambiguity, correcting and supplementing business relationships, and finally outputting multi-dimensional structured results such as <Company A, business transactions, Company B> and <API A, call dependency, API B>.

[0051] Optionally, in this embodiment, the usage flow of combining the methods in the four scenarios includes: Scenario 1: Extraction of core resource information for authentication relationships (joint learning + remote supervision) Core logic: First, use remote supervision to batch process massive amounts of weakly labeled data, and then use joint learning to refine the data, achieving a dual improvement in "efficiency + accuracy". The specific process is as follows: 1. First stage: Remote supervision of batch extraction of candidate relations (processing weakly labeled data) Data preparation: Collect weakly labeled data, such as basic network-level authentication corpora such as website application user databases, permission databases, and resource access logs, and use the social engineering relation resource database as a remote knowledge base.

[0052] 2. Remote supervision and annotation: Match the "personnel-resource access records" in the logs with the "personnel-permission" association data in the knowledge base, automatically annotate preliminary relationship tags such as "personnel-access permissions" and "personnel-department", and generate weakly annotated corpus (including noise such as invalid access records).

[0053] 3. PCNN Model Processing: Construct a fusion feature of word vectors (person ID, resource name) and location vectors (access time, resource path location), input it into the PCNN model, segment the sentence by entity pair and perform segmented pooling, and filter out some noisy data.

[0054] 4. Attention Weight Allocation: Treat all access records of the same person as a "package". Through a multi-instance attention mechanism, assign high weight to logs that "successfully access and meet the permission rules", filter out core records, and output candidate triples such as <person, candidate access permission, resource> and <person, candidate affiliation, department> in batches.

[0055] 5. Second stage: Joint learning fine-tuning (processing high-quality data) High-quality data input: collect high-quality data, including organizational rosters, personnel attendance records and other manually verifiable authentication-type basic corpora, and associate them with candidate triples output by remote supervision.

[0056] 6. Joint Model Training and Entity Localization: Construct a joint bidirectional LSTM + tree-structured LSTM model. Input text that integrates hierarchical information of "organization-department-personnel". Use the Span method to accurately locate the boundaries of entities such as organizational structure, department, personnel, and permission type (solving the problem of "fuzzy department boundaries").

[0057] 7. Joint Classification and Error Correction: The model classifies both entity and relation types simultaneously, using a tree-based LSTM to capture the hierarchical dependencies of "organization-department-personnel" and correcting errors in remote supervision (such as mislabeling "temporary access" as "permanent permission"). A bias loss function is used to strengthen the effective association between "personnel and permissions" while ignoring interference from invalid access records.

[0058] 8. Third stage: Results fusion and output: Integrate the results of the two stages and output the final structured triple, such as <Zhang San, permanent access, financial system> <R&D department, affiliated, XX Technology Company> <Li Si, temporary access, project documents>, to support resource access authentication decisions.

[0059] Scenario 2: Extraction of core resource information of communication relationships (basic deep learning + remote supervision) Core logic: Basic methods are used to quickly build a database from structured labeled data, while remote supervision is used to supplement and improve the database from massive amounts of unstructured data. Frequency weighting is then used to calculate affinity. The process is as follows: 1. First stage: Building a basic relational database using basic deep learning (processing labeled data) Preparation of labeled data: Select a portion of the address book (including "name-phone number" association), core call records (manually labeled "caller-called" relationship) and other basic communication data as high-quality labeled data.

[0060] 2. Feature Construction and Model Training: Word vectors for "name" and "phone number" are generated using word2vec. These are then combined with the position vectors for "call time" and "call duration" to construct feature vectors. These vectors are input into an LSTM neural network, where weighted features are extracted using softmax, and the entity relationship extraction model is trained.

[0061] 3. Basic Relationship Output: The model quickly identifies the communication relationships between "persons" and outputs basic triples such as <Zhang San, communication, Li Si> and <Wang Wu, communication, Zhao Liu> to build an initial communication relationship library.

[0062] 4. Second stage: Remote supervision and supplementation and intimacy calculation (processing massive data) Weakly labeled data processing: Collect massive amounts of basic communication data such as text messages and email records (without manual annotation), use the basic relationship database of the first stage as a remote knowledge base, automatically label "person pairs with matching mobile phone numbers" as communication relationships, and generate weakly labeled data (including redundant data such as spam text messages).

[0063] 5. PCNN and Attention Mechanism Filtering: Construct word vectors and position vectors for "phone number-SMS content", input them into a PCNN model for segmented pooling, and filter out noise such as spam SMS messages. Treat all communication records of the same person pair as "packets", and assign weights according to "call / SMS / email frequency" through an attention mechanism—the higher the frequency, the greater the weight.

[0064] 6. Intimacy Level Classification: Calculate the frequency of high-weight records within the package, set thresholds (e.g., ≥30 times per month is high, 10-29 times is medium, <10 times is low), and generate enhanced triples such as <Person A, Contact (Intimacy: High), Person B>.

[0065] 7. Third stage: Results integration: The basic relationship database and the enhanced ternary database are merged, duplicates are removed, and the final output is structured data of communication relationships containing intimacy information to support social relationship analysis.

[0066] Scenario 3: Extraction of core resource information from social relationships (contrastive learning + collaborative learning) Core logic: Based on contrastive learning to enhance semantics, a joint extraction process is embedded to address semantic ambiguity. The process is as follows: 1. First stage: Data preprocessing and sample construction Data collection: Integrate basic communication data such as household registration information, address book (including notes such as "Brother Wang" and "Sister Li"), and social group chat records, and clean up invalid characters (such as emoticons and garbled characters).

[0067] 2. Entity Context Extraction: For each text, extract the person entity and its context (e.g., in “Brother Wang brought food for Xiao Wang yesterday”, the context of “Brother Wang” and “Xiao Wang” is “brought food yesterday”).

[0068] 3. Comparison Sample Generation: Generate positive sample pairs by randomly masking the context of the same entity (e.g., changing "yesterday help" to "_day help"); select the context of other text entities (e.g., the context of "Li Jie" in "Li Jie asked for leave today") as negative sample pairs.

[0069] 4. Second stage: Contrastive learning semantic enhancement encoder optimization: Input positive and negative sample pairs into the pre-trained encoder (such as BERT), calculate the semantic similarity loss through the contrastive learning module, optimize the encoder parameters, so that the model can distinguish the semantic difference between "Brother Wang-Xiao Wang" and "Brother Wang-Sister Li", as well as the semantic association between "husband" and "sir".

[0070] 5. Enhanced semantic output: Output entities and context with enhanced semantic representations, providing a precise semantic foundation for subsequent joint extraction.

[0071] 6. Third stage: Joint extraction and relation classification of candidate entities and tuple generation: Based on enhanced semantic representation, candidate entities (such as "Brother Wang", "Little Wang", "Sister Li") and Span pairs (such as <Brother Wang, Little Wang> <Brother Wang, Sister Li>) are generated using the Span method.

[0072] 7. Joint classification: The end-to-end model simultaneously identifies entity types (all of which are people) and relationship types, and combines household registration information for verification to output triples such as <Brother Wang, relative, Xiao Wang> and <Brother Wang, colleague, Sister Li>.

[0073] 8. Fourth stage: Results verification and output: Further verify the relationship types by combining social group types (such as family groups, company groups), and finally output accurate structured data of social relationships.

[0074] Scenario 4: Extraction of core resource information related to business relationships (joint learning + comparative learning) Core logic: Joint learning lays the foundation for processing structured data, while contrastive learning refines the details of unstructured data, resolving the "multiple entities + ambiguity" problem. The process is as follows: 1. First stage: Joint learning to extract basic business relationships (processing structured data) Structured data input: collect basic business data such as internal organizational roster and interface call logs (including explicit records of "interface A-call-interface B"), and label some samples as training set.

[0075] 2. Joint Model Training and Entity Localization: Construct a joint model of bidirectional LSTM and tree-structured LSTM, and accurately locate entities such as organizations, personnel, business interfaces, and code modules using the Span method (e.g., identify “Payment Interface V1” and “Order Interface V2” as interface entities from interface logs).

[0076] 3. Basic Relationship Output: The model simultaneously classifies entity relationships and outputs basic business relationship triples such as <XX Company, Business Transactions, YY Company> and <Payment Interface V1, Call Dependency, Order Interface V2>, thus constructing an initial business relationship library.

[0077] 4. Second stage: Contrastive learning to enhance semantics and correct relationships (processing unstructured data) Unstructured data processing: collect basic business corpora such as business documents (such as contracts, requirements documents), code files, etc., and extract text containing ambiguous terms (such as "payment interface" pointing to "payment interface V1" in the contract and "payment interface V2" in the requirements document).

[0078] 5. Contrast Sample Construction and Semantic Enhancement: Using entities in the basic business relational database as the core, contrast samples are constructed—positive samples are different expressions of the same interface (e.g., "Payment Interface V1" and "V1 Version Payment Interface"), and negative samples are expressions of similar interfaces (e.g., "Payment Interface V1" and "Payment Interface V2"). Semantic representation is enhanced through contrastive learning to distinguish differences between similar terms.

[0079] 6. Relationship Correction and Supplementation: Enhance the semantic input joint extraction sub-model, correct errors in the basic relation library (e.g., correct "Payment Interface V1-Call-Order Interface V3" to "Payment Interface V2-Call-Order Interface V3"), and supplement new relations (e.g., extract <XX Company, Cooperative Development, ZZ Company> from the contract).

[0080] 7. Third stage: Results integration and output: Integrate the basic relational database with the corrected relational data, classify the relational types according to "business transactions", "interface calls", "collaborative development", etc., and output multi-dimensional business relational structured triples to support business process analysis.

[0081] Optionally, in this embodiment, the entity relation extraction method includes: 1) Entity Relation Extraction Based on Deep Learning (Reference) Figure 2 ) This method uses "data annotation - feature representation - model training - relation output" as its core link, and relies on neural networks to achieve end-to-end relation extraction. The complete process is as follows: 1. Corpus preparation stage: Labeled text corpus is obtained in two ways: first, manual annotation by professionals to manually mark entity relationships in the text to form a high-quality annotation set; second, automatic alignment with remote knowledge base to match entities in the text with entities in the knowledge base to indirectly generate relationship tags and reduce manual costs.

[0082] 2. Feature Representation Stage: The word2vec model is used to process the labeled corpus, constructing multi-dimensional feature vectors to represent the semantic information of words. Specifically, this includes: ① word vectors, capturing the semantic connotation of the words themselves; ② position vectors, identifying the relative positional relationship between words and target entities, strengthening entity association features; ③ grammatical relation vectors, incorporating the grammatical roles of words in sentences (such as subject-verb, verb-object), enriching the semantic dimension. The three types of vectors are fused to form the final feature vector, which serves as the input to the neural network.

[0083] 3. Model Training and Feature Extraction Stage: The feature vectors are input into a pre-defined neural network (such as CNN, LSTM, etc.) for feature extraction, uncovering hidden entity relationship patterns in the text. The extracted features are processed by the softmax activation function, and the weights for different relationship types are calculated to highlight semantic features strongly correlated with entity relationships.

[0084] 4. Result Output Stage: Based on the weight distribution output by softmax, determine the relationship type between entities, and finally output the structured entity relationship pair <Entity 1, Relationship Type, Entity 2>.

[0085] 2) Named entity relation extraction based on joint learning method The core of this method is to "jointly complete entity recognition and relationship classification using a single model" to avoid cascading errors. This patent adopts a Span-based sequence labeling approach, and the complete process is as follows: 1. Model Architecture Construction Phase: An end-to-end joint neural network model is constructed, integrating bidirectional LSTM and tree-structured LSTM modules. The bidirectional LSTM is responsible for capturing sequence dependency information in the text for entity recognition tasks; the tree-structured LSTM is responsible for parsing the dependency tree substructure of the text and mining semantic relationships between entities for relation extraction tasks. Simultaneously, a bias loss function is designed to enhance the relation weights of valid entity pairs and ignore the interference of invalid entity labels.

[0086] 2. Candidate Entity and Relational Tuple Generation Stage: The input text is processed using a Span-based method—① The text is divided into continuous character segments (i.e., text Spans) as candidate entity sets, covering all possible entity boundaries; ② Candidate entities are paired to form Span pairs as candidate relational tuples, ensuring that no potential entity associations are missed.

[0087] 3. Joint Training and Classification Stage: Candidate entities (Spans) and Span pairs are simultaneously input into the model. A bidirectional LSTM identifies and classifies the entity Spans, outputting the entity type (e.g., personnel, departments, organizations). Its output word embeddings and label embeddings are used as input to a tree-structured LSTM, achieving parameter sharing. The tree-structured LSTM classifies the Span pairs according to relationships, outputting the relationship type (e.g., "belonging" or "call dependency"). The model parameters are optimized using a bias loss function to improve the accuracy of joint classification.

[0088] 4. Result optimization stage: Combining entity preprocessing (filtering invalid candidate entities) and pre-sampling (strengthening training samples of high-frequency valid entity pairs) strategies, the model performance is further improved, and the final output is a structured triple that contains both entities and their corresponding relationships.

[0089] 3) Entity Relation Extraction Combining Deep Learning and Remote Supervision This method, centered on "remote supervision for data supplementation - PCNN noise reduction - attention mechanism for accuracy improvement," addresses the problem of scarce labeled data. The complete process is as follows: 1. Remote supervised corpus construction stage: Introducing a remote knowledge base (such as a social worker relationship resource database or enterprise information database), matching entity pairs in the text with known relation entity pairs in the knowledge base. If the text entity pairs match the knowledge base entity pairs, the corresponding relation labels are automatically assigned to the text, generating a large-scale weakly labeled corpus. This process requires no manual intervention, but may introduce noise (such as discrepancies between the context of entity pairs in the text and their relation in the knowledge base).

[0090] 2. Feature Vector Construction Stage: For weakly labeled corpora, a feature embedding that fuses word vectors and position vectors is constructed. Word vectors capture the semantics of words, while position vectors identify the location information of entities in sentences. The combination of these two types of vectors forms the input features of the model.

[0091] 3. PCNN Model Feature Extraction Stage: PCNN (Piecewise Convolutional Neural Network) is used to process the feature vectors. The sentence containing the entity pair is divided into three segments according to the entity position, and pooling operations are performed on each segment. This avoids the excessive neglect of hidden layer nodes in traditional CNNs, retains more contextual information of entity pairs, effectively reduces the interference of noisy data, and improves the targeting of feature extraction.

[0092] 4. Multi-instance attention mechanism optimization stage: All sentences containing the same entity pair are treated as a "bag". A multi-instance learning method is used to select the sentence with the highest semantic relationship probability within the bag as the core indicator sentence for that entity pair. At the same time, an attention mechanism is introduced to assign weights to sentences within the bag according to the degree of association between sentences and entity relationships—sentences with a high degree of matching with the correct relationship label have higher weights and greater contribution rates; sentences associated with incorrect labels have lower weights and smaller contribution rates.

[0093] 5. Output Stage: Combining the feature output of PCNN with the weight allocation of the attention mechanism, the classifier determines the relationship type between entities and outputs accurate entity relationship triples.

[0094] 4) Joint entity relation extraction based on contrastive learning for text semantic enhancement (see reference) Figure 3 ) This method, based on joint extraction, addresses the problem of insufficient contextual attention through "contrastive learning to enhance semantics." The complete process is as follows: 1. Pre-trained Encoder and Sample Construction Stage: A pre-trained language model (such as BERT) is introduced as the encoder, and the input text is used. Constructing contrastive learning samples—① Positive sample pairs: The same text and entity context are repeatedly input into the encoder, and text embeddings with similar semantics but different expressions are generated through standard random masking (masking some characters); ② Negative sample pairs: The entity context of other texts in the same training batch is selected and combined with the entity context of the current text to form sample pairs with significant semantic differences.

[0095] 2. Contrastive Learning Semantic Enhancement Stage: Positive and negative sample pairs are input into the contrastive learning module. By calculating the semantic similarity loss function between samples, the encoder parameters are optimized. This makes the semantic representations of positive sample pairs more similar and the semantic representations of negative sample pairs more dissimilar, thereby enhancing the model's ability to understand the semantic context of the text, especially improving its ability to distinguish ambiguous expressions.

[0096] 3. Joint Extraction Stage: The enhanced semantic representation is input into the joint extraction sub-model, and the Span method, similar to "named entity relation extraction based on joint learning", is adopted to generate candidate entity Span and Span pairs. The end-to-end model simultaneously completes entity recognition (e.g., recognizing "Brother Wang" and "Sister Li" as personnel entities) and relation classification (e.g., determining that "Brother Wang" and "Xiao Wang" are relatives).

[0097] 4. Output Stage: Filter out invalid entities and relations, and output accurate triples of <entity1, relation type, entity2> to solve the problem of misjudgment of relations caused by fuzzy semantics.

[0098] Step S104: Construct a trust relationship knowledge graph based on triples for each type of target triple.

[0099] Optionally, constructing a trust relationship knowledge graph based on triples includes: The target triples of various types are aggregated into a unified database, and entities pointing to the same real-world object are identified and merged using methods such as string similarity, alias library, and context comparison. Relationship deduplication and conflict resolution are also performed: identical relation triples are merged, and if the same pair of entities has a conflicting relationship (such as being labeled as both "colleague" and "relative"), a decision is made based on the confidence level of the data source or manual rules. After defining the type system of entities and relations in the knowledge graph, the merged triples are converted into graph structure data and stored in a dedicated graph database.

[0100] During the construction process, based on the identification of entities and the extraction of entity relationships from various multi-source heterogeneous target network information data, the relationship between entities is represented by a triplet in the form of <entity1, relation, entity2>, with each triplet representing a factual information. By viewing entities as nodes in a network and relationships as directed edges between entities, the aforementioned triples are used to construct a trust relationship data knowledge graph.

[0101] This trust relationship knowledge network can aggregate a large number of fact triple entries, while providing hierarchical relationships between entities and constraint information on the entity attributes at both ends of the relationship. It reveals email trust relationships with key industry figures with high-frequency communication, displays target network trust domains and subdomains, and provides multiple trusted social accounts and social relationships that have been registered for more than a year, have a large number of effective followers, and post frequently. It can provide flexible technical support and data support for subsequent trust relationship knowledge fusion and knowledge reasoning.

[0102] As described above, the personnel relationship shaping method based on intelligent entity relationship extraction provided in this application can start from the trust relationship analysis and shaping needs of specific scenarios in different industry fields such as education, law, and catering. It can realize entity recognition of massive multi-source heterogeneous target network situation data such as Internet data, passwords, public software tools, proprietary software tools, and technical knowledge bases, analyze and extract the association relationship of trust entities such as authentication, communication, social, and business, and construct a trust relationship knowledge graph based on the relationship graph model to realize the comprehensive shaping of network structure relationships and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network situation data.

[0103] To achieve comprehensive shaping of network structure and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network intelligence data, this application provides an embodiment of a personnel relationship shaping device based on intelligent entity relationship extraction, which implements all or part of the aforementioned personnel relationship shaping method based on intelligent entity relationship extraction. See [link to embodiment]. Figure 4 The personnel relationship shaping device based on intelligent entity relationship extraction specifically includes the following: Resource classification module 10 is used to obtain the core resources on which trust relationships are built; to classify the core resources and to determine the information extraction rules for each type of core resource; the types of core resources include authentication relationships, communication relationships, social relationships, and business relationships; The entity recognition module 20 is used to collect raw text data from core resources of various types according to information extraction rules, and to perform named entity recognition on the raw text data based on a deep learning model to obtain the basic corpus of raw text data of various types. The relation extraction module 30 is used to extract target triples from various types of basic corpora using entity relation extraction methods, either individually or in combination. Specifically, authentication relations employ named entity relation extraction based on joint learning, as well as entity relation extraction combining deep learning and remote supervision; connectivity relations employ entity relation extraction based on deep learning, as well as entity relation extraction combining deep learning and remote supervision; social relations employ joint entity relation extraction based on contrastive learning and text semantic enhancement; and business relations employ named entity relation extraction based on joint learning, as well as joint entity relation extraction based on contrastive learning and text semantic enhancement. The output module 40 is used to construct a trust relationship knowledge graph based on triples of various types of target triples.

[0104] As described above, the personnel relationship shaping device based on intelligent entity relationship extraction provided in this application embodiment can start from the trust relationship analysis and shaping needs of specific scenarios in different industry fields such as education, law, and catering. It can realize entity recognition of massive multi-source heterogeneous target network situation data such as Internet data, passwords, public software tools, proprietary software tools, and technical knowledge bases, analyze and extract the association relationship of trust entities such as authentication, communication, social, and business, and construct a trust relationship knowledge graph based on the relationship graph model to realize the comprehensive shaping of network structure relationships and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network situation data.

[0105] From a hardware perspective, in order to achieve comprehensive shaping of network structure relationships and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network sentiment data, this application provides an embodiment of an electronic device for implementing all or part of the personnel relationship shaping method based on intelligent entity relationship extraction. The electronic device specifically includes the following components: The system comprises a processor, memory, a communications interface, and a bus; wherein the processor, memory, and communications interface communicate with each other via the bus; the communications interface is used to realize information transmission between the personnel relationship shaping device based on intelligent entity relationship extraction and core business systems, user terminals, and related databases and other related devices; the logic controller can be a desktop computer, tablet computer, or mobile terminal, etc., and this embodiment is not limited to these. In this embodiment, the logic controller can be implemented with reference to the embodiments of the personnel relationship shaping method based on intelligent entity relationship extraction and the embodiments of the personnel relationship shaping device based on intelligent entity relationship extraction in the embodiments, the contents of which are incorporated herein, and repeated details will not be described again.

[0106] It is understood that the user terminal may include smartphones, tablet computers, network set-top boxes, portable computers, desktop computers, personal digital assistants (PDAs), in-vehicle devices, smart wearable devices, etc. Among these, the smart wearable devices may include smart glasses, smartwatches, smart bracelets, etc.

[0107] In practical applications, parts of the personnel relationship shaping method based on intelligent entity relationship extraction can be executed on the electronic device side as described above, or all operations can be completed in the client device. The choice can be made based on the processing power of the client device and the limitations of the user's usage scenario. This application does not impose any limitations on this. If all operations are completed in the client device, the client device may further include a processor.

[0108] The aforementioned client device may have a communication module (i.e., a communication unit) that can communicate with a remote server to achieve data transmission with the server. The server may include a server on the task scheduling center side; in other implementation scenarios, it may also include a server on an intermediate platform, such as a server on a third-party server platform that has a communication link with the task scheduling center server. The server may include a single computer device, a server cluster consisting of multiple servers, or a distributed server structure.

[0109] Figure 5This is a schematic block diagram illustrating the system configuration of the electronic device 9600 according to an embodiment of this application. Figure 5 As shown, the electronic device 9600 may include a central processing unit 9100 and a memory 9140; the memory 9140 is coupled to the central processing unit 9100. It is worth noting that... Figure 5 This is an example; other types of structures can also be used to supplement or replace this structure to achieve telecommunications functions or other functions.

[0110] In one embodiment, the personnel relationship shaping method based on intelligent entity relationship extraction can be integrated into the central processing unit 9100. The central processing unit 9100 can be configured to perform the following control: Step S101: Obtain the core resources upon which the building of trust relationships depends; classify the core resources and determine the information extraction rules for each type of core resource; the types of core resources include authentication relationships, communication relationships, social relationships, and business relationships; Step S102: Collect raw text data from each type of core resource according to the information extraction rules, and perform named entity recognition on the raw text data based on the deep learning model to obtain the basic corpus of each type of raw text data; Step S103: Extract target triples from the basic corpora of each type using entity relation extraction methods, either individually or in combination. Specifically, authentication relations are extracted using named entity relation extraction based on joint learning, and entity relation extraction combining deep learning and remote supervision; connectivity relations are extracted using entity relation extraction based on deep learning, and entity relation extraction combining deep learning and remote supervision; social relations are extracted using joint entity relation extraction based on contrastive learning and text semantic enhancement; and business relations are extracted using named entity relation extraction based on joint learning, and entity relation extraction based on contrastive learning and text semantic enhancement. Step S104: Construct a trust relationship knowledge graph based on triples for each type of target triple.

[0111] As can be seen from the above description, the electronic device provided in this application embodiment, based on the network structure and personnel relationship shaping method of intelligent entity relationship extraction, can start from the trust relationship analysis and shaping needs of specific scenarios in different industry fields such as education, law, and catering. It can realize entity recognition of massive multi-source heterogeneous target network situation data such as Internet data, passwords, public software tools, proprietary software tools, and technical knowledge bases, analyze and extract the association relationship of trust entities such as authentication, communication, social, and business, and construct a trust relationship knowledge graph based on the relationship graph model. It can realize the comprehensive shaping of network structure relationships and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network situation data.

[0112] In another embodiment, the personnel relationship shaping device based on intelligent entity relationship extraction can be configured separately from the central processing unit 9100. For example, the personnel relationship shaping device based on intelligent entity relationship extraction can be configured as a chip connected to the central processing unit 9100, and the personnel relationship shaping method function based on intelligent entity relationship extraction can be realized through the control of the central processing unit.

[0113] like Figure 5 As shown, the electronic device 9600 may further include: a communication module 9110, an input unit 9120, an audio processor 9130, a display 9160, and a power supply 9170. It is worth noting that the electronic device 9600 does not necessarily need to include these components. Figure 5 All components shown; in addition, the electronic device 9600 may also include Figure 5 For components not shown, please refer to existing technologies.

[0114] like Figure 5 As shown, the central processing unit 9100, sometimes also referred to as a controller or operating control, may include a microprocessor or other processor device and / or logic device, which receives inputs and controls the operation of various components of the electronic device 9600.

[0115] The memory 9140 may be, for example, one or more of a cache, flash memory, hard drive, removable media, volatile memory, non-volatile memory, or other suitable devices. It may store the aforementioned failure-related information, and also store a program for executing that information. The central processing unit 9100 may execute the program stored in the memory 9140 to perform information storage or processing, etc.

[0116] Input unit 9120 provides input to central processing unit 9100. Input unit 9120 may be, for example, a keypad or touch input device. Power supply 9170 provides power to electronic device 9600. Display 9160 displays images and text. Display may be, for example, an LCD display, but is not limited thereto.

[0117] The memory 9140 can be a solid-state memory, such as a read-only memory (ROM), random access memory (RAM), a SIM card, etc. It can also be a memory that retains information even when power is off, can be selectively erased, and contains more data; examples of this type of memory are sometimes referred to as EPROMs. The memory 9140 can also be some other type of device. The memory 9140 includes a buffer memory 9141 (sometimes referred to as a buffer). The memory 9140 may include an application / function storage unit 9142 for storing application programs and function programs or processes for executing the operation of the electronic device 9600 via the central processing unit 9100.

[0118] The memory 9140 may also include a data storage unit 9143 for storing data, such as contacts, digital data, pictures, sounds, and / or any other data used by the electronic device. The driver storage unit 9144 of the memory 9140 may include various drivers for the electronic device for communication functions and / or for performing other functions of the electronic device (such as messaging applications, address book applications, etc.).

[0119] The communication module 9110 is a transmitter / receiver that sends and receives signals via the antenna 9111. The communication module 9110 (transmitter / receiver) is coupled to the central processing unit 9100 to provide input signals and receive output signals, which is the same as in a conventional mobile communication terminal.

[0120] Based on different communication technologies, multiple communication modules 9110 can be configured in the same electronic device, such as cellular network modules, Bluetooth modules, and / or wireless LAN modules. The communication module 9110 (transmitter / receiver) is also coupled to a speaker 9131 and a microphone 9132 via an audio processor 9130 to provide audio output via the speaker 9131 and receive audio input from the microphone 9132, thereby realizing typical telecommunications functions. The audio processor 9130 may include any suitable buffer, decoder, amplifier, etc. Additionally, the audio processor 9130 is coupled to a central processing unit 9100, enabling on-device recording via the microphone 9132 and on-device playback of stored audio via the speaker 9131.

[0121] Embodiments of this application also provide a computer-readable storage medium capable of implementing all steps of the personnel relationship shaping method based on intelligent entity relationship extraction, where the execution subject is a server or client, as described in the above embodiments. The computer-readable storage medium stores a computer program that, when executed by a processor, implements all steps of the personnel relationship shaping method based on intelligent entity relationship extraction, where the execution subject is a server or client, as described in the above embodiments. For example, when the processor executes the computer program, it implements the following steps: Step S101: Obtain the core resources upon which the building of trust relationships depends; classify the core resources and determine the information extraction rules for each type of core resource; the types of core resources include authentication relationships, communication relationships, social relationships, and business relationships; Step S102: Collect raw text data from each type of core resource according to the information extraction rules, and perform named entity recognition on the raw text data based on the deep learning model to obtain the basic corpus of each type of raw text data; Step S103: Extract target triples from the basic corpora of each type using entity relation extraction methods, either individually or in combination. Specifically, authentication relations are extracted using named entity relation extraction based on joint learning, and entity relation extraction combining deep learning and remote supervision; connectivity relations are extracted using entity relation extraction based on deep learning, and entity relation extraction combining deep learning and remote supervision; social relations are extracted using joint entity relation extraction based on contrastive learning and text semantic enhancement; and business relations are extracted using named entity relation extraction based on joint learning, and entity relation extraction based on contrastive learning and text semantic enhancement. Step S104: Construct a trust relationship knowledge graph based on triples for each type of target triple.

[0122] As can be seen from the above description, the computer-readable storage medium provided in this application embodiment, based on the network structure and personnel relationship shaping method of intelligent entity relationship extraction, can start from the trust relationship analysis and shaping needs of specific scenarios in different industry fields such as education, law, and catering. It can realize entity identification of massive multi-source heterogeneous target network situation data such as Internet data, passwords, public software tools, proprietary software tools, and technical knowledge bases, analyze and extract the association relationship of trust entities such as authentication, communication, social, and business, and construct a trust relationship knowledge graph based on the relationship graph model to realize the comprehensive shaping of network structure relationships and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network situation data.

[0123] Embodiments of this application also provide a computer program product capable of implementing all steps in the personnel relationship shaping method based on intelligent entity relationship extraction, where the execution subject is a server or client, as described in the above embodiments. When executed by a processor, this computer program / instruction implements the steps of the personnel relationship shaping method based on intelligent entity relationship extraction. For example, the computer program / instruction implements the following steps: Step S101: Obtain the core resources upon which the building of trust relationships depends; classify the core resources and determine the information extraction rules for each type of core resource; the types of core resources include authentication relationships, communication relationships, social relationships, and business relationships; Step S102: Collect raw text data from each type of core resource according to the information extraction rules, and perform named entity recognition on the raw text data based on the deep learning model to obtain the basic corpus of each type of raw text data; Step S103: Extract target triples from the basic corpora of each type using entity relation extraction methods, either individually or in combination. Specifically, authentication relations are extracted using named entity relation extraction based on joint learning, and entity relation extraction combining deep learning and remote supervision; connectivity relations are extracted using entity relation extraction based on deep learning, and entity relation extraction combining deep learning and remote supervision; social relations are extracted using joint entity relation extraction based on contrastive learning and text semantic enhancement; and business relations are extracted using named entity relation extraction based on joint learning, and entity relation extraction based on contrastive learning and text semantic enhancement. Step S104: Construct a trust relationship knowledge graph based on triples for each type of target triple.

[0124] As can be seen from the above description, the computer program product provided in this application embodiment, based on the network structure and personnel relationship shaping method of intelligent entity relationship extraction, can start from the trust relationship analysis and shaping needs of specific scenarios in different industry fields such as education, law, and catering. It can realize entity recognition of massive multi-source heterogeneous target network situation data such as Internet data, passwords, public software tools, proprietary software tools, and technical knowledge bases, analyze and extract the association relationship of trust entities such as authentication, communication, social, and business, and construct a trust relationship knowledge graph based on the relationship graph model to realize the comprehensive shaping of network structure relationships and personnel relationships, provide users with reliable and effective target trust relationship data, and improve the utilization rate of network situation data.

[0125] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0126] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (devices), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0127] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0128] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0129] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for shaping personnel relationships based on intelligent entity relationship extraction, characterized in that, The method includes: Acquire the core resources upon which trust relationships are built; categorize these core resources and determine the information extraction rules for each type; the types of core resources include authentication relationships, communication relationships, social relationships, and business relationships; Based on the information extraction rules, raw text data is collected from core resources of each type, and named entity recognition is performed on the raw text data based on a deep learning model to obtain the basic corpus of raw text data of each type. Entity relation extraction methods, either individually or in combination, are used to extract target triples from various types of basic corpora. Specifically, authentication relations are extracted using named entity relation extraction based on joint learning, as well as entity relation extraction combining deep learning and remote supervision; connectivity relations are extracted using entity relation extraction based on deep learning, as well as entity relation extraction combining deep learning and remote supervision; social relations are extracted using joint entity relation extraction based on contrastive learning and text semantic enhancement; and business relations are extracted using named entity relation extraction based on joint learning, as well as joint entity relation extraction based on contrastive learning and text semantic enhancement. Construct a trust relationship knowledge graph based on triples of various types.

2. The method for shaping personnel relationships based on intelligent entity relationship extraction according to claim 1, characterized in that, The core resources for authentication relationships include personnel attendance records, organizational rosters, and contact lists; Core resources for communication relationships include address book, call logs, SMS sending and receiving records, and email sending and receiving records; Core resources for social relationships include household registration information or contact information, and social groups; Core resources for business relationships include internal organizational rosters, address books, business documents, API call logs, and code files.

3. The method for shaping personnel relationships based on intelligent entity relationship extraction according to claim 1, characterized in that, The rules for extracting information about authentication relationships include: Grouping data from different organizations using organizational tag data from the social worker relationship resource database; Further grouping based on the data tags of each organization to obtain data at the departmental level; Based on information related to personnel's work and business, determine the authentication relationships of personnel at the resource access level; At the network level, the registered user list, access permissions, and historical records are obtained through the website application user database, permission database, resource access logs, and network domain / workgroup records. The rules for extracting core resource information about communication relationships include: Extract communication relationships between people from communication records; The communication intimacy level is extracted by communication frequency information, which includes call frequency, SMS sending and receiving frequency, and email sending and receiving frequency. The rules for extracting core resource information from social relationships include: Extract social relationship information from notes; Use social group data to identify and classify different types of social relationships; The rules for extracting core resource information related to business relationships include: Business relationships are obtained through business association information; the business association information includes communication and interaction records, business transaction records, and program releases. Analyze entity relationships in business processes based on business documents and API call logs.

4. The method for shaping personnel relationships based on intelligent entity relationship extraction according to claim 1, characterized in that, The steps for obtaining basic corpora of various types of raw text data by performing named entity recognition on raw text data based on a deep learning model include: The original text data is vectorized to obtain vectorized text by dynamically training word vectors using a pre-trained language model. The vectorized text is input into the feature extractor to further extract text features and obtain a high-level feature vector. The high-level feature vectors are input into the decoding network to obtain the entity label sequence; Based on entity label sequences, content containing corresponding entities is filtered from the original text data and assembled into a target sentence set, which serves as the basic corpus for relation extraction.

5. The method for shaping personnel relationships based on intelligent entity relationship extraction according to claim 1, characterized in that: Methods for extracting authentication relations include those employing joint learning-based named entity relation extraction and entity relation extraction combining deep learning and remote supervision: Weakly labeled data in the authentication corpus are correlated and matched with social engineering relationship resource database to generate weakly labeled corpus. Segmented convolutional neural network is used to filter noise in the weakly labeled corpus to obtain candidate triples. Candidate triples are associated with verifiable authentication-type basic corpora to obtain hierarchical information text; the hierarchical information text is then input into a joint model of bidirectional neural network and tree structure neural network for joint classification to obtain corrected triples; the hierarchical information includes organizations, departments, and personnel; Integrate candidate triples and modified triples to obtain the target triple; Methods for extracting connectivity relationships using entity relation extraction based on deep learning, and entity relation extraction combining deep learning and remote supervision, include: The basic corpus of communication categories containing communication relationships is input into the neural network as a high-quality labeled corpus for feature extraction to obtain a basic relation library; Weakly labeled corpora are generated by matching basic corpora without connectivity relationships with basic relation databases, and noise in the weakly labeled corpora is filtered out using a segmented convolutional neural network to obtain enhanced triples. Integrate the basic relational database with enhanced triples to obtain the target triples; Methods for social relation extraction that employ text semantic enhancement based on contrastive learning for joint entity relation extraction include: Contrast samples are generated from basic social corpora and input into a pre-trained encoder. The encoder parameters are optimized through a contrastive learning module to obtain enhanced semantic representations. Based on enhanced semantic representation, candidate entities and candidate relation tuples are generated by the Span method. The candidate entities and candidate relation tuples are then input into an end-to-end joint neural network model for joint classification, and the target triplet is output. Methods for extracting business relationships include joint learning-based named entity relation extraction and text semantic enhancement-based entity relation joint extraction based on contrastive learning. The basic business corpus is input into an end-to-end joint neural network model for joint classification to obtain a basic business relationship database. Extract texts containing ambiguous terms from the basic business corpus, and construct contrast samples with entities in the basic business relation database as the core. Obtain enhanced semantic representations through contrastive learning. Input the enhanced semantic representations into the joint extraction sub-model to correct errors in the basic business relation database and supplement new relations. Integrate the basic business relationship database with the corrected relationship data to output the target triplet.

6. The method for shaping personnel relationships based on intelligent entity relationship extraction according to claim 1, characterized in that, The deep learning-based entity relation extraction method includes: The entity relations of the text are annotated to obtain entity relation annotated corpus; the methods for annotating entity relations include manual annotation and generating relation tags by matching with a remote knowledge base; The entity relation annotation corpus is processed using a neural network model to construct multi-dimensional feature vectors that represent the semantic information of words; the feature vectors include word vectors, position vectors, and grammatical relation vectors. The feature vectors are input into the neural network for feature extraction; weights are calculated on the extracted features, and the possible relationship types between entities are determined based on the semantic features of the weights, and entity relationship pairs are output. The named entity relation extraction method based on joint learning includes: An end-to-end joint neural network model is constructed, which integrates bidirectional neural network and tree structure neural network modules. The bidirectional neural network is used for entity recognition tasks, and the tree structure neural network is used for relation extraction tasks. The text is segmented into continuous text fragments as candidate entities; the candidate entities are then combined in pairs to form candidate relation tuples. The candidate entity and candidate relation tuples are input into the end-to-end joint neural network model. The candidate entity and candidate relation tuples are jointly classified by the bidirectional neural network and the tree structure neural network, and the relation type is output. The entity relation extraction method combining deep learning and remote supervision includes: The text is matched with an existing knowledge base, and entity pairs in the text that are consistent with the existing knowledge base are labeled to generate weakly labeled corpus. For weakly labeled corpora, feature embeddings that fuse word vectors and position vectors are constructed, and the two types of vectors are combined to form the model input features; A segmented convolutional neural network is used to divide the model input features of entity pairs into three segments according to the entity position, and pooling operations are performed on each segment to obtain sentence feature vectors. The feature vectors of sentences containing the same pair of entities are packaged together, and a multi-instance learning method is used to select the sentence with the highest semantic relation probability within the package as the core indicator sentence of the entity pair, thereby obtaining the relation type of the entity pair. The text semantic enhancement entity relation joint extraction method based on contrastive learning includes: The same text and entity context are repeatedly input into the pre-trained language model. Text embeddings with similar semantics but different expressions are generated through standard random masks to obtain positive sample pairs. The entity contexts of other texts in the same training batch are selected and combined with the entity context of the current text to form differential sample pairs to obtain negative sample pairs. By inputting positive and negative sample pairs into the contrast learning module, the semantic representations of positive sample pairs become more similar, while the semantic representations of negative sample pairs become more distant, thus obtaining an enhanced semantic representation. The enhanced semantic representation is segmented into continuous text fragments as candidate entities; candidate entities are combined in pairs to form candidate relation tuples; candidate entities and candidate relation tuples are input into an end-to-end joint neural network model for joint classification, and the relation type is output.

7. The method for shaping personnel relationships based on intelligent entity relationship extraction according to claim 1, characterized in that, The steps for constructing a trust relationship knowledge graph based on triples of various types include: The target triples of various types are aggregated into a unified database, and knowledge fusion is performed through entity alignment and relation deduplication; the fused triples are converted into graph structure data and stored in the graph database.

8. A personnel relationship shaping device based on intelligent entity relationship extraction, characterized in that, The device includes: The resource classification module is used to acquire the core resources upon which trust relationships are built; it classifies the core resources and determines the information extraction rules for each type of core resource; the types of core resources include authentication relationships, communication relationships, social relationships, and business relationships; The entity recognition module is used to collect raw text data from core resources of various types according to information extraction rules, and to perform named entity recognition on the raw text data based on a deep learning model to obtain the basic corpus of raw text data of various types. The relation extraction module is used to extract target triples from various types of basic corpora using individual or combined entity relation extraction methods. Specifically, authentication relations are extracted using named entity relation extraction based on joint learning, as well as entity relation extraction combining deep learning and remote supervision; connectivity relations are extracted using entity relation extraction based on deep learning, as well as entity relation extraction combining deep learning and remote supervision; social relations are extracted using joint entity relation extraction based on contrastive learning and text semantic enhancement; and business relations are extracted using named entity relation extraction based on joint learning, as well as joint entity relation extraction based on contrastive learning and text semantic enhancement. The results output module is used to construct a trust relationship knowledge graph based on triples of various types of target triples.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the personnel relationship shaping method based on intelligent entity relationship extraction as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the personnel relationship shaping method based on intelligent entity relationship extraction as described in any one of claims 1 to 7.

Images