Method for mining relationship between multi-source entity attributes based on knowledge graph

Abstract

The application provides a kind of multi-source entity attribute relationship mining method based on knowledge graph, by combining knowledge graph, community detection, semantic embedding and clustering technology, realize the efficient mining and integration of multi-source entity attribute relationship.Method includes four main stages, multi-source entity attribute extraction: extract entity and its attribute information from domain text, obtain structured attribute data through BIO annotation and feature extraction, multi-entity alignment: adopt knowledge graph embedding method to vectorize multi-source entity, realize the unified identification of same entity in different data sources, multi-source entity attribute relationship mining: build intra-cluster attribute relationship knowledge graph, true value discovery: true value discovery is carried out through community detection algorithm and Expectation-Maximization algorithm.The method can accurately identify the attribute relationship of multi-source entity in large-scale and complex data environment, effectively improve the accuracy of multi-source data integration and the efficiency of true value discovery, suitable for intelligent question answering, information extraction and other application scenarios.

Description

Technical Field

[0001] This invention relates to a method for mining relationships between attributes of multi-source entities based on knowledge graphs, belonging to the fields of knowledge graphs, data fusion, and artificial intelligence. Background Technology

[0002] With the rapid development of information technology, the scale and complexity of multi-source data are growing exponentially. Entity attribute information in fields such as healthcare, financial risk control, and the Internet of Things is often provided by multiple independent sources and contains significant conflicts. For example, diagnostic records for the same patient may differ significantly between different hospitals, descriptions of the same event on social media may contradict each other, and sensor network monitoring data may deviate from the true value due to noise interference. Against this backdrop, how to extract the truth value of entity attributes from conflicting information has become a core challenge in the field of data intelligence. Traditional truth discovery methods are usually based on statistical principles or simple aggregation strategies, but they expose a series of limitations in complex, multi-source, heterogeneous data relationships.

[0003] Traditional methods often treat different attributes of an entity (such as "author" and "publication year" of a book) as independent variables, neglecting the semantic relationships between attributes, leading to information fragmentation and decision-making bias. For example, in medical diagnosis, relying solely on symptom descriptions while ignoring genetic testing results may result in misdiagnosis of hereditary diseases; in product authenticity verification, isolated analysis of "production date" and "packaging batch" makes it difficult to identify counterfeiting patterns. However, existing methods largely rely on manually predefined static rules, making it difficult to adapt to dynamic knowledge updates. Diagnostic criteria in the medical field iterate frequently with research progress. In the early stages of the pandemic, the static rule model failed in early data fusion due to the failure to integrate new indicators such as "loss of smell" in a timely manner.

[0004] To address these issues, academia has proposed improvement schemes such as probabilistic graphical models, weighted voting, and deep learning, but all have significant limitations. Probabilistic graphical models, by iteratively optimizing the joint probability distribution of source credibility and attribute truth values, can model soft dependencies, but their computational complexity increases exponentially with the number of attributes, and they struggle to incorporate domain-specific prior knowledge. Voting mechanisms are simple to implement but are easily misled by coordinated fraud, exhibiting a prominent cold-start problem. Deep learning models perform exceptionally well in processing unstructured text, but their black-box nature leads to poor interpretability and limited cross-domain transferability, restricting their application in high-risk scenarios such as healthcare and finance.

[0005] Knowledge graphs, as a type of structured semantic network, offer a novel approach to multi-source truth discovery. They explicitly encode domain knowledge through entity relationships, supporting dynamic expansion and multimodal fusion. However, current research has not fully realized their potential: static graphs struggle to adapt to real-time data streams, entity mapping between unstructured data and graphs relies on manually generated rules, and large-scale reasoning faces efficiency bottlenecks. The core challenge of current truth discovery technologies lies in the complex relationships between multi-source heterogeneous data and the insufficient semantic understanding capabilities of traditional methods. A novel framework is urgently needed to deeply integrate domain knowledge, uniformly process multimodal data, and balance accuracy and efficiency. Summary of the Invention

[0006] This invention proposes a multi-source entity attribute relationship mining method based on knowledge graphs, aiming to solve the problems of entity attribute conflicts, strong relationship implicitness, and low truth inference efficiency in multi-source data. For similar entities from multiple sources, this method systematically achieves attribute relationship mining and truth discovery by combining semantic association modeling and dynamic alignment techniques of knowledge graphs with cluster analysis and statistical optimization. This method is applicable to scenarios such as intelligent question answering, knowledge graph construction, and information extraction, effectively improving the accuracy and reliability of multi-source data integration.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for mining relationships between attributes of multi-source entities based on knowledge graphs, the method comprising the following steps:

[0008] A. Multi-source entity attribute extraction stage: Given source text, entities, and predefined attribute names, the task scenario is to extract attribute values ​​from the source text.

[0009] B. In the multi-source entity alignment stage, the same entity from different data sources is identified and merged. An entity alignment method is established, comprising three modules: an embedding module, an interaction module, and an alignment module. The embedding module is responsible for embedding knowledge graphs using knowledge representation learning techniques. The interaction module is responsible for mapping the embedding spaces of different knowledge graphs to the same vector space based on the aligned entity pairs. The alignment module is responsible for obtaining the entity alignment result based on the distance or similarity between entities in the vector space. By reducing the high-dimensional entities and relationships, a low-dimensional vector numerical representation is obtained. Furthermore, the semantic relationship between two objects is characterized by the distance between two low-dimensional dense vectors.

[0010] In the C stage of mining relationships between multi-source entity attributes, entities from different sources are clustered, grouping entities with similar attributes into the same cluster. The K-means algorithm is used, and features are derived based on the multi-dimensional attributes of the entities. Within each cluster, potential attribute relationships are extracted by analyzing the similarity and common attributes between entities, constructing a knowledge graph within the cluster. Nodes represent attributes within the cluster, and edges represent relationships between different attributes.

[0011] D utilizes the relationships between attributes of multi-source entities for truth discovery. Through a community detection algorithm, it identifies communities with close relationships between attributes in the graph. If certain attributes appear isolated or significantly conflict with other attributes throughout the graph, these attributes are considered erroneous. Combining the weights of the data sources, attribute value similarity, and inter-source relationships, a confidence level is assigned to each attribute value. The Expectation-Maximization algorithm continuously updates the reliability of the data sources and the confidence levels of the attribute values ​​to converge to the final truth.

[0012] The specific steps of the multi-source entity attribute extraction stage are as follows:

[0013] A1. Select a field and collect a large amount of text data for that field;

[0014] A2. Tag the text data into BIO tag format, where B represents the beginning of an entity or attribute value, I represents the inside of an entity, and O represents the non-entity part, thus obtaining the training dataset. Each sample is a sequence of words in a sentence and its corresponding BIO tag.

[0015] A3. Extract features from the text that can help the model classify, including lexical features, part-of-speech features, word morphological features, and contextual features. Combine these features into a feature vector for each word, which is then used by the machine learning algorithm for classification.

[0016] A4. Train a Conditional Random Field (CRF) model based on the training dataset. Define the input as a sequence of feature vectors (in the format of a list, where each list item is the feature and label of a word), and the output as the corresponding BIO label sequence. CRF will learn the probability of labeling each word through the feature vectors based on the sequence context, calculate the weight of each label, and optimize the label sequence by combining global path information.

[0017]

[0018] A5. Optimize the CRF model by adjusting the regularization parameters c1 and c2 to balance the model's complexity and fit. Observe the changes in model performance by selecting different combinations of features through experiments. Evaluate the model's generalization ability through K-fold cross-validation to ensure that the model performs generally consistently across different datasets.

[0019] A6. Input new text, extract its features, and use the trained CRF model to make predictions;

[0020] A7. Process the output BIO tags and extract the corresponding attribute values;

[0021] A8. When multiple candidates exist for the same attribute, conflict resolution is performed based on contextual information or confidence scores, and the relevant attribute values ​​are retained.

[0022] A9. Use the `classification_report` tool in sklearn to evaluate the performance of the CRF model using precision, recall, and F1 score. Based on the evaluation results, further optimize model features, hyperparameters, or perform more data annotation.

[0023] The specific steps of the multi-source entity alignment stage are as follows:

[0024] B1. Embedding a knowledge graph using a translation model: Given a training set S consisting of a triple (h, l, t) of two entities h, t∈E (entity set) and a relation l∈L (relation set), construct the objective function L(h, l, t)=||h+lt||, where ||| represents the distance between vectors, using L1 or L2 distance. Use random initialization to initialize all entities and relations as low-dimensional vectors, and use stochastic gradient descent (SGD) to minimize the objective function. After multiple iterations, the final embedding vector is obtained.

[0025] B2. Based on the aligned entities and combined with the mapping function, Procrustes analysis based on aligned entity pairs is used to map the embedding spaces of different knowledge graphs into the same spatial vector;

[0026]

[0027] B3. Within a unified embedding space, calculate the cosine similarity between entity vectors in different knowledge graphs. Based on the calculation results, set a similarity threshold, perform cosine similarity calculation on all entities, and generate the final entity alignment result.

[0028] .

[0029] The specific steps in the stage of constructing a knowledge graph to mine the relationships between multiple entity attributes are as follows:

[0030] C1. Clean and normalize multi-source entity data to ensure that all attribute values ​​are comparable. In the data cleaning stage, remove duplicate entity data, check and delete records with many missing or abnormal values, and correct spelling errors and non-standard naming in attribute values. In the data normalization stage, unify the names of the same attributes in multi-source attributes, unify the format of numerical data (such as dates, prices, etc.), and standardize the data when necessary to eliminate the influence of units.

[0031]

[0032] in, It is standardized ,

[0033] C2. Vectorize the multidimensional attributes of each entity. For categorical attributes, use one-hot encoding to convert them into numerical vectors. For numerical attributes, directly use the normalized values. Thus, the attributes of each entity are combined into a multidimensional vector, where... It is the value of the i-th attribute;

[0034]

[0035] C3. Determine the optimal number of clusters K using the silhouette coefficient method. The silhouette coefficient ranges from [−1, 1], with a larger value indicating better clustering performance.

[0036]

[0037] , The measure is the relative difference between the density of a sample within its cluster and its separation from neighboring clusters.

[0038] The cluster number with the highest silhouette coefficient is selected as K. Based on the multidimensional attributes of entities, the K-means clustering algorithm is applied to group entities with similar attributes into the same cluster. The objective function is shown below, where... It is a cluster. It is the cluster center;

[0039]

[0040] First, randomly select K initial cluster centers. For each sample Calculate its Euclidean distance from the center of each cluster and assign it to the nearest cluster;

[0041]

[0042] Update the center of each cluster to the mean of all samples in that cluster;

[0043]

[0044] Repeat the above steps until cluster allocation no longer changes or the objective function is achieved. convergence,

[0045] C4. Construct an attribute graph for each cluster, where nodes represent attributes and edges represent relationships between attributes. The edges connecting attributes in the graph represent their correlation, which is measured by the frequency of co-occurrence. Higher frequencies indicate stronger relationships between attribute pairs.

[0046]

[0047] in, These are the two attributes of the relationship that are being measured. Refers to the simultaneous occurrence in the dataset Attributes and The number of entities with attributes It is the total number of all entities in the dataset.

[0048] The truth discovery phase, which utilizes the relationships between attributes of multi-source entities, specifically includes the following steps:

[0049] D1. Using a community detection algorithm, community detection is performed on a knowledge graph. Multiple communities are identified based on the attribute relationships in the graph. Each community has high consistency or dependency in the attribute relationships, while the attribute relationships between different communities are relatively weak.

[0050] D2. Assign weights based on the credibility of each data source, and evaluate the credibility of each attribute by the similarity of edges in the attribute relationship graph. Isolated or conflicting attribute values ​​can be regarded as potentially erroneous data and should have a lower credibility.

[0051] D3. In each iteration, combining the weights of the data sources and the similarity of the attribute values, update the confidence of each attribute value based on the confidence of the attribute value relative to the current truth value. Further, update the weights of each data source based on the confidence of the attribute values. The closer the truth value provided by the data source is to the current truth value, the higher the weight. Repeat this iteration until the confidence of the attribute values ​​and the data sources converges to a stable value.

[0052] D4. Based on the confidence level of the final attribute values, select the attribute value with the higher confidence level as the final truth value.

[0053] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the aforementioned method for mining relationships between multi-source entity attributes based on knowledge graphs.

[0054] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the aforementioned knowledge graph-based method for mining relationships between multi-source entity attributes.

[0055] Beneficial Effects: The multi-source entity attribute relationship mining method based on knowledge graphs provided in this invention has the following beneficial effects compared with existing technologies:

[0056] (1) By performing attribute relationship analysis on entities within a cluster after clustering, this invention can identify potential relationships and commonalities between attributes and construct a more fine-grained attribute relationship map. This process can effectively extract attribute combinations with consistency or dependency, improving the accuracy of attribute association identification. Compared with general analysis methods based on association rules or simple similarity, the results are better.

[0057] (2) In the truth-finding stage, this invention utilizes a community detection algorithm and an Expectation-Maximization (EM) algorithm, combined with the weights of the data sources, attribute value similarity, and inter-source relationships for iterative optimization. Each update considers the reliability of the data sources and the relationships between attributes. Compared to traditional voting or simple weighted average truth-finding methods, the method of this invention can dynamically adjust the weights of the data sources, making the final truth more reliable.

[0058] (3) This invention, by introducing efficient clustering and knowledge graph construction, is suitable for processing large-scale, multi-source, and heterogeneous entity data. Compared with existing truth discovery methods for small-scale, single-source data, this invention is suitable for complex entity attribute processing tasks in the context of big data, and has good scalability and universality. Specific scenarios include, but are not limited to: In the medical field, through attribute clustering analysis, the relationships between disease-related attributes (such as symptoms, causes, and treatment methods) are mined to identify potential commonalities between different diseases (such as multiple diseases sharing certain symptoms or treatment plans). In the truth discovery stage, data source weights and attribute relationships are optimized to reliably determine the therapeutic effect relationship between diseases and drugs, thereby constructing a more accurate medical knowledge graph and providing scientific support for intelligent diagnosis and treatment systems. In the field of smart cities, this invention, through clustering analysis of the attribute relationships of urban entities (such as buildings, roads, and vehicles), mines potential commonalities such as building density and traffic congestion, and dynamically optimizes data source weights to improve the accuracy of attribute data, ultimately constructing a high-quality urban knowledge graph to provide decision support for traffic planning, population flow prediction, etc. Attached Figure Description

[0059] Figure 1 This is a flowchart of the method of the present invention.

[0060] Figure 2 This is a schematic diagram of the entity alignment stage of the method of the present invention.

[0061] Figure 3 This is a schematic diagram of the hierarchical relationship of the method of the present invention.

[0062] Figure 4 This is a diagram illustrating BIO annotation.

[0063] Figure 5 This is a detailed flowchart illustrating the method of the present invention. Detailed Implementation

[0064] The present invention will be further illustrated below with reference to the accompanying drawings and specific implementation examples. It should be understood that these examples are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0065] Example: Figure 1 The diagram shows a flowchart of a multi-source entity attribute relationship mining method based on knowledge graphs. When performing truth discovery, knowledge graphs are used to construct the relationships between multi-source entity attributes, thereby enhancing prediction accuracy and efficiency.

[0066] A. Multi-source entity attribute extraction stage:

[0067] A1. Domain Selection and Data Collection

[0068] The domain is selected as product information in e-commerce platforms. Product data is collected from platforms such as Taobao, JD.com and Pinduoduo, including fields such as product name, price, brand, description, and category.

[0069] A2. Text annotation

[0070] Text is labeled in BIO format: "B" (beginning): indicates the beginning of an attribute (such as product name, brand), "I" (inner): indicates the inner part of an attribute (such as the multi-word part of a product name), "O" (other): non-attribute parts. This labeling helps the model distinguish between entity and non-entity parts.

[0071] For example, product name: "Apple iPhone 15" → B-product name I-product name; brand: "Apple" → B-brand I-brand; price: "6999 yuan" → B-price I-price.

[0072] A3. Feature Extraction

[0073] Features are extracted for each word in the text, including: lexical features (such as the word itself), part-of-speech features (such as verbs and nouns), contextual features (such as the words before and after it), and word form features (such as uppercase letters and numbers).

[0074] A4. Training a Conditional Random Field (CRF) Model

[0075] The labeled BIO tags and feature vectors are input into the CRF model, which learns the context and features of the text to train a model that can predict attribute values.

[0076] A5. CRF Model Tuning

[0077] K-fold cross-validation was used to evaluate the model's performance, and the generalization ability of the model on different datasets was optimized by adjusting the regularization parameters c1 and c2.

[0078] A6. New Text Input

[0079] Input new product text data, use a trained CRF model to make predictions, and label the entities and attributes in the product information.

[0080] A7. BIO Tag Processing and Attribute Extraction

[0081] Extract attribute values ​​from the BIO tags output by the model. For example, such as... Figure 4 As shown, “Apple iPhone 12 was released in 2020”, its BIO label is: “Apple” → B-brand, “iPhone” → I-brand, “12” → I-brand, “in” → O-non-physical, “2020” → B-time, “year” → I-time, “release” → O-non-physical.

[0082] A8. Attribute Conflict Resolution

[0083] When the same attribute values ​​extracted from different data sources are inconsistent, the decision to retain a value is made based on the context or the confidence score of the data source. For example, if the same product is labeled as "iPhone 14" and "iPhone 15" on different platforms, the conflict can be resolved by combining contextual information (such as price and release time) or confidence scores, and the attribute value with higher confidence can be retained.

[0084] A9. Model Performance Evaluation

[0085] The performance of the CRF model was evaluated using the sklearn's classification_report tool by precision, recall, and F1 score, and further optimization was performed based on the evaluation results.

[0086] B. Multi-source entity alignment stage:

[0087] B1. Knowledge Graph Embedding

[0088] Use translation models (such as TransE) to embed entities in knowledge graphs. For example, map the representations of "Apple" and "iPhone 15" across multiple platforms into the same embedding space. Figure 2 As shown, entities in KG1 and KG2 are mapped to a unified embedding space through the embedding module.

[0089] B2. Embedded Space Mapping

[0090] According to the embedding space, map product attributes from different platforms (such as "Apple" and "苹果") to a unified space, so that the same product has similar representations in different data sources.

[0091] B3. Entity Alignment

[0092] In the unified embedding space, calculate the cosine similarity between entities. For example, the cosine similarity between "Apple" and "苹果" is relatively high, so they are considered to be the same entity. As Figure 2 shown, the aligned entity pairs are fed back to the embedding module and the interaction module for further optimization.

[0093] C. Stage of constructing a knowledge graph to mine relationships between multiple entity attributes:

[0094] C1. Data Cleaning and Normalization

[0095] Clean the product data of each platform to ensure that the format of the same attribute value is consistent. For example, unify the price fields of different platforms into the format of "Price (Yuan)".

[0096] C2. Vectorized Representation

[0097] Convert multi-dimensional attributes of products (such as price, brand, description) into vector representations. Each attribute is represented by a vector, and the dimension of the vector can be designed based on the number of attribute features.

[0098] C3. K-means Clustering

[0099] Use the K-means clustering algorithm to cluster similar products into one category according to the multi-dimensional attributes (such as price and brand) of the products.

[0100] C4. Constructing Attribute Relationship Graph

[0101] Within each cluster, construct an attribute relationship graph, where nodes represent product attributes (such as "Price" and "Brand"), and edges represent the relationships between them (such as the correlation between price and brand). For example, an edge is established between the brand "Apple" and the price range "5000-7000 Yuan" to indicate that there is a strong correlation between them. As Figure 3 shown, the process of constructing the knowledge graph demonstrates how to extract objects and attributes from multiple data sources and construct them into a graph structure.

[0102] D. Stage of truth discovery using the relationships between multi-source entity attributes:

[0103] D1. Community Detection

[0104] Community detection algorithms (such as the Louvain algorithm) are used to perform community detection on the constructed knowledge graph, identifying communities between attributes such as price, brand, and category. For example, price and brand may form one community, while description and category may form another.

[0105] D2. Credibility Assessment

[0106] The credibility of attribute values ​​is assessed by combining the credibility weights of each data source with the similarity in the attribute relationship graph. Isolated attribute values ​​or those that conflict with other attributes (e.g., a product's price is significantly higher than other similar products) may be considered erroneous data with low credibility.

[0107] D3. Initial Truth Value Calculation

[0108] For numerical attributes, a weighted average is used; for categorical attributes, the most frequent value is selected through weighted voting.

[0109] D4. Iterative Update

[0110] The EM algorithm updates the weights of data sources based on the credibility of attribute values ​​in each iteration. For example, if a data source shows a significant difference between the product price and the actual market price, the weight of that data source will be reduced, thus affecting the final attribute credibility. Figure 5 As shown, the iterative process of the EM algorithm includes initial confidence assignment, algorithm iteration, and convergence check.

[0111] D5. Choosing the final truth value

[0112] Based on the updated attribute credibility, the attribute value with higher credibility is selected as the final truth value. For example, price information from multiple platforms will eventually converge to a relatively consistent price, which will be used as the final truth value for that product.

[0113] Through the above steps, product attributes are extracted from different data sources, and clustering and knowledge graph construction are performed based on the relationships between multi-source entity attributes. Finally, through community detection and EM algorithm optimization, truth discovery is achieved from multi-source data. This method can effectively integrate product information from multiple e-commerce platforms, discover the most reliable product attributes, and improve the accuracy and efficiency of data integration.

[0114] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A method for mining relationships between attributes of multi-source entities based on knowledge graphs, characterized in that, The method includes the following steps: In the multi-source entity attribute extraction stage, the research domain is first determined, and a large amount of domain-related data is collected from public data sources, domain databases, or crawled datasets as source text. The text data is labeled into BIO tag format to obtain the training dataset. Each sample is a sequence of words in a sentence and its corresponding BIO tag. Given an entity and predefined attribute names, the task scenario is to extract attribute values ​​from the source text. In the multi-source entity alignment stage (B), the same entity from different data sources is identified and merged. An entity alignment method is established, comprising an embedding module, an interaction module, and an alignment module. The embedding module is responsible for embedding the knowledge graph using knowledge representation learning techniques. The interaction module is responsible for mapping the embedding spaces of different knowledge graphs to the same vector space based on the aligned entity pairs. The alignment module is responsible for obtaining the entity alignment result based on the distance or similarity between entities in the vector space. By reducing the high-dimensional entities and relationships, a low-dimensional vector numerical representation is obtained. Furthermore, the semantic relationship between two objects is characterized by the distance between two low-dimensional dense vectors. In the C stage of mining relationships between multi-source entity attributes, entities from different sources are clustered, grouping entities with similar attributes into the same cluster. The K-means algorithm is used, and features are derived based on the multi-dimensional attributes of the entities. Within each cluster, potential attribute relationships are extracted by analyzing the similarity and common attributes between entities, constructing a knowledge graph within the cluster. Nodes represent attributes within the cluster, and edges represent relationships between different attributes. D utilizes the relationships between attributes of multi-source entities for truth discovery, employing a community detection algorithm to perform community detection on a knowledge graph, identifying multiple communities based on attribute relationships within the graph. Weights are assigned based on the credibility of each data source, and the credibility of each attribute is evaluated using the similarity of edges in the attribute relationship graph. Isolated or conflicting attribute values ​​are considered potentially erroneous data and should have lower credibility. In each iteration, combining the weights of the data sources and the similarity of attribute values, the credibility of each attribute value is updated based on its relevance to the current truth value, and the weights of each data source are updated based on the credibility of the attribute values. The closer the truth value provided by the data source is to the current truth value, the higher its weight. This process is repeated iteratively, using the Expectation-Maximization algorithm to update the reliability of the data sources and the credibility of the attribute values ​​until both converge to stable values. Based on the final attribute value credibility, the attribute value with higher credibility is selected as the final truth value.

2. The method for mining relationships between multi-source entity attributes based on knowledge graphs according to claim 1, characterized in that: The specific steps for extracting attributes from multi-source entities are as follows: A1. Select a field and collect a large amount of text data for that field; A2. Tag the text data into BIO tag format, where B represents the beginning of an entity or attribute value, I represents the inside of an entity, and O represents the non-entity part, thus obtaining the training dataset. Each sample is a sequence of words in a sentence and its corresponding BIO tag. A3. Extract features from the text that can help the model classify, including lexical features, part-of-speech features, word morphological features, and contextual features. Combine these features into a feature vector for each word, which is then used by the machine learning algorithm for classification. A4. Train a Conditional Random Field (CRF) model based on the training dataset. Define the input as a sequence of feature vectors in list format, where each list item represents the features and label of a word. The output is the corresponding BIO label sequence. CRF will learn the probability of labeling each word based on the sequence context and through the feature vectors, calculate the weight of each label, and optimize the label sequence by combining global path information. It is a probabilistic graphical model that calculates the joint probability distribution of a sequence of labels, using the following formula: X: Input sequence, Y: Label sequence, : Characteristic function Feature weights Current location Current location tag Previous position label, Normalization factor A5. Optimize the CRF model by adjusting the regularization parameters c1 and c2 to balance the model's complexity and fit. Observe the changes in model performance by selecting different combinations of features through experiments. Evaluate the model's generalization ability through K-fold cross-validation to ensure that the model performs generally consistently across different datasets. A6. Input new text, extract its features, and use the trained CRF model to make predictions; A7. Process the output BIO tags and extract the corresponding attribute values; A8. When multiple candidates exist for the same attribute, conflict resolution is performed based on contextual information or confidence scores, and the relevant attribute values ​​are retained. A9. Use the `classification_report` tool in sklearn to evaluate the performance of the CRF model using precision, recall, and F1 score. Based on the evaluation results, further optimize model features, hyperparameters, or perform more data annotation.

3. The method for mining relationships between multi-source entity attributes based on knowledge graphs according to claim 1, characterized in that: The specific steps of the multi-source entity alignment stage are as follows: B1. Embedding a knowledge graph using a translation model: Given a training set S consisting of a triple (h, l, t) of two entities h, t∈E and a relation l∈L, construct the objective function L(h, l, t)=||h+lt||, where ||| represents the distance between vectors, using L1 or L2 distance. Use random initialization to initialize all entities and relations as low-dimensional vectors, and use stochastic gradient descent (SGD) to minimize the objective function. After multiple iterations, the final embedding vector is obtained. B2. Based on the aligned entities and combined with the mapping function, Procrustes analysis based on aligned entity pairs is used to map the embedding spaces of different knowledge graphs into the same spatial vector; : Two are just embedding vectors of the graph. : Mapping matrix, B3. Within a unified embedding space, calculate the cosine similarity between entity vectors in different knowledge graphs. Based on the calculation results, set a similarity threshold, perform cosine similarity calculation on all entities, and generate the final entity alignment result. 。 4. The method for mining relationships between multi-source entity attributes based on knowledge graphs according to claim 1, characterized in that: The specific steps in the stage of constructing a knowledge graph to mine the relationships between multiple entity attributes are as follows: C1. Clean and standardize multi-source entity data to ensure that all attribute values ​​are comparable. In the data cleaning stage, remove duplicate entity data, check and delete records with many missing or abnormal values, and correct spelling errors and non-standard naming in attribute values. In the data standardization stage, unify the names of the same attributes in multi-source attributes, unify the format of numerical data, and standardize the data when necessary to eliminate the influence of units. in, It is standardized , C2. Vectorize the multidimensional attributes of each entity. For categorical attributes, use one-hot encoding to convert them into numerical vectors. For numerical attributes, directly use the normalized values. Thus, the attributes of each entity are combined into a multidimensional vector, where... It is the value of the i-th attribute; C3. Determine the optimal number of clusters K using the silhouette coefficient method. The silhouette coefficient ranges from [−1, 1], with a larger value indicating better clustering performance. For the sample The average distance to other samples in the same cluster, For the sample The average distance to the nearest neighbor cluster, The measure is the relative difference between the density of a sample within its cluster and its separation from neighboring clusters. The cluster number with the highest silhouette coefficient is selected as K. Based on the multidimensional attributes of entities, the K-means clustering algorithm is applied to group entities with similar attributes into the same cluster. The objective function is shown below, where... It is a cluster. It is the cluster center; First, randomly select K initial cluster centers. For each sample Calculate its Euclidean distance from the center of each cluster and assign it to the nearest cluster; Update the center of each cluster to the mean of all samples in that cluster; Repeat the above steps until cluster allocation no longer changes or the objective function is achieved. convergence, C4. Construct an attribute graph for each cluster, where nodes represent attributes and edges represent relationships between attributes. The edges connecting attributes in the graph represent their correlation, which is measured by the frequency of co-occurrence. Higher frequencies indicate stronger relationships between attribute pairs. in, These are the two attributes of the relationship that are being measured. Refers to the simultaneous occurrence in the dataset Attributes and The number of entities with attributes It represents the total number of all entities in the dataset.

5. 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 knowledge graph-based multi-source entity attribute relationship mining method as described in any one of claims 1 to 4.

6. A computer-readable storage medium storing computer instructions thereon, characterized in that: When executed by a processor, the computer instructions implement the knowledge graph-based method for mining relationships between attributes of multi-source entities as described in any one of claims 1-4.

Images