Deep learning-based creative application knowledge graph construction method

Abstract

The invention discloses a deep learning-based credential application knowledge graph construction method, and belongs to the technical field of knowledge graph construction, and the method comprises the steps: firstly obtaining unstructured data in the credential field, and carrying out the data preprocessing; then performing entity recognition and relationship extraction on the preprocessed unstructured data, representing the extracted entities and relationships in a graph structure form to construct a creative application knowledge graph, and storing the creative application knowledge graph in a graph database; embedding learning and relation optimization are carried out on the constructed knowledge graph; meanwhile, updating the knowledge graph in real time, and automatically updating the data and adjusting the information weight in the knowledge graph when new data related to the credential and creative field is detected; and finally, constructing a query function of the creative application knowledge graph through the composite index, the word frequency matching degree, the inverse document frequency and a comprehensive scoring mechanism, and performing result display according to comprehensive score sorting.

Description

Technical Field

[0001] The present invention relates to the technical field of knowledge graph construction, and specifically to a method for constructing a knowledge graph for information technology innovation applications based on deep learning. Background Art

[0002] With the rapid development of information technology and the continuous growth of the information and innovation field, enterprises, scientific research institutions and government departments are facing unprecedented information challenges. The information and innovation field covers many cutting-edge technology directions such as cloud computing, big data, artificial intelligence, and the Internet of Things. The rapid development of these technologies has not only promoted the transformation and upgrading of the industry, but also brought massive and complex data resources; however, how to efficiently and accurately retrieve the required information from these huge data has become one of the key issues restricting the development of the information and innovation field.

[0003] Traditional information retrieval methods mainly rely on keyword matching and text similarity calculation. This method performs well when processing structured data, but its effectiveness is often greatly reduced when faced with unstructured or semi-structured data that is widely present in the field of information technology. Unstructured data such as academic papers, technical documents, industry reports, etc. contain rich professional terms, complex semantic relationships and implicit knowledge structures. Traditional methods find it difficult to effectively capture and represent this information, resulting in low accuracy and relevance of retrieval results.

[0004] In addition, the knowledge in the field of information innovation is updated very quickly, and new technologies and new products emerge in an endless stream. Traditional information retrieval systems are difficult to track and reflect these changes in real time, resulting in poor timeliness of information retrieval. At the same time, the knowledge system in the field of information innovation is complex and highly specialized, requiring a deep understanding of the professional knowledge and logical relationships in the field in order to make accurate retrieval and judgment. Summary of the invention

[0005] The purpose of this invention is to provide a method for constructing a knowledge graph for information technology applications based on deep learning to solve the problems raised in the above-mentioned background technology.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A method for constructing a knowledge graph of information technology applications based on deep learning, the method comprising the following steps:

[0008] Step S100: Obtain unstructured data in the information innovation field and perform data preprocessing;

[0009] Step S200: performing entity recognition and relationship extraction on the pre-processed unstructured data;

[0010] Step S300: Represent the extracted entities and relationships in the form of a graph structure to construct a knowledge graph for information technology innovation applications, and store it in a graph database;

[0011] Step S400: performing embedding learning and relationship optimization on the constructed knowledge graph;

[0012] Step S500: Update the knowledge graph in real time. When new data related to the information innovation field is detected, automatically update the data and adjust the information weight in the knowledge graph;

[0013] Step S600: Construct the query function of the knowledge graph of trusted computing applications through compound indexing, word frequency matching, inverse document frequency and comprehensive scoring mechanism, and display the results according to the comprehensive score sorting.

[0014] The step S100 collects multi-source data in the information and innovation field through different collection methods and captures unstructured data therefrom; performs preprocessing operations on the captured data, including data cleaning, unified data format, word segmentation and part-of-speech tagging, stop word processing, and stem extraction, and finally obtains a data document set D = {d1, d2, ..., d N}, where N is the total number of documents, d N Represents the Nth document contained in the dataset D.

[0015] In the above technical solution, unstructured data in the field of information technology innovation are collected from multiple source channels to ensure the comprehensiveness and diversity of the data; through pre-processing operations such as data cleaning, unified format, word segmentation, part-of-speech tagging, stop word processing and stem extraction, the quality and consistency of the data are improved, laying the foundation for subsequent steps.

[0016] The step S200 identifies named entities in unstructured data as nodes in the knowledge graph through the trained BERT model. Based on the identified entities, the Transformer model is used to extract semantic relationships between entities as edges in the knowledge graph to connect different entity nodes. During the extraction process, if different descriptions of the same entity or relationship are found from multiple sources, knowledge fusion is performed to integrate information from different sources to form a unified entity and relationship representation.

[0017] In the above technical solution, through entity recognition and relationship extraction, the key entities in the text can be accurately identified, the semantic relationships in the text can be deeply understood, and the deep connections between entities can be excavated, providing high-quality nodes and rich edge information for the construction of the knowledge graph; through knowledge fusion processing of entity and relationship descriptions from different sources, information redundancy and conflict are eliminated, a unified entity and relationship representation is formed, and the accuracy and reliability of the knowledge graph are improved.

[0018] The step S300 assigns a unique identifier code to each extracted entity and relationship, and adds a timestamp to record the time when each entity and relationship was created or last updated. The encoded entities and relationships are organized in the form of a graph structure to construct an initial knowledge graph framework and store it in a graph database; the entity set in the knowledge graph is set to E = {e1, e2, ..., e i}, where e i Represents the i-th entity in the knowledge graph, and the relationship set is R = {r1, r2, …, r g}, where r g represents the g-th relation in the knowledge graph.

[0019] In the above technical solution, entities and relationships are represented in the form of a graph structure, which can intuitively display the complex knowledge network in the field of information technology innovation and facilitate understanding and analysis; using a graph database to store knowledge graphs can efficiently manage and query graph structured data; at the same time, a unique identifier and timestamp are assigned to each entity and relationship, recording their creation and update time, providing a basis for subsequent dynamic updates and weight adjustments.

[0020] The specific steps of step S400 are as follows: Step S401 performs embedding learning on the initial knowledge graph, generates an initial embedding vector for each entity in the graph, and defines e j For entity e i Neighbor entity, indicating the relationship with e i The directly connected entities in the knowledge graph optimize the embedding vector through the attention mechanism and assign different weights to adjacent entities. The specific calculation formula is as follows:

[0021]

[0022] Where W ij Represents entity e i and its neighbor entity e j The attention weights between and Entity e i and e j The initial eigenvector of , W is the weight matrix, is the weight vector of the attention mechanism, N i It is entity e i The set of neighbor entities, LeakyReLU is the activation function;

[0023] Step S402 compares the embedding vectors of each pair of entities that are not directly connected in the knowledge graph, calculates the similarity between them through the optimized embedding vectors, and decides whether to add a new relationship or modify the existing relationship according to the set similarity threshold. The similarity calculation formula is as follows:

[0024]

[0025] in Represents entity e i With e j The similarity between embedded vectors, n represents the dimension of the vector, and the similarity threshold σ is set. If Then in entity e i and entity e j Add or modify relationships between.

[0026] In the above technical solution, by introducing the attention mechanism in the embedding learning process, the model can automatically assign different weights to different neighbor entities. This weight assignment mechanism is based on the semantic relevance and importance between entities, and can highlight key information, thereby further improving the quality and accuracy of the knowledge graph.

[0027] When new information innovation field-related data is detected, step S500 automatically triggers the data processing process including data preprocessing, entity recognition, relationship extraction and knowledge fusion, and assigns a current timestamp to each newly added or updated entity and relationship to record the latest status; the information weight of each entity and relationship in the knowledge graph is calculated according to the function formula:

[0028]

[0029] where w t is the weight at the current time t, λ is the decay rate, and t0 is the timestamp of information creation or last update; when adding or updating, the weights of all entities and relationships are updated according to the function.

[0030] In the above technical solution, by introducing timestamps and weight decay rates, the system can automatically adjust the information weights of entities and relationships in the knowledge graph; as time goes by, the weight of old information gradually decreases, while the weight of new information gradually increases. This mechanism helps to maintain the timeliness and accuracy of the knowledge graph and avoid the impact of outdated information on query results.

[0031] The step S600 constructs a composite index K according to the attributes of each entity and relationship in the knowledge graph; after receiving a query request, the composite index K is used to locate the relevant entity and relationship subset, which is represented as the target document d i , calculate each target document d i The degree of match with the query statement;

[0032] Step S601 calculates the word frequency matching degree and sets the query statement

[0033] Q={q1,q2,…,q M}, set qi represents the i-th word in the query, where i = {1, 2, ..., M}, M is the total number of words in the query; for each word q in the query Q i In the target document d i The word frequency F in q Calculated by formula:

[0034]

[0035] Where n qi It is the word q i The number of times it appears in the document, K is the number of times it appears in the document d i The total number of words in ;

[0036] Step S602 calculates the inverse document frequency. For each word q in the query statement Q i , calculate the inverse document frequency DF by the formula q :

[0037]

[0038] where x qi is a word containing q i The number of documents;

[0039] Step S603 calculates the query statement Q and the target document d i The matching degree TM between them is based on the formula:

[0040]

[0041] According to the matching results and weight w t Calculate the comprehensive score S according to the formula:

[0042] S=α*TM+β*w t ;

[0043] Among them, α and β are coefficients for adjusting the matching degree and weight, and the query results are displayed in descending order according to the comprehensive score results.

[0044] In the above technical solution, a composite index is constructed by multiple attributes of entities and relationships, so that a data subset related to the query request can be quickly located; a query matching calculation system is constructed by term frequency matching and inverse document frequency, which not only considers the frequency of occurrence of query terms in the target document, but also considers the prevalence and importance of the query terms in the entire data set, and by combining these factors, the system can more accurately evaluate the relevance between the query statement and the target document; at the same time, the query results are sorted and displayed according to the comprehensive score, which can ensure that users see the most relevant and important information first.

[0045] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0046] The present invention uses a deep learning model to perform entity recognition and relationship extraction, significantly improving the accuracy and efficiency of extracting useful information from unstructured data. At the same time, in the process of entity recognition and relationship extraction, knowledge fusion is performed to integrate information from different sources to form a unified entity and relationship representation, ensuring the consistency and integrity of the knowledge graph.

[0047] The present invention optimizes the embedding vector through the attention mechanism, assigns different weights to adjacent entities, and further improves the accuracy of knowledge graph construction; ensures the timeliness and accuracy of the knowledge graph through real-time detection and automatic update of new data in the information innovation field; introduces timestamp and information weight decay mechanism to dynamically adjust the importance of entities and relationships in the knowledge graph, making the knowledge graph closer to the development status of the current field;

[0048] The present invention provides an intelligent and efficient query function by constructing a composite index, word frequency matching, inverse document frequency and a comprehensive scoring mechanism, and sorts and displays the query results according to the comprehensive score, thereby improving user experience and query efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0050] Figure 1 It is a flow chart of a method for constructing a knowledge graph for trusted computing applications based on deep learning. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] See also Figure 1 The present invention provides a technical solution: a method for constructing a knowledge graph of information innovation applications based on deep learning, the method comprising the following steps:

[0053] Step S100 collects multi-source data in the field of information technology innovation through different collection methods and captures unstructured data from it; preprocesses the captured data including data cleaning, unified data format, word segmentation and part-of-speech tagging, stop word processing, and stem extraction, and finally obtains a data document set D = {d1, d2, ..., dN}, where N is the total number of documents, d N Represents the Nth document contained in the dataset D.

[0054] In the above technical solution, unstructured data in the field of information technology innovation are collected from multiple source channels to ensure the comprehensiveness and diversity of the data; through pre-processing operations such as data cleaning, unified format, word segmentation, part-of-speech tagging, stop word processing and stem extraction, the quality and consistency of the data are improved, laying a solid foundation for subsequent steps.

[0055] Step S200 uses the trained BERT model to identify named entities in unstructured data as nodes in the knowledge graph. Based on the identified entities, the Transformer model is used to extract the semantic relationships between entities as edges in the knowledge graph to connect different entity nodes. During the extraction process, if different descriptions of the same entity or relationship are found from multiple sources, knowledge fusion is performed to integrate information from different sources to form a unified entity and relationship representation.

[0056] In the above technical solution, through entity recognition and relationship extraction, the key entities in the text can be accurately identified, and the semantic relationships in the text can be deeply understood, and the deep connections between entities can be excavated, providing high-quality nodes and rich edge information for the construction of the knowledge graph; through knowledge fusion processing of entity and relationship descriptions from different sources, information redundancy and conflict are eliminated, forming a unified and consistent entity and relationship representation, and improving the accuracy and reliability of the knowledge graph.

[0057] Step S300 assigns a unique identifier code to each extracted entity and relationship, and adds a timestamp to record the time when each entity and relationship was created or last updated. The encoded entities and relationships are organized in the form of a graph structure to construct an initial knowledge graph framework and store it in the graph database; the entity set in the knowledge graph is set to E = {e1, e2, ..., e i}, where e i Represents the i-th entity in the knowledge graph, and the relationship set is R = {r1, r2, …, r g}, where r g represents the g-th relation in the knowledge graph.

[0058] In the above technical solution, entities and relationships are represented in the form of graph structures, which can intuitively display the complex knowledge network in the field of information technology innovation and facilitate understanding and analysis; using graph databases to store knowledge graphs can efficiently manage and query graph structured data and support complex graph traversal and pattern matching operations; at the same time, each entity and relationship is assigned a unique identifier and timestamp, which records their creation and update time, providing a basis for subsequent dynamic updates and weight adjustments.

[0059] Step S400 performs embedding learning and relationship optimization on the constructed knowledge graph. The specific steps are as follows: Step S401 performs embedding learning on the initial knowledge graph, generates an initial embedding vector for each entity in the graph, and defines e j For entity e i Neighbor entity, indicating the relationship with e i The directly connected entities in the knowledge graph optimize the embedding vector through the attention mechanism and assign different weights to adjacent entities. The specific calculation formula is as follows:

[0060]

[0061] Where W ij Represents entity e i and its neighbor entity e j The attention weights between and Entity e i and e j The initial eigenvector of , W is the weight matrix, is the weight vector of the attention mechanism, N i It is entity e i The set of neighbor entities, LeakyReLU is the activation function;

[0062] Step S402 compares the embedding vectors of each pair of entities that are not directly connected in the knowledge graph, calculates the similarity between them through the optimized embedding vectors, and decides whether to add a new relationship or modify the existing relationship according to the set similarity threshold. The similarity calculation formula is as follows:

[0063]

[0064] in Represents entity e i With e j The similarity between embedded vectors, n represents the dimension of the vector, and the similarity threshold σ is set. If Then in entity e i and entity e j Add or modify relationships between.

[0065] In the above technical solution, by introducing the attention mechanism in the embedding learning process, the model can automatically assign different weights to different neighbor entities. This weight assignment mechanism is based on the semantic relevance and importance between entities, and can highlight key information, thereby further improving the quality and accuracy of the knowledge graph.

[0066] Step S500: When new information innovation related data is detected, the data processing flow is automatically triggered, including data preprocessing, entity recognition, relationship extraction and knowledge fusion. At the same time, a current timestamp is assigned to each newly added or updated entity and relationship to record the latest status; the information weight of each entity and relationship in the knowledge graph is calculated according to the function formula:

[0067]

[0068] where w t is the weight at the current time t, λ is the decay rate, and t0 is the timestamp of information creation or last update; when adding or updating, the weights of all entities and relationships are updated according to the function.

[0069] In the above technical solution, by introducing timestamps and weight decay rates, the system can automatically adjust the information weights of entities and relationships in the knowledge graph; as time goes by, the weight of old information gradually decreases, while the weight of new information gradually increases. This mechanism helps to maintain the timeliness and accuracy of the knowledge graph and avoid the impact of outdated information on query results.

[0070] Step S600 constructs a composite index K based on the attributes of each entity and relationship in the knowledge graph; after receiving a query request, locates the relevant entity and relationship subset through the composite index K, which is represented as the target document d i , calculate each target document d i The degree of match with the query statement;

[0071] Step S601 calculates the word frequency matching degree and sets the query statement

[0072] Q={q1,q2,…,q M}, set q i represents the i-th word in the query, where i = {1, 2, ..., M}, M is the total number of words in the query; for each word q in the query Q i In the target document d i The word frequency F in q Calculated by formula:

[0073]

[0074] Where n qi It is the word q i The number of times it appears in the document, K is the number of times it appears in the document d i The total number of words in ;

[0075] Step S602 calculates the inverse document frequency. For each word q in the query statement Q i , calculate the inverse document frequency DF by the formula q :

[0076]

[0077] where x qi is a word containing q i The number of documents;

[0078] Step S603 calculates the query statement Q and the target document d i The matching degree TM between them is based on the formula:

[0079]

[0080] According to the matching results and weight w t Calculate the comprehensive score S according to the formula:

[0081] S=α*TM+β*w t ;

[0082] Among them, α and β are coefficients for adjusting the matching degree and weight, and the query results are displayed in descending order according to the comprehensive score results.

[0083] In the above technical solution, a composite index is constructed by multiple attributes of entities and relationships, so that a data subset related to the query request can be quickly located; a query matching calculation system is constructed by term frequency matching and inverse document frequency, which not only considers the frequency of occurrence of query terms in the target document, but also considers the prevalence and importance of the query terms in the entire data set, and by combining these factors, the system can more accurately evaluate the relevance between the query statement and the target document; at the same time, the query results are sorted and displayed according to the comprehensive score, which can ensure that users see the most relevant and important information first.

[0084] In this embodiment, step S100 uses a web crawler to crawl unstructured data in the field of information technology from major news websites, academic databases, technical blog platforms, etc., including news reports, academic papers, technical blogs, social media posts, etc.; pre-processing operations are performed on the crawled data, including data cleaning, unified data format, word segmentation and part-of-speech tagging, stop word processing, and stem extraction, for example:

[0085] Original text: "Recently, a company released a new operating system based on Xinchuang technology, which has excellent performance in security, stability and compatibility";

[0086] After preprocessing: "Recently, a company released a new operating system based on Xinchuang technology, which performed well in terms of system security, stability and compatibility";

[0087] Step S200 identifies entities in the unstructured data through the trained BERT model: "A certain company, Xinchuang technology, a brand-new operating system"; uses the Transformer model to extract the semantic relationships between entities: A certain company - releases - a brand-new operating system;

[0088] Step S300 represents the extracted entities and relationships in the form of a graph structure, constructs a Xinchuang application knowledge graph, assigns unique identifiers and timestamps to each entity and relationship, and stores them in a graph database storage;

[0089] Step S400 initializes a random embedding vector for each entity e i ∈E; The dimension of the embedding vector is 100; uses the attention mechanism to assign weights W i to each pair of adjacent entities e j and e ij ; according to these weights, updates the embedding vector of each entity, and repeats the update process to optimize the embedding vector through multiple iterations; for each pair of entities in the knowledge graph that are not directly connected, such as "Xinchuang technology" and "security", uses the optimized embedding vector to calculate the similarity between them:

[0090]

[0091] Sets a similarity threshold σ to 0.8, then adds a relationship between entity Xinchuang technology and entity security: Xinchuang technology - affects - security; according to the result of relationship optimization, updates the entity and relationship information in the knowledge graph. Since a new relationship is added, new edges also need to be added to the knowledge graph accordingly;

[0092] Step S500 detects new data related to the Xinchuang field, automatically triggers a data processing flow including data preprocessing, entity recognition, relationship extraction, and knowledge fusion, adds a new entity "the latest Xinchuang product" and a relationship "the latest Xinchuang product - adopts - Xinchuang technology"; assigns a unique identifier and the current timestamp to "the latest Xinchuang product", adds the relationship "adopts" between "the latest Xinchuang product" and "Xinchuang technology" to the knowledge graph, and assigns a unique relationship identifier and the current timestamp;

[0093] Obtains that the timestamp of the last update of the "Xinchuang technology" entity is January 1, 2024, and the timestamp is simplified to t0 = 1. The current time is July 1, 2024, and the timestamp is simplified to t0 = 7. The decay rate λ = 0.01; calculates according to the formula:

[0094] w7 = e -0.01(7-1) = e -0.06 ≈0.941;

[0095] For the newly added "Latest Trusted Innovation Product - Adoption - Trusted Innovation Technology" relationship, its weight is initialized to 1 and gradually decays over time; the weight of "Trusted Innovation Technology" is updated to 0.941 in the graph database, and a new entity "Latest Trusted Innovation Product" and a relationship "Latest Trusted Innovation Product - Adoption - Trusted Innovation Technology" are added at the same time;

[0096] Step S600: The user submits a query request Q: "Query the latest operating system related to the information and innovation technology and its performance";

[0097] A composite index K is constructed based on the attributes of each entity and relationship in the knowledge graph (such as entity type, relationship type, timestamp, etc.); after receiving a query request, the composite index K is used to quickly locate the entity and relationship subsets related to "trusted innovation technology" and "operating system". These entities include "new operating system", "latest trusted innovation product", etc., and the relationships include "a company-released-new operating system" and "latest trusted innovation product-adopted-trusted innovation technology", etc.;

[0098] Calculate word frequency matching:

[0099] Query statement Q = {"trusted innovation technology", "latest", "operating system", "performance"}; set the target document to "latest trusted innovation product", document content: "This latest trusted innovation product uses advanced trusted innovation technology and is the latest generation of operating system. It is particularly outstanding in performance, especially in terms of stability and compatibility."

[0100] The content of the pre-processed document is: "The latest trusted innovation products use trusted innovation technology and the latest generation of operating systems with excellent performance, stability and compatibility"

[0101] Calculate word frequency: F 最新 =2 / 34≈0.06; F 操作系统 =1 / 34≈0.03;

[0102] F 信创技术 =1 / 34≈0.03; F 性能表现 =1 / 34≈0.03;

[0103] In the data document collection, there are 100 documents containing "latest", 500 documents containing "operating system", 200 documents containing "trust innovation technology", 300 documents containing "performance", and the total number of documents N is 1000;

[0104] Calculating Inverse Document Frequency: DF 最新 =log(1000 / (1+100))≈1.0;

[0105] DF 操作系统=log(1000 / (1+500))≈0.37;

[0106] DF 信创技术 =log(1000 / (1+200))≈0.69;

[0107] DF 性能表现 =log(1000 / (1+300))≈0.52;

[0108] Calculate the matching degree:

[0109] TM=0.06*1.0+0.03*0.37+0.03*0.69+0.03*0.52≈0.11;

[0110] In this embodiment, α=0.4, β=0.6 is set. Since the “latest trusted innovation product” is just added, its weight is w t =1;

[0111] The comprehensive score S = 0.4*0.11+0.6*1 = 0.644. Based on this score, the "latest trusted innovation products" and their related information are displayed in the query results in descending order according to the score.

[0112] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for constructing a knowledge graph for information technology applications based on deep learning, characterized in that: The method comprises the following steps: Step S100: Obtain unstructured data in the information innovation field and perform data preprocessing; Step S200: performing entity recognition and relationship extraction on the pre-processed unstructured data; Step S300: Represent the extracted entities and relationships in the form of a graph structure to construct a knowledge graph for information technology innovation applications, and store it in a graph database; Step S400: performing embedding learning and relationship optimization on the constructed knowledge graph; Step S500: Update the knowledge graph in real time. When new data related to the information innovation field is detected, automatically update the data and adjust the information weight in the knowledge graph; Step S600: Construct the query function of the knowledge graph of trusted computing applications through compound indexing, word frequency matching, inverse document frequency and comprehensive scoring mechanism, and display the results according to the comprehensive score sorting.

2. According to a method for constructing a knowledge graph of information innovation applications based on deep learning according to claim 1, it is characterized by: The step S100 collects multi-source data in the information and innovation field through different collection methods and captures unstructured data therefrom; performs preprocessing operations on the captured data, including data cleaning, unified data format, word segmentation and part-of-speech tagging, stop word processing, and stem extraction, and finally obtains a data document set D = {d1, d2, ..., d N }, where N is the total number of documents, d N Represents the Nth document contained in the dataset D.

3. According to a method for constructing a knowledge graph of information innovation applications based on deep learning according to claim 1, it is characterized by: The step S200 identifies named entities in unstructured data as nodes in the knowledge graph through the trained BERT model. Based on the identified entities, the Transformer model is used to extract semantic relationships between entities as edges in the knowledge graph to connect different entity nodes. During the extraction process, if different descriptions of the same entity or relationship are found from multiple sources, knowledge fusion is performed to integrate information from different sources to form a unified entity and relationship representation.

4. According to a method for constructing a knowledge graph of information innovation applications based on deep learning according to claim 1, it is characterized by: The step S300 assigns a unique identifier code to each extracted entity and relationship, and adds a timestamp to record the time when each entity and relationship was created or last updated. The encoded entities and relationships are organized in the form of a graph structure to construct an initial knowledge graph framework and store it in a graph database; the entity set in the knowledge graph is set to E = {e1, e2, ..., e i }, where e i Represents the i-th entity in the knowledge graph, and the relationship set is R = {r1, r2, …, r g }, where r g Represents the g-th relationship in the knowledge graph.

5. According to a method for constructing a knowledge graph of information innovation applications based on deep learning according to claim 1, it is characterized in that: The specific steps of step S400 are as follows: Step S401 performs embedding learning on the initial knowledge graph, generates an initial embedding vector for each entity in the graph, and defines e j For entity e i Neighbor entity, indicating the relationship with e i The directly connected entities in the knowledge graph optimize the embedding vector through the attention mechanism and assign different weights to adjacent entities. The specific calculation formula is as follows: Where W ij Represents entity e i and its neighbor entity e j The attention weights between and Entity e i and e j The initial eigenvector of , W is the weight matrix, is the weight vector of the attention mechanism, N i It is entity e i The set of neighbor entities, LeakyReLU is the activation function; Step S402 compares the embedding vectors of each pair of entities that are not directly connected in the knowledge graph, calculates the similarity between them through the optimized embedding vectors, and decides whether to add a new relationship or modify an existing relationship based on the set similarity threshold. The similarity calculation formula is as follows: in Represents entity e i With e j The similarity between embedded vectors, n represents the dimension of the vector, and the similarity threshold σ is set. If Then in entity e i and entity e j Add or modify relationships between them.

6. According to a method for constructing a knowledge graph of information innovation applications based on deep learning according to claim 1, it is characterized by: When new information innovation field-related data is detected, step S500 automatically triggers the data processing process including data preprocessing, entity recognition, relationship extraction and knowledge fusion, and assigns a current timestamp to each newly added or updated entity and relationship to record the latest status; The information weight of each entity and relationship in the knowledge graph is calculated according to the function formula: where w t is the weight at the current time t, λ is the decay rate, and t0 is the timestamp of information creation or last update; when adding or updating, the weights of all entities and relationships are updated according to the function.

7. According to a method for constructing a knowledge graph of information innovation applications based on deep learning according to claim 1, it is characterized by: The step S600 constructs a composite index K according to the attributes of each entity and relationship in the knowledge graph; after receiving a query request, the composite index K is used to locate the relevant entity and relationship subset, which is represented as the target document d i , calculate each target document d i The degree of match with the query statement; Step S601 calculates the word frequency matching degree, setting the query sentence Q = {q1, q2, ..., q M }, set q i represents the i-th word in the query, where i = {1, 2, ..., M}, M is the total number of words in the query; for each word q in the query Q i In the target document d i The word frequency F in q Calculated by formula: where n qi It is the word q i The number of times it appears in the document, K is the number of times it appears in the document d i The total number of words in ; Step S602 calculates the inverse document frequency. For each word q in the query statement Q i , calculate the inverse document frequency DF by the formula q : where x qi is the word containing q i The number of documents; Step S603 calculates the query statement Q and the target document d i The matching degree TM between them is based on the formula:

8. According to a method for constructing a knowledge graph of information innovation applications based on deep learning according to claim 1, it is characterized by: According to the matching results and weight w t Calculate the comprehensive score S according to the formula: S=α*TM+β*w t ; Among them, α and β are coefficients for adjusting the matching degree and weight, and the query results are displayed in descending order according to the comprehensive score results.

Images